CN109657229A - A kind of intention assessment model generating method, intension recognizing method and device - Google Patents
A kind of intention assessment model generating method, intension recognizing method and device Download PDFInfo
- Publication number
- CN109657229A CN109657229A CN201811290240.4A CN201811290240A CN109657229A CN 109657229 A CN109657229 A CN 109657229A CN 201811290240 A CN201811290240 A CN 201811290240A CN 109657229 A CN109657229 A CN 109657229A
- Authority
- CN
- China
- Prior art keywords
- vector
- text
- region
- sample
- word
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 97
- 239000011159 matrix material Substances 0.000 claims description 96
- 238000012549 training Methods 0.000 claims description 26
- 238000000605 extraction Methods 0.000 claims description 16
- 230000008859 change Effects 0.000 claims description 6
- 241000208340 Araliaceae Species 0.000 claims description 5
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 claims description 5
- 235000003140 Panax quinquefolius Nutrition 0.000 claims description 5
- 235000008434 ginseng Nutrition 0.000 claims description 5
- 238000005516 engineering process Methods 0.000 abstract description 5
- 230000008569 process Effects 0.000 description 13
- 230000007246 mechanism Effects 0.000 description 11
- 238000010586 diagram Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 230000008901 benefit Effects 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 5
- 238000004590 computer program Methods 0.000 description 5
- 238000012545 processing Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 235000013399 edible fruits Nutrition 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000001755 vocal effect Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000001035 drying Methods 0.000 description 1
- 239000011521 glass Substances 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000013518 transcription Methods 0.000 description 1
- 230000035897 transcription Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/279—Recognition of textual entities
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/22—Procedures used during a speech recognition process, e.g. man-machine dialogue
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/26—Speech to text systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Theoretical Computer Science (AREA)
- Human Computer Interaction (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Machine Translation (AREA)
Abstract
The present invention provides a kind of intention assessment model generating method, intension recognizing method and devices, belong to field of computer technology.Specifically, each sample text can be inputted initial intention assessment model by the generation method, and according to the sample areas vector in the corresponding each sample text region of each sample text, sample areas phonetic vector and sample areas label vector, determine the prediction intention value of each sample text, prediction intention value and true intention value based on sample text calculate penalty values, when penalty values within a preset range, using initial intention assessment model as intention assessment model, in this way, when carrying out intention assessment using the intention assessment model, it can be based on the region vector of statement text to be sorted, region phonetic vector and area label vector are identified, since region phonetic vector sum area label vector can reduce ambiguity deviation caused by accuracy existing for the word and domanial words of identification mistake, and then the standard of intention assessment can be improved True rate.
Description
Technical field
The invention belongs to field of computer technology, more particularly to a kind of intention assessment model generating method, intention assessment
Method and device.
Background technique
With the fast development of computer technology, novel product of the intelligent assistant as a kind of combination machine intelligence and dialogue
Form, by user be intended to understand based on, help user complete particular task, for user bring hommization service experience and
Convenient for servicing.
Currently, intelligent assistant based on man machine language's interactive mode, in the prior art, usually obtains in sample text
Term vector and context term vector are trained, and obtain intention assessment model, in this way, when carrying out intention assessment, Ke Yili
With the term vector and context term vector of each word in the intention assessment model extraction of the training statement text to be sorted
Matrix, then, term vector and context term vector matrix based on each word of extraction determine that user inputs corresponding meaning
Figure, and then a series of behaviors and strategy are generated and execute, realize the interaction with user.But statement text to be sorted is often
What speech recognition obtained is carried out to the voice of user's input, since the precision of speech recognition is limited, in this way, statement text to be sorted
In it is possible that wrong word, simultaneously, during actual use, the word that will appear ambiguity may be understood in text, in turn
It will lead to and be not inconsistent using the intention of intention assessment model identification and the practical intention of user of the training, it is intended that the accuracy rate of identification
It is lower.
Summary of the invention
The present invention provides a kind of intention assessment model generating method, intension recognizing method and device, to solve to be intended to know
The lower problem of other accuracy rate.
According to the present invention in a first aspect, provide a kind of intention assessment model generating method, this method comprises:
Each sample text that training sample is concentrated inputs initial intention assessment model;
According to sample areas vector, the sample areas phonetic in the corresponding each sample text region of each sample text
Vector and sample areas label vector, and the initial intention assessment model is utilized, determine the pre- of each sample text
Survey intention value;
The true intention value of prediction intention value and each sample text based on each sample text, described in calculating
The penalty values of initial intention assessment model;
When the penalty values within a preset range, using the initial intention assessment model as intention assessment model.
Optionally, after the penalty values for calculating the initial intention assessment model, the method also includes:
When the penalty values are not in the preset range, the parameter of the initial intention assessment model is adjusted, and is based on
The training sample set continues to train to initial intention assessment model adjusted.
Optionally, each sample text that training sample is concentrated inputs before initial intention assessment model, described
Method further include:
The parameter of each layer in initial intention assessment model is initialized;Wherein, the initial intention assessment model packet
Include region vector aligned layer, region vector attention layer, full articulamentum and maximum value softmax layers soft;
The parameter of the region vector aligned layer includes at least: the term vector of each sample word in each sample text,
The context term vector matrix of phonetic term vector and context term vector matrix and each entity tag.
Optionally, the sample areas vector according to the corresponding each sample text region of each sample text,
Sample areas phonetic vector and sample areas label vector, and the initial intention assessment model is utilized, it determines each described
The prediction intention value of sample text, comprising:
For each sample text, which is inputted into the region vector aligned layer, is based on the region
Vector aligned layer divides the sample text, obtains at least one corresponding sample text region of the sample text;
For each sample text, extract the sample areas in the corresponding each sample text region of the sample text to
Amount, sample areas phonetic vector and sample areas label vector;
Sample areas vector, sample areas based on the corresponding each sample text region of each sample text
Phonetic vector and sample areas label vector determine the prediction intention value of each sample text.
Optionally, described for each sample text, extract the corresponding each sample text region of the sample text
Sample areas vector, sample areas phonetic vector and sample areas label vector, comprising:
Term vector by the region vector aligned layer based on each sample word in each sample text region,
Calculate the sample areas vector in each sample text region;
Pinyin word by the region vector aligned layer based on each sample word in each sample text region
Vector calculates the sample areas phonetic vector in each sample text region;
The sample word for belonging to target domain in each sample text region is determined by the region vector aligned layer
The corresponding entity tag of language, and based on the corresponding entity of sample word for belonging to target domain in each sample text region
Label calculates the sample areas label vector in each sample text region.
Optionally, described that each sample word in each sample text region is based on by the region vector aligned layer
The term vector of language calculates the sample areas vector in each sample text region, comprising:
For each sample text region, by the region vector aligned layer by the sample text region interposition
The sample word set determines the term vector of the core word and the context term vector square of the core word as core word
Battle array;
By the context term vector matrix and each sample in text filed is abutted by the region vector aligned layer
The term vector of word carries out the first dot product operation, and carries out maximum pond to the result of first dot product operation, obtains the sample
This text filed region vector.
Optionally, described that each sample word in each sample text region is based on by the region vector aligned layer
The phonetic term vector of language calculates the sample areas phonetic vector in each sample text region, comprising:
For each sample text region, by the region vector aligned layer by the sample text region interposition
The sample word set determines the phonetic term vector of the core word and the context pinyin word of the core word as core word
Vector matrix;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in
The phonetic term vector of each sample word carries out the second dot product operation, and carries out maximum pond to the result of second dot product operation
Change, obtains the region phonetic vector in the sample text region.
Optionally, described based on the corresponding entity of sample word for belonging to target domain in each sample text region
Label calculates the sample areas label vector in each sample text region, comprising:
For each sample text region, determined by the region vector aligned layer every in the sample text region
The context term vector matrix of a entity tag;
By the context term vector matrix of each entity tag and text is abutted by the region vector aligned layer
The term vector of each sample word carries out the operation of third dot product in region, and carries out to the result of third dot product operation maximum
Chi Hua obtains the corresponding maximum pond result of each entity tag;
The average value of the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer,
Obtain the area label term vector in the sample text region.
Second aspect according to the present invention provides a kind of intension recognizing method, this method comprises:
According to the input information of user, statement text to be sorted is determined;
By the statement text input intention assessment model to be sorted, by the intention assessment model to described to be sorted
It states text and carries out intention assessment, obtain the intention classification of the statement text to be sorted;Wherein, the intention assessment model is
It is generated using method described in first aspect.
Optionally, described that intention assessment is carried out to the statement text to be sorted by the intention assessment model, it obtains
The intention classification of the statement text to be sorted, comprising:
The statement text to be sorted is divided by the intention assessment model, obtains the statement text to be sorted
This is corresponding, and at least one is text filed;
By the corresponding each text filed region of statement to be sorted text described in the intention assessment model extraction to
Amount, region phonetic vector and area label vector;
By the intention assessment model based on each text filed region vector, region phonetic vector and area
Domain label vector determines the intention classification of the statement text to be sorted.
Optionally, the intention assessment model includes region vector aligned layer;
It is described to pass through the corresponding each text filed area of statement to be sorted text described in the intention assessment model extraction
Domain vector, region phonetic vector and area label vector, comprising:
By the region vector aligned layer based on it is each it is described it is text filed in each word term vector, calculate each
The text filed region vector;
By the region vector aligned layer based on it is each it is described it is text filed in each word phonetic term vector, calculate
Each text filed region phonetic vector;
By the region vector aligned layer determine it is each it is described it is text filed in belong to the domanial words pair of target domain
The entity tag answered, and based on it is each it is described it is text filed in the corresponding entity tag of each domanial words, calculate each described
Text filed area label vector.
Optionally, the intention assessment model further includes region vector attention layer, full articulamentum and softmax layers;
It is described by the intention assessment model based on each text filed region vector, region phonetic vector with
And area label vector, determine the intention classification of the statement text to be sorted, comprising:
By the region vector aligned layer by each text filed region vector, region phonetic vector and area
Domain label vector is spliced, and each text filed final area vector is obtained;
The weight of each text filed final area vector is determined by the region vector attention layer;
By the full articulamentum based on each text filed the final area vector and its weight, determine it is described to
The class vector of classification statement text;
The intention classification of the statement text to be sorted is determined based on the class vector by described softmax layers.
Optionally, described that the statement text to be sorted is divided by the intention assessment model, it obtains described
Statement text to be sorted is corresponding, and at least one is text filed, comprising:
Participle operation is carried out to the statement text to be sorted by the region vector aligned layer, is obtained described to be sorted
The m word that statement text includes;
I-th to the N+i word in the m word is divided into a text by the region vector aligned layer
Region, obtaining the statement text to be sorted, corresponding at least one is text filed;
Wherein, the i=1...m-N, the N and the m are the positive integer greater than 1.
Optionally, it is described by the region vector aligned layer based on it is each it is described it is text filed in each word word to
Amount calculates each text filed region vector, comprising:
For each described text filed, by the region vector aligned layer by the word in this article one's respective area middle position
As core word, the term vector of the core word and the context term vector matrix of the core word are determined;
By the context term vector matrix and each word in text filed is abutted by the region vector aligned layer
Term vector carry out the operation of the 4th dot product, and maximum pond is carried out to the result of the 4th dot product operation, obtains this article local area
The region vector in domain.
Optionally, it is described by the region vector aligned layer based on it is each it is described it is text filed in each word phonetic
Term vector calculates each text filed region phonetic vector, comprising:
For each described text filed, by the region vector aligned layer by the word in this article one's respective area middle position
As core word, the phonetic term vector of the core word and the context pinyin word vector matrix of the core word are determined;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in
The phonetic term vector of each word carries out the operation of the 5th dot product, and carries out maximum pond to the result of the 5th dot product operation,
Obtain the region phonetic vector of this article one's respective area.
Optionally, it is described by the region vector aligned layer determine it is each it is described it is text filed in belong to target domain
The corresponding entity tag of domanial words, comprising:
For each described text filed, preset target domain dictionary is based on by the region vector aligned layer, it will
The word for belonging to the target domain dictionary in the word that this article one's respective area includes is determined as domanial words;
The corresponding label of each domanial words is searched in the target domain dictionary by the region vector aligned layer,
Obtain the entity tag of each domanial words.
Optionally, it is described based on it is each it is described it is text filed in the corresponding entity tag of each domanial words, calculate each
The text filed area label vector, comprising:
For each described text filed, each entity mark in this article one's respective area is determined by the region vector aligned layer
The context term vector matrix of label;
By the context term vector matrix of each entity tag and text is abutted by the region vector aligned layer
Each term vector carries out the operation of the 6th dot product in region, and carries out maximum pond to the result of the 6th dot product operation, obtains
The corresponding maximum pond result of each entity tag;
The average value of the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer,
Obtain the area label term vector of this article one's respective area.
Optionally, the input information according to user, determines statement text to be sorted, comprising:
When the input information is voice messaging, the voice messaging is converted into text, and by the text and
Formerly the corresponding text of input information is determined as statement text to be sorted;
Wherein, formerly input information is received in the preset time period before receiving the input information.
The third aspect according to the present invention, provides a kind of intention assessment model generating means, which includes:
Input module, each sample text for concentrating training sample input initial intention assessment model;
First determining module, for the sample areas according to the corresponding each sample text region of each sample text
Vector, sample areas phonetic vector and sample areas label vector, and the initial intention assessment model is utilized, it determines each
The prediction intention value of the sample text;
Computing module, the true meaning for prediction intention value and each sample text based on each sample text
Map values calculate the penalty values of the initial intention assessment model;
Second determining module, for when the penalty values within a preset range, using the initial intention assessment model as
Intention assessment model.
Optionally, described device further include:
Module is adjusted, for working as the penalty values not in the preset range, adjusts the initial intention assessment model
Parameter, and initial intention assessment model adjusted is continued to train based on the training sample set.
Optionally, described device further include:
Initialization module is initialized for the parameter to each layer in initial intention assessment model;Wherein, described initial
Intention assessment model includes region vector aligned layer, region vector attention layer, full articulamentum and soft maximum value softmax
Layer;
The parameter of the region vector aligned layer includes at least: the term vector of each sample word in each sample text,
The context term vector matrix of phonetic term vector and context term vector matrix and each entity tag.
Optionally, first determining module, comprising:
First divides submodule, for for each sample text, which to be inputted the region vector
Aligned layer divides the sample text based on the region vector aligned layer, obtains the sample text corresponding at least one
A sample text region;
First extracting sub-module, for extracting the corresponding each sample of the sample text for each sample text
Text filed sample areas vector, sample areas phonetic vector and sample areas label vector;
First determines submodule, for the sample based on the corresponding each sample text region of each sample text
One's respective area vector, sample areas phonetic vector and sample areas label vector determine the prediction meaning of each sample text
Map values.
Optionally, first extracting sub-module, comprising:
First computing unit, it is each in each sample text region for being based on by the region vector aligned layer
The term vector of sample word calculates the sample areas vector in each sample text region;
Second computing unit, it is each in each sample text region for being based on by the region vector aligned layer
The phonetic term vector of sample word calculates the sample areas phonetic vector in each sample text region;
Third computing unit belongs to for being determined in each sample text region by the region vector aligned layer
The corresponding entity tag of sample word of target domain, and based on the sample for belonging to target domain in each sample text region
The corresponding entity tag of this word calculates the sample areas label vector in each sample text region.
Optionally, first computing unit, is used for:
For each sample text region, by the region vector aligned layer by the sample text region interposition
The sample word set determines the term vector of the core word and the context term vector square of the core word as core word
Battle array;
By the context term vector matrix and each sample in text filed is abutted by the region vector aligned layer
The term vector of word carries out the first dot product operation, and carries out maximum pond to the result of first dot product operation, obtains the sample
This text filed region vector.
Optionally, second computing unit, is used for:
For each sample text region, by the region vector aligned layer by the sample text region interposition
The sample word set determines the phonetic term vector of the core word and the context pinyin word of the core word as core word
Vector matrix;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in
The phonetic term vector of each sample word carries out the second dot product operation, and carries out maximum pond to the result of second dot product operation
Change, obtains the region phonetic vector in the sample text region.
Optionally, the third computing unit, is used for:
For each sample text region, determined by the region vector aligned layer every in the sample text region
The context term vector matrix of a entity tag;
By the context term vector matrix of each entity tag and text is abutted by the region vector aligned layer
The term vector of each sample word carries out the operation of third dot product in region, and carries out to the result of third dot product operation maximum
Chi Hua obtains the corresponding maximum pond result of each entity tag;
The average value of the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer,
Obtain the area label term vector in the sample text region.
Fourth aspect according to the present invention, provides a kind of intention assessment device, which includes:
Third determining module determines statement text to be sorted for the input information according to user;
Identification module, for passing through the intention assessment mould for the statement text input intention assessment model to be sorted
Type carries out intention assessment to the statement text to be sorted, obtains the intention classification of the statement text to be sorted;Wherein, described
Intention assessment model is generated using device described in the third aspect.
Optionally, the identification module, comprising:
Second divides submodule, for being divided by the intention assessment model to the statement text to be sorted,
Obtaining the statement text to be sorted, corresponding at least one is text filed;
Second extracting sub-module, for corresponding every by statement text to be sorted described in the intention assessment model extraction
A text filed region vector, region phonetic vector and area label vector;
Second determine submodule, for by the intention assessment model based on each text filed region to
Amount, region phonetic vector and area label vector determine the intention classification of the statement text to be sorted.
Optionally, the intention assessment model includes region vector aligned layer;
Second extracting sub-module, comprising:
4th computing unit, for by the region vector aligned layer be based on it is each it is described it is text filed in each word
Term vector, calculate each text filed region vector;
5th computing unit, for by the region vector aligned layer be based on it is each it is described it is text filed in each word
Phonetic term vector, calculate each text filed region phonetic vector;
6th computing unit, for by the region vector aligned layer determine it is each it is described it is text filed in belong to target
The corresponding entity tag of the domanial words in field, and based on it is each it is described it is text filed in the corresponding entity mark of each domanial words
Label calculate each text filed area label vector.
Optionally, the intention assessment model further includes region vector attention layer, full articulamentum and softmax layers;
Described second determines submodule, is used for:
By the region vector aligned layer by each text filed region vector, region phonetic vector and area
Domain label vector is spliced, and each text filed final area vector is obtained;
The weight of each text filed final area vector is determined by the region vector attention layer;
By the full articulamentum based on each text filed the final area vector and its weight, determine it is described to
The class vector of classification statement text;
The intention classification of the statement text to be sorted is determined based on the class vector by described softmax layers.
Optionally, described second submodule is divided, is used for:
Participle operation is carried out to the statement text to be sorted by the region vector aligned layer, is obtained described to be sorted
The m word that statement text includes;
I-th to the N+i word in the m word is divided into a text by the region vector aligned layer
Region, obtaining the statement text to be sorted, corresponding at least one is text filed;
Wherein, the i=1...m-N, the N and the m are the positive integer greater than 1.
Optionally, the 4th computing unit, is used for:
For each described text filed, by the region vector aligned layer by the word in this article one's respective area middle position
As core word, the term vector of the core word and the context term vector matrix of the core word are determined;
By the context term vector matrix and each word in text filed is abutted by the region vector aligned layer
Term vector carry out the operation of the 4th dot product, and maximum pond is carried out to the result of the 4th dot product operation, obtains this article local area
The region vector in domain.
Optionally, the 5th computing unit, is used for:
For each described text filed, by the region vector aligned layer by the word in this article one's respective area middle position
As core word, the phonetic term vector of the core word and the context pinyin word vector matrix of the core word are determined;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in
The phonetic term vector of each word carries out the operation of the 5th dot product, and carries out maximum pond to the result of the 5th dot product operation,
Obtain the region phonetic vector of this article one's respective area.
Optionally, the 6th computing unit, is used for:
For each described text filed, preset target domain dictionary is based on by the region vector aligned layer, it will
The word for belonging to the target domain dictionary in the word that this article one's respective area includes is determined as domanial words;
The corresponding label of each domanial words is searched in the target domain dictionary by the region vector aligned layer,
Obtain the entity tag of each domanial words.
Optionally, the 6th computing unit, is used for:
For each described text filed, each entity mark in this article one's respective area is determined by the region vector aligned layer
The context term vector matrix of label;
By the context term vector matrix of each entity tag and text is abutted by the region vector aligned layer
Each term vector carries out the operation of the 6th dot product in region, and carries out maximum pond to the result of the 6th dot product operation, obtains
The corresponding maximum pond result of each entity tag;
The average value of the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer,
Obtain the area label term vector of this article one's respective area.
Optionally, the third determining module, is used for:
When the input information is voice messaging, the voice messaging is converted into text, and by the text and
Formerly the corresponding text of input information is determined as statement text to be sorted;
Wherein, formerly input information is received in the preset time period before receiving the input information.
For first technology, the present invention has following advantage: each sample text that training sample can be concentrated inputs
Initial intention assessment model, then, according to sample areas vector, the sample in the corresponding each sample text region of each sample text
One's respective area phonetic vector and sample areas label vector, and initial intention assessment model is utilized, determine each sample text
It predicts intention value, then the prediction intention value based on each sample text and the true intention value of each sample text, calculates
The penalty values of initial intention assessment model, when penalty values within a preset range, can be using initial intention assessment model as intention
Identification model.In the embodiment of the present invention, by extract the corresponding each sample text region of sample text sample areas vector,
Sample areas phonetic vector and sample areas label vector generate intention assessment model, so that in subsequent use process, benefit
It, can be literary based on the statement to be sorted with the intention assessment model of the generation when treating classification statement text and carrying out intention assessment
This region vector, region phonetic vector and area label vector are identified, since region phonetic vector can reduce
Occur word deviation caused by classification accuracy of identification mistake in statement text to be sorted, it can be with by area label vector
To domanial words, there are ambiguities to lead to classification results and the practical deviation for being intended to occur of user for reduction, and then intention can be improved and know
Other accuracy rate.
The above description is only an overview of the technical scheme of the present invention, in order to better understand the technical means of the present invention,
And it can be implemented in accordance with the contents of the specification, and in order to allow above and other objects of the present invention, feature and advantage can
It is clearer and more comprehensible, the followings are specific embodiments of the present invention.
Detailed description of the invention
By reading the following detailed description of the preferred embodiment, various other advantages and benefits are common for this field
Technical staff will become clear.The drawings are only for the purpose of illustrating a preferred embodiment, and is not considered as to the present invention
Limitation.And throughout the drawings, the same reference numbers will be used to refer to the same parts.In the accompanying drawings:
Fig. 1 is a kind of step flow chart of intention assessment model generating method provided in an embodiment of the present invention;
Fig. 2 is the step flow chart of another intention assessment model generating method provided in an embodiment of the present invention;
Fig. 3 is a kind of step flow chart of intension recognizing method provided in an embodiment of the present invention;
Fig. 4-1 is the step flow chart of another intension recognizing method provided in an embodiment of the present invention;
Fig. 4-2 is that the embodiment of the present invention provides a kind of structural schematic diagram of intention assessment model;
Fig. 5 is the step flow chart of another intension recognizing method provided in an embodiment of the present invention;
Fig. 6 is a kind of block diagram of intention assessment model generating means provided in an embodiment of the present invention;
Fig. 7 is a kind of block diagram of intention assessment device provided in an embodiment of the present invention.
Specific embodiment
The exemplary embodiment that the present invention will be described in more detail below with reference to accompanying drawings.Although showing the present invention in attached drawing
Exemplary embodiment, it being understood, however, that may be realized in various forms the present invention without should be by embodiments set forth here
It is limited.It is to be able to thoroughly understand the present invention on the contrary, providing these embodiments, and can be by the scope of the present invention
It is fully disclosed to those skilled in the art.
Fig. 1 is a kind of step flow chart of intention assessment model generating method provided in an embodiment of the present invention, such as Fig. 1 institute
Show, this method may include:
Step 101, each sample text for concentrating training sample input initial intention assessment model.
In the embodiment of the present invention, it is preparatory according to actual needs that the sample text which concentrates can be developer
It chooses, exemplary, when which can be user's progress voice control, the sentence that may be used, the present invention is implemented
Example is not construed as limiting this.Further, which it is preparatory based on neural network model to can be developer
Building, which may include multilayered structure, and different processing may be implemented in every layer of structure.
Step 102, according to sample areas vector, the sample in the corresponding each sample text region of each sample text
Region phonetic vector and sample areas label vector, and the initial intention assessment model is utilized, determine each sample
The prediction intention value of text.
In the embodiment of the present invention, it can use initial intention assessment model and the sample text of input handled, specifically
, initial intention assessment model can first obtain the corresponding sample text region of each sample text, then determine each sample
Sample areas vector, sample areas phonetic vector and the sample areas label in the corresponding each sample text region of text to
Amount, wherein sample areas vector can be based on the term vector determination in sample text region, and sample phonetic region vector can
Being determined based on the phonetic term vector in sample text region, sample areas label vector be can be based on sample text
Belong to the corresponding entity tag determination of sample word of target domain in region, then, each sample text pair can be based on
Sample areas vector, sample areas phonetic vector and the sample areas label vector answered, determine the prediction of each sample text
Intention value.
In this way, being spelled by the sample areas vector in the corresponding each sample text region of extraction sample text, sample areas
Sound vector and sample areas label vector generate intention assessment model, in subsequent use process, utilize the intention of the generation
Identification model treat classification statement text carry out intention assessment when, can based on this it is to be sorted statement text region vector,
Region phonetic vector and area label vector are identified, since region phonetic vector can reduce statement text to be sorted
The middle word deviation caused by classification accuracy for identification mistake occur, can be reduced by area label vector to domanial words
There are ambiguities to lead to classification results and the practical deviation for being intended to occur of user, it is thus possible to improve the accuracy rate of intention assessment
The true intention value of step 103, the prediction intention value based on each sample text and each sample text,
Calculate the penalty values of the initial intention assessment model.
In the embodiment of the present invention, the true intention value of each sample text can be and predefine when collecting sample text
, which can indicate the extent of deviation between the prediction intention value of sample text and the true intention value of sample text.
Step 104, when the penalty values within a preset range, using initial intention assessment model as intention assessment model.
In the embodiment of the present invention, which can set according to practical application scene and actual demand, and the present invention is real
It is without restriction to its to apply example.Further, if penalty values within a preset range, it may be considered that each sample text is pre-
The deviation surveyed between intention value and its true intention value is very small, at this point it is possible to think the prediction intention value of sample text with
Its true intention value is consistent, and the true intention which can correctly predict text correspondingly can
Using by the initial intention assessment model as intention assessment model.
In conclusion a kind of intention assessment model generating method provided in an embodiment of the present invention, it can be by training sample set
In each sample text input initial intention assessment model, then, according to the corresponding each sample text of each sample text
Sample areas vector, sample areas phonetic vector and the sample areas label vector in region, and utilize initial intention assessment mould
Type determines the prediction intention value of each sample text, then the prediction intention value based on each sample text and each sample
The true intention value of text calculates the penalty values of initial intention assessment model, when penalty values within a preset range, will can initially anticipate
Figure identification model is as intention assessment model.In the embodiment of the present invention, by extracting the corresponding each sample text of sample text
Sample areas vector, sample areas phonetic vector and the sample areas label vector in region generates intention assessment model, makes
It obtains in subsequent use process, using the intention assessment model of the generation when treating classification statement text progress intention assessment, energy
It is enough to be identified based on region vector, region phonetic vector and the area label vector of the statement text to be sorted, due to
Region phonetic vector, which can reduce in statement text to be sorted, there is word deviation caused by classification accuracy of identification mistake,
By area label vector can reduce to domanial words there are ambiguity cause classification results and user it is practical be intended to occur it is inclined
Difference, and then the accuracy rate of intention assessment can be improved.
Fig. 2 is the step flow chart of another intention assessment model generating method provided in an embodiment of the present invention, such as Fig. 2 institute
Show, this method may include:
Step 201 initializes the parameter of each layer in initial intention assessment model.
In this step, initial intention assessment model may include multilayered structure, which may include
Region vector aligned layer, region vector attention layer, full articulamentum and softmax layers.Further, the parameter of each layer can be with
Be the realization according to needed for every layer function it is pre-selected, specifically, for region vector aligned layer, the parameter of this layer is at least wrapped
It includes: the term vector of each sample word, phonetic term vector and context term vector matrix in each sample text, and, often
The context term vector matrix of a entity tag, wherein the sample word comprising corresponding different entities label in sample word.Tool
Body, in initialization, it can be and first generate initial value of the number as parameter at random, or being based on experience is each ginseng
Number setting initial value, for example, be directed to the context term vector matrix of sample word, can in conjunction with the word it is common indicate up and down
Literary context sets initial value, and the embodiment of the present invention is not construed as limiting this.
Step 202, each sample text for concentrating training sample input initial intention assessment model.
Specifically, the implementation of this step can refer to above-mentioned steps 101, the embodiment of the present invention is not construed as limiting this.
Step 203, according to sample areas vector, the sample in the corresponding each sample text region of each sample text
Region phonetic vector and sample areas label vector, and the initial intention assessment model is utilized, determine each sample
The prediction intention value of text.
Specifically, this step can be realized by 2031~step 2033 of following step:
The sample text input region vector aligned layer is based on each sample text by step 2031
The region vector aligned layer divides the sample text, obtains at least one corresponding sample text area of the sample text
Domain.
In this step, participle operation first can be carried out to the sample text, obtain the p sample word that the sample text includes
Language, wherein word segmentation processing refers to the process that continuous word sequence is reassembled into word sequence according to certain specification.Right
Sample text carry out participle operation when, can using the segmenting method based on string matching, the segmenting method based on understanding and
Segmenting method, etc. based on statistics, the embodiment of the present invention is not construed as limiting this.It then, can will be in the p sample word
Xth to the Q+x sample word is divided into a sample text region, obtains corresponding at least one sample text of the sample text
One's respective area.Wherein, x=1...n-Q, n and Q are the positive integer greater than 1.Wherein, the occurrence of Q can be according to practical need
Ask preset, the embodiment of the present invention is not construed as limiting this.
Step 2032, for each sample text, extract the corresponding each sample text region of the sample text
Sample areas vector, sample areas phonetic vector and sample areas label vector.
Specifically, this step can be realized by following sub-steps (1)~sub-step (3):
Sub-step (1): word by region vector aligned layer based on each sample word in each sample text region to
Amount, calculates the sample areas vector in each sample text region.
It, can will be in the sample text region by region vector aligned layer specifically, for each sample text region
Between position sample word as core word, determine the term vector of the core word and the context term vector square of the core word
Battle array.Specifically, when determining the sample word in sample text region middle position, the sample that can include in sample text region
When this word number is odd number, determine that (n+1)/2 sample word in the sample text region is the sample word in middle position,
When the sample word number for including in sample text region is even number, the n-th/2 sample word in the sample text region is determined
Or n-th/2+1 sample word is the sample word in middle position.Further, in the term vector and core for determining core word
When the context term vector matrix of heart word, the core word can be searched in the preset parameter of region vector aligned layer
Term vector and context term vector matrix, it is then possible to pass through region vector aligned layer for context term vector matrix and neighbour
The term vector for connecing each word in sample text region carries out the first dot product operation, and carries out to the result of first dot product operation
Maximum pond, that is, take the maximum point of local acceptance region intermediate value, obtain the region vector in the sample text region.
Sub-step (2): each sample word in each sample text region is based on by the region vector aligned layer
The phonetic term vector of language calculates the sample areas phonetic vector in each sample text region.
It, can will be in the sample text region by region vector aligned layer specifically, for each sample text region
Between position sample word as core word, determine the phonetic term vector of the core word and the context pinyin word of the core word
Vector matrix.It correspondingly, can in the context pinyin word vector matrix of the phonetic term vector and core word that determine core word
With searched in the preset parameter of region vector aligned layer the core word phonetic term vector and context pinyin word to
Moment matrix, it is then possible to by context pinyin word vector matrix and be abutted each in text filed by region vector aligned layer
The phonetic term vector of sample word carries out the second dot product operation, and carries out maximum pond to the result of the second dot product operation, that is, takes
The local maximum point of acceptance region intermediate value, obtains the region phonetic vector in the sample text region.
Sub-step (3): it is determined by the region vector aligned layer and belongs to target neck in each sample text region
The corresponding entity tag of sample word in domain, and based on the sample word for belonging to target domain in each sample text region
Corresponding entity tag calculates the sample areas label vector in each sample text region.
Specifically, can determine the sample text region by region vector aligned layer for each sample text region
In each entity tag context term vector matrix, then, by region vector aligned layer will each entity tag up and down
Cliction vector matrix and the term vector for abutting each sample word in text filed carry out the operation of third dot product, and to third dot product
The result of operation carries out maximum pond, obtain the corresponding maximum pond of each entity tag as a result, finally, pass through region to
Amount aligned layer calculates the average value of the corresponding maximum pond result of each entity tag, obtains the region mark in the sample text region
Sign term vector.Wherein, the first dot product operation above-mentioned, the operation of the second dot product and the operation of third dot product are referred to the prior art
The mode of middle vector dot carries out, each sample word in the context term vector matrix of entity tag, adjacent sample text region
The term vector and phonetic term vector of language can be to be obtained from the preset parameter of region vector aligned layer, the present invention
This will not be repeated here for embodiment.
Step 2033, the sample areas based on the corresponding each sample text region of each sample text to
Amount, sample areas phonetic vector and sample areas label vector, determine the prediction intention value of each sample text.
Specifically, region vector aligned layer can be first passed through by the sample areas in each sample text region in this step
Vector, sample areas phonetic vector and sample areas label vector are spliced, and the sample in each sample text region is obtained
Then final area vector determines the sample final area vector in each sample text region by region vector attention layer
Weight, specifically, region vector attention layer can use attention mechanism algorithm, according to sample final area vector sum sample
Relevance between text determines each sample final area vector weight.Then, each sample can be based on by full articulamentum
This text filed sample final area vector and its weight determines the sample classification vector of each sample text, finally, passing through
The softmax layers of prediction intention value for determining sample text based on sample classification vector, wherein the prediction intention value can be sample
It is intended to the corresponding value of classification belonging to text, specifically, softmax layers can calculate the sample text for each sample text
The distance between corresponding sample classification vector vector corresponding with each intention classification, apart from smaller, it may be considered that the sample
The probability that this text belongs to the intention classification is bigger, and the intention classification which can finally be corresponded to maximum probability is corresponding
Value, the prediction intention value as the sample text.
The true intention value of step 204, the prediction intention value based on each sample text and each sample sentence,
Calculate the penalty values of the initial intention assessment model.
Specifically, the implementation of this step can refer to above-mentioned steps 104, the embodiment of the present invention is not construed as limiting this.
In this step, loss function can be used as using entropy function is intersected, by the prediction intention value of each sample text and
True intention value is substituted into the cross entropy loss function and is calculated, using the value being calculated as penalty values, further,
It can be calculated using unknown losses function, the embodiment of the present invention is not construed as limiting this.It is of course also possible to calculate each sample
Euclidean distance between the prediction intention value and true intention value of text, obtains the corresponding penalty values of each sample text, then
The mean value for calculating the corresponding penalty values of each text obtains the penalty values of initial intention assessment model.
Step 205, when the penalty values within a preset range, using the initial intention assessment model as intention assessment mould
Type.
Specifically, the implementation of this step can refer to above-mentioned steps 105, the embodiment of the present invention is not construed as limiting this.
Step 206, when the penalty values are not in the preset range, adjust the ginseng of the initial intention assessment model
Number, and initial intention assessment model adjusted is continued to train based on the training sample set.
In this step, if penalty values are not within a preset range, it may be considered that the prediction intention value of each sample text
Deviation between its true intention value is larger, which can't correctly predict the true of text
Therefore sincere figure can be adjusted according to parameter of the preset step-length to initial intention assessment model, and want to propagate after utilizing to calculate
Method continues to train to initial intention assessment model adjusted, that is, by modifying parameter, is iterated training, makes initially to be intended to
The recognition result of identification model is more nearly legitimate reading, correspondingly, during mostly wheel repetitive exercise, if a certain wheel is first
The penalty values control of beginning intention assessment model within a preset range, then can be using the initial intention assessment model of the wheel as intention
Identification model.
In conclusion another kind intention assessment model generating method provided in an embodiment of the present invention, it can be by training sample
The each sample text concentrated inputs initial intention assessment model, then, according to the corresponding each sample text of each sample text
Sample areas vector, sample areas phonetic vector and the sample areas label vector of one's respective area, and utilize initial intention assessment
Model determines the prediction intention value of each sample text, then the prediction intention value based on each sample text and each sample
The true intention value of this text calculates the penalty values of initial intention assessment model, when penalty values within a preset range, can will be initial
Intention assessment model as intention assessment model, when penalty values not within a preset range, initial intention assessment model can be adjusted
Parameter continues to train.In the embodiment of the present invention, by the sample areas for extracting the corresponding each sample text region of sample text
Vector, sample areas phonetic vector and sample areas label vector generate intention assessment model, so that subsequent use process
In, it, can be to be sorted based on this using the intention assessment model of the generation when treating classification statement text progress intention assessment
Region vector, region phonetic vector and the area label vector of text are stated to be identified, since region phonetic vector can
With reduce it is to be sorted statement text in occur identification mistake word deviation caused by classification accuracy, by area label to
Amount can be reduced to domanial words there are the deviation that ambiguity leads to classification results and the practical intention appearance of user, and then can be improved
The accuracy rate of intention assessment.
Fig. 3 is a kind of step flow chart of intension recognizing method provided in an embodiment of the present invention, as shown in figure 3, this method
May include:
Step 301, the input information according to user, determine statement text to be sorted.
In the embodiment of the present invention, which can be the voice of user's input, for example, user inputs language to client
Sound " search fellow No.9 door ", then the voice " search fellow No.9 door " is input information.It is possible to further extract input voice
Vocal print feature, the vocal print feature for being then based on extraction, which is realized, is converted to text for the voice that user inputs.For example, terminal by pair
Voice " search fellow No.9 door " is converted, and the text being converted to " search fellow No.9 door " can be determined as statement text to be sorted
This.Certainly, which is also possible to the text of user's input, and correspondingly, the text of user input is table to be sorted
State text.Wherein, which can be mounted in the electronics such as terminal, mobile phone, computer, set-top box, television set, intelligent glasses and set
Standby upper client.
Step 302, by the statement text input intention assessment model to be sorted, by the intention assessment model to institute
It states statement text to be sorted and carries out intention assessment, obtain the intention classification of the statement text to be sorted.
In this step, which, which can be, utilizes the generation of aforementioned intention assessment model generating method, due to
The intention assessment model is by extracting the sample areas vector in the corresponding each sample text region of sample text, sample areas
What phonetic vector and sample areas label vector generated, in this way, in this step, using the intention assessment model to be sorted
It, can be based on region vector, region phonetic vector and the area of the statement text to be sorted when stating text progress intention assessment
Domain label vector is identified the word of identification mistake occur since region phonetic vector can reduce in statement text to be sorted
Language deviation caused by classification accuracy can be reduced to cause to classify there are ambiguity to domanial words by area label vector and be tied
Fruit and the practical deviation for being intended to occur of user, and then the accuracy rate of intention assessment can be improved.
In conclusion a kind of intension recognizing method provided in an embodiment of the present invention, it can be according to the input information of user, really
Fixed statement text to be sorted is treated point then by statement text input intention assessment model to be sorted by intention assessment model
Class states text and carries out intention assessment, obtains the intention classification of statement text to be sorted, wherein the intention assessment model is to pass through
Extract sample areas vector, sample areas phonetic vector and the sample areas in the corresponding each sample text region of sample text
Label vector, and the generation of this these vector based on extraction, therefore, in the embodiment of the present invention, utilize the intention assessment model
It, can be based on region vector, the region phonetic of the statement text to be sorted when treating classification statement text progress intention assessment
Vector and area label vector are identified, are known since region phonetic vector can reduce in statement text to be sorted
Not wrong word deviation caused by classification accuracy, can reduce that there are ambiguities to domanial words by area label vector
Lead to classification results and the practical deviation for being intended to occur of user, and then the accuracy rate of intention assessment can be improved.
Fig. 4-1 is the step flow chart of another intension recognizing method provided in an embodiment of the present invention, as shown in Fig. 4-1,
This method may include:
Step 401, the input information according to user, determine statement text to be sorted.
In actual application scenarios, user may repeatedly be interacted with terminal whithin a period of time, and two people are at one section
There may be association between time interior more words of speaking, therefore, in this step, terminal is determining statement text to be sorted
When, can input information be voice messaging when, the voice messaging is first converted into text, then by the text and formerly it is defeated
Enter the corresponding text of information as statement text to be sorted, wherein formerly input information can be receive input information it
It is received in preceding preset time period, which can be presets according to actual needs, and exemplary, this is default
Period can be 1 minute, which may be half a minute, and the embodiment of the present invention is not construed as limiting this.The present invention
In embodiment, pass through the corresponding text of input information that the user received is inputted into the corresponding text of information and is received before
This can make full use of the relevance between language as statement text to be sorted, increase entrained by statement text to be sorted
Effective information, and then improve the accuracy rate that intention assessment is carried out based on the statement text to be sorted.
Step 402 divides the statement text to be sorted by the intention assessment model, obtains described wait divide
Class states text, and corresponding at least one is text filed.
In practical application scene, statement text to be sorted is often the sentence being made of multiple words, and in sentence not
The meaning stated with word is often different, and each word is not also identical to the semantic influence degree of entire sentence, therefore, this
In inventive embodiments, it is text filed statement text to be sorted can be divided at least one, wherein each text filed middle packet
Multiple words are included.In this way, it is multiple text filed by the way that statement text to be sorted to be divided into, so that can be right in subsequent process
It is each it is text filed be respectively processed analysis, and then can more accurately analyze language represented by entire statement text to be sorted
Justice.
Participle operation is carried out specifically, intention assessment model can be first passed through and treat classification statement text, is obtained to be sorted
The m word that statement text includes.In this step, word segmentation processing is referred to continuous word sequence according to certain specification again
It is combined into the process of word sequence.It, can be using based on string matching when treating classification statement text and carrying out participle operation
Segmenting method, the segmenting method based on understanding and segmenting method based on statistics, etc., the embodiment of the present invention does not limit this
It is fixed.
Then, by i-th in m word to the N+i word be divided into one it is text filed, obtain statement to be sorted text
This is corresponding, and at least one is text filed.
In this step, i=1...m-N, m, N are the positive integer greater than 1.Wherein, the occurrence of N can be according to reality
Demand is preset, exemplary, it is assumed that N 2 treats classification chart text and carries out after segmenting operation operation, obtains 5 words
Language: a, b, c, d, e, then can be by the 1st to the 3rd word, that is, word a, b, c are divided into one text filed [a, b, c],
By the 2nd to the 4th word, that is, word b, c, d are divided into one text filed [b, c, d], by the 3rd to the 5th word, that is, word
Language c, d, e are divided into one text filed [c, d, e], obtain 3 it is text filed.
It is step 403, corresponding each text filed by statement text to be sorted described in the intention assessment model extraction
Region vector, region phonetic vector and area label vector.
In this step, which may include region vector aligned layer, correspondingly, can pass through following sub-steps
Suddenly (4)~sub-step (6) is realized:
Sub-step (4): by the region vector aligned layer based on it is each it is described it is text filed in each word word to
Amount calculates each text filed region vector.
Specifically, region vector aligned layer can be first passed through by text filed middle position for each text filed
Word determines the term vector of core word and the context term vector matrix of core word, wherein determining text as core word
When the word in region middle position, when the word number that can include in text filed is odd number, the is determined in this article one's respective area
(n+1)/2 word is that the word in middle position determines this article one's respective area when the word number for including is even number in text filed
In the n-th/2 word or n-th/2+1 word be middle position word, in the term vector and core word for determining core word
Context term vector matrix when, due to training intention assessment model when, it has been determined that go out different terms term vector and
Phonetic term vector, the context term vector matrix of word and context pinyin word vector matrix therefore, can be in this step
The corresponding term vector of each core word and context term vector matrix are searched from the parameter of region vector aligned layer.
Then, each word in text filed by context term vector matrix and can be abutted by region vector aligned layer
Term vector carry out the 4th dot product operation, that is, take the maximum point of local acceptance region intermediate value and to the 4th dot product operation result into
Row maximum pond, obtains the region vector of this article one's respective area.Wherein, it is adjacent it is text filed can be with before this article one's respective area and
Later adjacent text filed, it is adjacent it is text filed in the term vector of each word can be the ginseng from region vector aligned layer
Lookup obtains in number, and the mode that the operation of the 4th dot product is referred to vector dot in the prior art carries out, and the present invention implements to exist
This is not repeated them here.
Sub-step (5): by the region vector aligned layer based on it is each it is described it is text filed in each word phonetic
Term vector calculates each text filed region phonetic vector.
It, can be by region vector aligned layer by this article one's respective area middle position for each text filed in this step
Word as core word, determine the phonetic term vector of core word and the context pinyin word vector matrix of core word, specifically
, the corresponding phonetic term vector of each core word and context pinyin word can be searched from the parameter of region vector aligned layer
Vector matrix.
Then, it by context pinyin word vector matrix and can be abutted each in text filed by region vector aligned layer
The phonetic term vector of word carries out the operation of the 5th dot product, and carries out maximum pond to the result of the 5th dot product operation, obtains this article
The region phonetic vector of one's respective area.Wherein, the phonetic term vector for abutting each word in text filed can be from region vector
Lookup obtains in the parameter of aligned layer, and the mode that the operation of the 5th dot product is referred to vector dot in the prior art carries out,
This will not be repeated here for present invention implementation.
Due to different word may corresponding phonetic it is identical, for example, phonetic of the word " fellow No.9's door " with " wine is laid one's hand on "
It is identical, but the two is to indicate the different terms of different meanings, therefore, there are the words of transcription error in statement text to be sorted
When language, information represented by the word that is carried of region vector that is determined based on the term vector of word cannot be represented accurately
Meaning represented by text filed middle word.And the intention assessment model in the embodiment of the present invention is true using the term vector of word
While determining region vector, the phonetic term vector in region can be also determined using the phonetic term vector of phonetic, since phonetic can
The pronunciation of the accurate voice for indicating user's input, therefore, the region phonetic vector determined based on the phonetic term vector of phonetic is held
Information represented by the phonetic of load can accurately represent meaning represented by text filed middle phonetic, so correct due to
Intention assessment deviation caused by speech recognition.
Sub-step (6): by the region vector aligned layer determine it is each it is described it is text filed in belong to target domain
The corresponding entity tag of domanial words, and based on it is each it is described it is text filed in the corresponding entity tag of each domanial words, meter
Calculate each text filed area label vector.
In the embodiment of the present invention, target domain can indicate field belonging to statement text to be sorted, due to same word
There is different meanings in different fields, can thus exist and understand ambiguity, for example, word " apple " may indicate hand
Machine, it is also possible to indicate fruit.Therefore, in this step, preset target domain dictionary can be based on by region vector aligned layer,
The word for belonging to target domain dictionary in word that this article one's respective area includes is determined as domanial words, then, by region to
Amount aligned layer searches the corresponding label of each domanial words in target domain dictionary, obtains the entity mark of each domanial words
Label.It wherein, may include all words for belonging to target domain and the corresponding entity of each word in the target domain dictionary
Label, wherein entity tag can be used to indicate that actual content represented by the domanial words, exemplary, it is assumed that this article local area
Domanial words in domain are " fellow No.9's door ", and the domanial words corresponding label in target domain dictionary is " TV play ", then
The entity tag of the available domanial words is " TV play ".
It is possible to further determine the cliction up and down of each entity tag in this article one's respective area by region vector aligned layer
Vector matrix by the context term vector matrix of each entity tag and is abutted every in text filed by region vector aligned layer
A term vector carries out the operation of the 6th dot product, and carries out maximum pond to the result of the 6th dot product operation, obtains each entity tag
Corresponding maximum pond is as a result, finally, calculate the corresponding maximum pond result of each entity tag by region vector aligned layer
Average value, obtain the area label term vector of this article one's respective area.Wherein, the context term vector matrix of entity tag can be
It searches and obtains from the parameter of region vector aligned layer, the operation of the 6th dot product is referred to vector dot in the prior art
Mode carries out, and this will not be repeated here for present invention implementation.
In the embodiment of the present invention, the corresponding entity of each domanial words in text filed is based on by region vector aligned layer
Label determines text filed area label vector, since the entity tag of domanial words can accurately define domanial words
Represented actual content eliminates the understanding ambiguity of domanial words, and therefore, the area label vector determined based on entity tag can
With the more accurate meaning represented by text filed middle word that indicates, and then corrects and understand caused by ambiguity since word exists
Intention assessment deviation.
Step 404, by the intention assessment model based on each text filed region vector, region phonetic to
Amount and area label vector determine the intention classification of the statement text to be sorted.
In this step, which can also include region vector attention layer, full articulamentum and softmax
Layer, Fig. 4-2 is that the embodiment of the present invention provides a kind of structural schematic diagram of intention assessment model, as shown in the Fig. 4-2, the intention assessment
Model may include region vector aligned layer, region vector attention layer, full articulamentum and softmax layers in total.Correspondingly,
It can be realized by following sub-steps (7)~sub-step (8):
Sub-step (7): by the region vector aligned layer by each text filed region vector, region phonetic
Vector and area label vector are spliced, and each text filed final area vector is obtained.
Specifically, can be by region vector by text filed region vector, region phonetic vector and area label
Vector Groups are combined into a vector, obtain the final area vector of this article one's respective area.Combination when, region vector, region phonetic to
Amount and the sequence of area label vector can be preset according to actual needs, and the embodiment of the present invention is not construed as limiting this.Show
Example, it is assumed that text filed region vector, region phonetic vector and area label vector is the vector of 100 dimensions, is passed through
Splice the final area vector that available dimension is 300.
Sub-step (8): each text filed final area vector is determined by the region vector attention layer
Weight.
Wherein, region vector attention layer can be using attention mechanism algorithm, wherein attention (Attention)
For the essence of mechanism from human visual attention's mechanism, visual attention mechanism is at brain signal specific to human vision
Reason mechanism, human vision obtain the target area for needing to pay close attention to, that is, general described by quickly scanning global image
' s focus of attention, then to the more attention resources of this regional inputs, with obtain it is more required for concern targets details
Information, and inhibit other garbages.Therefore, attention mechanism algorithm be based on simulation human attention mechanism and establish one
Kind network algorithm, region vector attention layer can be based on attention mechanism, capture final area vector sum statement text to be sorted
Relevance between this, the relevance can be attention weight, by by attention weight distribution to corresponding final area
To get to the weight for applying attention mechanism on vector, since the weight includes final area vector sum table to be sorted
State the relevance between text, therefore, it is subsequent using the weight carry out as classification in application, enable classification results more
It is accurate to add.
Specifically, attention mechanism algorithm used by the vector attention layer of region can be indicated by following formula:
Wherein, αiIndicate the weight of final area vector, usIndicate the corresponding matrix mapping of final area vector, ui=
tanh(Whi+bs) indicate the corresponding context-aware matrix mapping of the final area vector, hiIndicate final area vector, W, bsIt is pre-
First trained parameter.
Sub-step (9): by the full articulamentum based on each text filed the final area vector and its weight,
Determine the class vector of the statement text to be sorted.
In this step, full articulamentum can be calculated all based on each text filed final area vector and its weight
Text filed weighted sum, that is, the sum of products of each text filed final area vector and its weight obtains table to be sorted
State the class vector of text.Exemplary, class vector v can be indicated are as follows:
Sub-step (10): the statement text to be sorted is determined based on the class vector by described softmax layers
It is intended to classification.
It, can be by between softmax layers of calculating class vector vector corresponding with each intention classification in this step
Distance, wherein apart from smaller, it may be considered that the probability that the statement text to be sorted belongs to the intention classification is bigger, finally, can
Intention classification the statement text to be sorted to be corresponded to the intention classification of maximum probability, as the statement text to be sorted.
In conclusion it is provided in an embodiment of the present invention another kind intension recognizing method, can according to the input information of user,
It determines statement text to be sorted, classification statement text is then treated by intention assessment model and is divided, table to be sorted is obtained
State text it is corresponding at least one is text filed, it is then, corresponding every by intention assessment model extraction statement text to be sorted
A text filed region vector, region phonetic vector and area label vector, finally, being based on by intention assessment model every
A text filed region vector, region phonetic vector and area label vector determine the intention class of statement text to be sorted
Not, in the embodiment of the present invention, using the intention assessment model when treating classification statement text progress intention assessment, can be based on should
Region vector, region phonetic vector and the area label vector of statement text to be sorted are identified, due to region phonetic
Vector, which can reduce in statement text to be sorted, there is word deviation caused by classification accuracy of identification mistake, passes through region
Label vector can reduce to domanial words there are ambiguity cause classification results and user it is practical be intended to occur deviation, Jin Erke
To improve the accuracy rate of intention assessment.
Fig. 5 is the step flow chart of another intension recognizing method provided in an embodiment of the present invention, as shown in figure 5, the party
Method may include:
Step 501, each sample text for concentrating training sample input initial intention assessment model.
Specifically, the implementation of this step can refer to above-mentioned steps 101, the embodiment of the present invention is not construed as limiting this.
Step 502, according to sample areas vector, the sample in the corresponding each sample text region of each sample text
Region phonetic vector and sample areas label vector, and the initial intention assessment model is utilized, determine each sample
The prediction intention value of text.
Specifically, the implementation of this step can refer to above-mentioned steps 102, the embodiment of the present invention is not construed as limiting this.
The true intention value of step 503, the prediction intention value based on each sample text and each sample text,
Calculate the penalty values of the initial intention assessment model.
Specifically, the implementation of this step can refer to above-mentioned steps 103, the embodiment of the present invention is not construed as limiting this.
Step 504, when the penalty values within a preset range, using the initial intention assessment model as intention assessment mould
Type.
Specifically, the implementation of this step can refer to above-mentioned steps 104, the embodiment of the present invention is not construed as limiting this.
Step 505, the input information according to user, determine statement text to be sorted.
Specifically, the implementation of this step can refer to above-mentioned steps 301, the embodiment of the present invention is not construed as limiting this.
Step 506, by the statement text input intention assessment model to be sorted, by the intention assessment model to institute
It states statement text to be sorted and carries out intention assessment, obtain the intention classification of the statement text to be sorted.
Specifically, the implementation of this step can refer to above-mentioned steps 302, the embodiment of the present invention is not construed as limiting this.
In conclusion another intension recognizing method provided in an embodiment of the present invention, training sample can be concentrated every
A sample text inputs initial intention assessment model, then, according to the corresponding each sample text region of each sample text
Sample areas vector, sample areas phonetic vector and sample areas label vector, and initial intention assessment model is utilized, it determines
The prediction intention value of each sample text, then the prediction intention value based on each sample text and each sample text is true
Sincere map values calculate the penalty values of initial intention assessment model, when penalty values within a preset range, can be by initial intention assessment mould
Type is as intention assessment model, then, according to the input information of user, determines statement text to be sorted, by statement text to be sorted
This input intention assessment model carries out intention assessment to the statement text to be sorted by the intention assessment model, obtains
The intention classification of statement text to be sorted.In the embodiment of the present invention, classification chart is being treated using the intention assessment model of the generation
It, can be based on region vector, region phonetic vector and the region of the statement text to be sorted when stating text progress intention assessment
Label vector is identified the word of identification mistake occur since region phonetic vector can reduce in statement text to be sorted
The deviation caused by classification accuracy, can reduce that there are ambiguities to lead to classification results to domanial words by area label vector
And the practical deviation for being intended to occur of user, and then the accuracy rate of intention assessment can be improved.
Fig. 6 is a kind of block diagram of intention assessment model generating means provided in an embodiment of the present invention, as shown in fig. 6, the dress
Setting 60 may include:
Input module 601, each sample text for concentrating training sample input initial intention assessment model;
First determining module 602, for the sample according to the corresponding each sample text region of each sample text
Region vector, sample areas phonetic vector and sample areas label vector, and the initial intention assessment model is utilized, it determines
The prediction intention value of each sample text;
Computing module 602, for based on each sample text prediction intention value and each sample text it is true
Sincere map values calculate the penalty values of the initial intention assessment model;
Second determining module 604 makees the initial intention assessment model for working as the penalty values within a preset range
For intention assessment model.
In conclusion a kind of intention assessment model generating means provided in an embodiment of the present invention, it can be by training sample set
In each sample text input initial intention assessment model, then, according to the corresponding each sample text of each sample text
Sample areas vector, sample areas phonetic vector and the sample areas label vector in region, and utilize initial intention assessment mould
Type determines the prediction intention value of each sample text, then the prediction intention value based on each sample text and each sample
The true intention value of text calculates the penalty values of initial intention assessment model, when penalty values within a preset range, will can initially anticipate
Figure identification model is as intention assessment model.In the embodiment of the present invention, by extracting the corresponding each sample text of sample text
Sample areas vector, sample areas phonetic vector and the sample areas label vector in region generates intention assessment model, makes
It obtains in subsequent use process, using the intention assessment model of the generation when treating classification statement text progress intention assessment, energy
It is enough to be identified based on region vector, region phonetic vector and the area label vector of the statement text to be sorted, due to
Region phonetic vector, which can reduce in statement text to be sorted, there is word deviation caused by classification accuracy of identification mistake,
By area label vector can reduce to domanial words there are ambiguity cause classification results and user it is practical be intended to occur it is inclined
Difference, and then the accuracy rate of intention assessment can be improved.
Optionally, described device 60 further include:
Module is adjusted, for working as the penalty values not in the preset range, adjusts the initial intention assessment model
Parameter, and initial intention assessment model adjusted is continued to train based on the training sample set.
Optionally, described device 60 further include:
Initialization module is initialized for the parameter to each layer in initial intention assessment model;Wherein, described initial
Intention assessment model includes region vector aligned layer, region vector attention layer, full articulamentum and soft maximum value softmax
Layer;
The parameter of the region vector aligned layer includes at least: the term vector of each sample word in each sample text,
The context term vector matrix of phonetic term vector and context term vector matrix and each entity tag.
Optionally, first determining module 602, comprising:
First divides submodule, for for each sample text, which to be inputted the region vector
Aligned layer divides the sample text based on the region vector aligned layer, obtains the sample text corresponding at least one
A sample text region;
First extracting sub-module, for extracting the corresponding each sample of the sample text for each sample text
Text filed sample areas vector, sample areas phonetic vector and sample areas label vector;
First determines submodule, for the sample based on the corresponding each sample text region of each sample text
One's respective area vector, sample areas phonetic vector and sample areas label vector determine the prediction meaning of each sample text
Map values.
Optionally, first extracting sub-module, comprising:
First computing unit, it is each in each sample text region for being based on by the region vector aligned layer
The term vector of sample word calculates the sample areas vector in each sample text region;
Second computing unit, it is each in each sample text region for being based on by the region vector aligned layer
The phonetic term vector of sample word calculates the sample areas phonetic vector in each sample text region;
Third computing unit belongs to for being determined in each sample text region by the region vector aligned layer
The corresponding entity tag of sample word of target domain, and based on the sample for belonging to target domain in each sample text region
The corresponding entity tag of this word calculates the sample areas label vector in each sample text region.
Optionally, first computing unit, is used for:
For each sample text region, by the region vector aligned layer by the sample text region interposition
The sample word set determines the term vector of the core word and the context term vector square of the core word as core word
Battle array;
By the context term vector matrix and each sample in text filed is abutted by the region vector aligned layer
The term vector of word carries out the first dot product operation, and carries out maximum pond to the result of first dot product operation, obtains the sample
This text filed region vector.
Optionally, second computing unit, is used for:
For each sample text region, by the region vector aligned layer by the sample text region interposition
The sample word set determines the phonetic term vector of the core word and the context pinyin word of the core word as core word
Vector matrix;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in
The phonetic term vector of each sample word carries out the second dot product operation, and carries out maximum pond to the result of second dot product operation
Change, obtains the region phonetic vector in the sample text region.
Optionally, the third computing unit, is used for:
For each sample text region, determined by the region vector aligned layer every in the sample text region
The context term vector matrix of a entity tag;
By the context term vector matrix of each entity tag and text is abutted by the region vector aligned layer
The term vector of each sample word carries out the operation of third dot product in region, and carries out to the result of third dot product operation maximum
Chi Hua obtains the corresponding maximum pond result of each entity tag;
The average value of the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer,
Obtain the area label term vector in the sample text region.
In conclusion a kind of intention assessment model generating means provided in an embodiment of the present invention, it can be by training sample set
In each sample text input initial intention assessment model, then, according to the corresponding each sample text of each sample text
Sample areas vector, sample areas phonetic vector and the sample areas label vector in region, and utilize initial intention assessment mould
Type determines the prediction intention value of each sample text, then the prediction intention value based on each sample text and each sample
The true intention value of text calculates the penalty values of initial intention assessment model, when penalty values within a preset range, will can initially anticipate
Figure identification model as intention assessment model, when penalty values not within a preset range, the ginseng of initial intention assessment model can be adjusted
Number, continues to train.In the embodiment of the present invention, by extract the sample areas in the corresponding each sample text region of sample text to
Amount, sample areas phonetic vector and sample areas label vector generate intention assessment model, so that in subsequent use process,
Using the intention assessment model of the generation when treating classification statement text progress intention assessment, the statement to be sorted can be based on
Region vector, region phonetic vector and the area label vector of text is identified, since region phonetic vector can drop
There is word deviation caused by classification accuracy of identification mistake in low statement text to be sorted, it can by area label vector
With reduction, to domanial words, there are ambiguities to lead to classification results and the practical deviation for being intended to occur of user, and then intention can be improved
The accuracy rate of identification.
Fig. 7 is a kind of block diagram of intention assessment device provided in an embodiment of the present invention, as shown in fig. 7, the device 70 can be with
Include:
Third determining module 701 determines statement text to be sorted for the input information according to user;
Identification module 702, for passing through the intention assessment for the statement text input intention assessment model to be sorted
Model carries out intention assessment to the statement text to be sorted, obtains the intention classification of the statement text to be sorted;Wherein, institute
Stating intention assessment model is generated using device described in Fig. 6.
In conclusion a kind of intention assessment device provided in an embodiment of the present invention, it can be according to the input information of user, really
Fixed statement text to be sorted is treated point then by statement text input intention assessment model to be sorted by intention assessment model
Class states text and carries out intention assessment, obtains the intention classification of statement text to be sorted, wherein the intention assessment model is to pass through
Extract sample areas vector, sample areas phonetic vector and the sample areas in the corresponding each sample text region of sample text
Label vector, and the generation of this these vector based on extraction, therefore, in the embodiment of the present invention, utilize the intention assessment model
It, can be based on region vector, the region phonetic of the statement text to be sorted when treating classification statement text progress intention assessment
Vector and area label vector are identified, are known since region phonetic vector can reduce in statement text to be sorted
Not wrong word deviation caused by classification accuracy, can reduce that there are ambiguities to domanial words by area label vector
Lead to classification results and the practical deviation for being intended to occur of user, and then the accuracy rate of intention assessment can be improved.
Optionally, the identification module 702, comprising:
Second divides submodule, for being divided by the intention assessment model to the statement text to be sorted,
Obtaining the statement text to be sorted, corresponding at least one is text filed;
Second extracting sub-module, for corresponding every by statement text to be sorted described in the intention assessment model extraction
A text filed region vector, region phonetic vector and area label vector;
Second determine submodule, for by the intention assessment model based on each text filed region to
Amount, region phonetic vector and area label vector determine the intention classification of the statement text to be sorted.
Optionally, the intention assessment model includes region vector aligned layer;
Second extracting sub-module, comprising:
4th computing unit, for by the region vector aligned layer be based on it is each it is described it is text filed in each word
Term vector, calculate each text filed region vector;
5th computing unit, for by the region vector aligned layer be based on it is each it is described it is text filed in each word
Phonetic term vector, calculate each text filed region phonetic vector;
6th computing unit, for by the region vector aligned layer determine it is each it is described it is text filed in belong to target
The corresponding entity tag of the domanial words in field, and based on it is each it is described it is text filed in the corresponding entity mark of each domanial words
Label calculate each text filed area label vector.
Optionally, the intention assessment model further includes region vector attention layer, full articulamentum and softmax layers;
Described second determines submodule, is used for:
By the region vector aligned layer by each text filed region vector, region phonetic vector and area
Domain label vector is spliced, and each text filed final area vector is obtained;
The weight of each text filed final area vector is determined by the region vector attention layer;
By the full articulamentum based on each text filed the final area vector and its weight, determine it is described to
The class vector of classification statement text;
The intention classification of the statement text to be sorted is determined based on the class vector by described softmax layers.
Optionally, described second submodule is divided, is used for:
Participle operation is carried out to the statement text to be sorted by the region vector aligned layer, is obtained described to be sorted
The m word that statement text includes;
I-th to the N+i word in the m word is divided into a text by the region vector aligned layer
Region, obtaining the statement text to be sorted, corresponding at least one is text filed;
Wherein, the i=1...m-N, the N and the m are the positive integer greater than 1.
Optionally, the 4th computing unit, is used for:
For each described text filed, by the region vector aligned layer by the word in this article one's respective area middle position
As core word, the term vector of the core word and the context term vector matrix of the core word are determined;
By the context term vector matrix and each word in text filed is abutted by the region vector aligned layer
Term vector carry out the operation of the 4th dot product, and maximum pond is carried out to the result of the 4th dot product operation, obtains this article local area
The region vector in domain.
Optionally, the 5th computing unit, is used for:
For each described text filed, by the region vector aligned layer by the word in this article one's respective area middle position
As core word, the phonetic term vector of the core word and the context pinyin word vector matrix of the core word are determined;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in
The phonetic term vector of each word carries out the operation of the 5th dot product, and carries out maximum pond to the result of the 5th dot product operation,
Obtain the region phonetic vector of this article one's respective area.
Optionally, the 6th computing unit, is used for:
For each described text filed, preset target domain dictionary is based on by the region vector aligned layer, it will
The word for belonging to the target domain dictionary in the word that this article one's respective area includes is determined as domanial words;
The corresponding label of each domanial words is searched in the target domain dictionary by the region vector aligned layer,
Obtain the entity tag of each domanial words.
Optionally, the 6th computing unit, is used for:
For each described text filed, each entity mark in this article one's respective area is determined by the region vector aligned layer
The context term vector matrix of label;
By the context term vector matrix of each entity tag and text is abutted by the region vector aligned layer
Each term vector carries out the operation of the 6th dot product in region, and carries out maximum pond to the result of the 6th dot product operation, obtains
The corresponding maximum pond result of each entity tag;
The average value of the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer,
Obtain the area label term vector of this article one's respective area.
Optionally, the third determining module 701, is used for:
When the input information is voice messaging, the voice messaging is converted into text, and by the text and
Formerly the corresponding text of input information is determined as statement text to be sorted;
Wherein, formerly input information is received in the preset time period before receiving the input information.
In conclusion a kind of intention assessment device provided in an embodiment of the present invention, it can be according to the input information of user, really
Then fixed statement text to be sorted is treated classification statement text by intention assessment model and is divided, obtains statement to be sorted
Text is corresponding, and at least one is text filed, then, corresponding each by intention assessment model extraction statement text to be sorted
Text filed region vector, region phonetic vector and area label vector, finally, being based on by intention assessment model each
Text filed region vector, region phonetic vector and area label vector determines the intention classification of statement text to be sorted,
It, can be based on should be to using the intention assessment model when treating classification statement text and carrying out intention assessment in the embodiment of the present invention
Region vector, region phonetic vector and the area label vector of classification statement text identified, due to region phonetic to
Amount, which can reduce in statement text to be sorted, there is word deviation caused by classification accuracy of identification mistake, is marked by region
Label vector can reduce that there are the deviations that ambiguity leads to classification results and the practical intention appearance of user to domanial words, and then can be with
Improve the accuracy rate of intention assessment.
For above-mentioned apparatus embodiment, since it is basically similar to the method embodiment, so be described relatively simple,
The relevent part can refer to the partial explaination of embodiments of method.
In addition, the embodiment of the present invention also provides a kind of terminal, including processor, memory, storage is on a memory and can
The computer program run on the processor, the computer program realize above-mentioned intention assessment model when being executed by processor
Generation method, each process of intension recognizing method embodiment, and identical technical effect can be reached, to avoid repeating, here
It repeats no more.
The embodiment of the present invention also provides a kind of computer readable storage medium, and meter is stored on computer readable storage medium
Calculation machine program, the computer program realize above-mentioned intention assessment model generating method, intension recognizing method when being executed by processor
Each process of embodiment, and identical technical effect can be reached, to avoid repeating, which is not described herein again.Wherein, the meter
Calculation machine readable storage medium storing program for executing, such as read-only memory (Read-Only Memory, abbreviation ROM), random access memory (Random
Access Memory, abbreviation RAM), magnetic or disk etc..
All the embodiments in this specification are described in a progressive manner, the highlights of each of the examples are with
The difference of other embodiments, the same or similar parts between the embodiments can be referred to each other.
It would have readily occurred to a person skilled in the art that: any combination application of above-mentioned each embodiment is all feasible, therefore
Any combination between above-mentioned each embodiment is all embodiment of the present invention, but this specification exists as space is limited,
This is not just detailed one by one.
Intention assessment model generating method, intension recognizing method are not with any certain computer, virtually provided herein
System or other equipment are inherently related.Various general-purpose systems can also be used together with teachings based herein.According to above
Description, it is obvious for constructing structure required by the system with the present invention program.In addition, the present invention is not also for any
Certain programmed language.It should be understood that can use various programming languages realizes summary of the invention described herein, and above
The description done to language-specific is in order to disclose the best mode of carrying out the invention.
In the instructions provided here, numerous specific details are set forth.It is to be appreciated, however, that implementation of the invention
Example can be practiced without these specific details.In some instances, well known method, structure is not been shown in detail
And technology, so as not to obscure the understanding of this specification.
It should be appreciated that in order to simplify the present invention and help to understand one or more of the various inventive aspects, it is right above
In the description of exemplary embodiment of the present invention, each feature of the invention be grouped together into sometimes single embodiment, figure or
In person's descriptions thereof.However, the disclosed method should not be interpreted as reflecting the following intention: i.e. claimed hair
It is bright to require features more more than feature expressly recited in each claim.More precisely, such as claims institute
As reflection, inventive aspect is all features less than single embodiment disclosed above.Therefore, it then follows specific embodiment party
Thus claims of formula are expressly incorporated in the specific embodiment, wherein each claim itself is as of the invention
Separate embodiments.
Those skilled in the art will understand that can be carried out adaptively to the module in the equipment in embodiment
Change and they are arranged in one or more devices different from this embodiment.It can be the module or list in embodiment
Member or component are combined into a module or unit or component, and furthermore they can be divided into multiple submodule or subelement or subgroups
Part.Other than such feature and/or at least some of process or unit exclude each other, any combination can be used
To all features disclosed in this specification (including adjoint claim, abstract and attached drawing) and so disclosed any side
All process or units of method or equipment are combined.Unless expressly stated otherwise, this specification (is wanted including adjoint right
Ask, make a summary and attached drawing) disclosed in each feature can be replaced with an alternative feature that provides the same, equivalent, or similar purpose.
It will be appreciated by those of skill in the art that although some embodiments described herein include other embodiments is wrapped
Certain features for including rather than other feature, but the combination of the feature of different embodiments mean in the scope of the present invention it
It is interior and form different embodiments.For example, in detail in the claims, the one of any of embodiment claimed is ok
In any combination mode come using.
Various component embodiments of the invention can be implemented in hardware, or to run on one or more processors
Software module realize, or with they combine realize.It should be appreciated that microprocessor or number letter can be used in practice
Number processor (DSP) come realize intention assessment model generating method according to an embodiment of the present invention, one in intension recognizing method
The some or all functions of a little or whole components.The present invention is also implemented as executing method as described herein
Some or all device or device programs (for example, computer program and computer program product).Such realization
Program of the invention can store on a computer-readable medium, or may be in the form of one or more signals.This
The signal of sample can be downloaded from an internet website to obtain, and is perhaps provided on the carrier signal or mentions in any other forms
For.
The above-mentioned embodiments illustrate rather than limit the invention, and those skilled in the art exist
It can be designed alternative embodiment without departing from the scope of the appended claims.In the claims, it should not will be located at and include
Any reference symbol between number is configured to limitations on claims.Word "comprising", which does not exclude the presence of, is not listed in claim
In element or step.Word "a" or "an" located in front of the element does not exclude the presence of multiple such elements.The present invention
It can be realized by means of including the hardware of several different elements and by means of properly programmed computer.If listing
In the unit claim of equipment for drying, several in these devices, which can be, to be embodied by the same item of hardware.It is single
The use of word first, second, and third does not indicate any sequence.These words can be construed to title.
Claims (36)
1. a kind of intention assessment model generating method, which is characterized in that the described method includes:
Each sample text that training sample is concentrated inputs initial intention assessment model;
According to sample areas vector, the sample areas phonetic vector in the corresponding each sample text region of each sample text
And sample areas label vector, and the initial intention assessment model is utilized, determine the prediction meaning of each sample text
Map values;
The true intention value of prediction intention value and each sample text based on each sample text calculates described initial
The penalty values of intention assessment model;
When the penalty values within a preset range, using the initial intention assessment model as intention assessment model.
2. the method according to claim 1, wherein the penalty values for calculating the initial intention assessment model
Later, the method also includes:
When the penalty values are not in the preset range, the parameter of the initial intention assessment model is adjusted, and based on described
Training sample set continues to train to initial intention assessment model adjusted.
3. the method according to claim 1, wherein each sample text that training sample is concentrated inputs
Before initial intention assessment model, the method also includes:
The parameter of each layer in initial intention assessment model is initialized;Wherein, the initial intention assessment model includes area
Domain vector aligned layer, region vector attention layer, full articulamentum and maximum value softmax layers soft;
The parameter of the region vector aligned layer includes at least: term vector, the phonetic of each sample word in each sample text
The context term vector matrix of term vector and context term vector matrix and each entity tag.
4. according to the method described in claim 3, it is characterized in that, described according to the corresponding each sample of each sample text
This text filed sample areas vector, sample areas phonetic vector and sample areas label vector, and utilize described initial
Intention assessment model determines the prediction intention value of each sample text, comprising:
For each sample text, which is inputted into the region vector aligned layer, is based on the region vector
Aligned layer divides the sample text, obtains at least one corresponding sample text region of the sample text;
For each sample text, extract the corresponding each sample text region of the sample text sample areas vector,
Sample areas phonetic vector and sample areas label vector;
Sample areas vector, sample areas phonetic based on the corresponding each sample text region of each sample text
Vector and sample areas label vector determine the prediction intention value of each sample text.
5. according to the method described in claim 4, extracting the sample it is characterized in that, described for each sample text
Sample areas vector, sample areas phonetic vector and the sample areas label in the corresponding each sample text region of text to
Amount, comprising:
Term vector by the region vector aligned layer based on each sample word in each sample text region calculates
The sample areas vector in each sample text region;
Phonetic term vector by the region vector aligned layer based on each sample word in each sample text region,
Calculate the sample areas phonetic vector in each sample text region;
The sample word pair for belonging to target domain in each sample text region is determined by the region vector aligned layer
The entity tag answered, and based on the corresponding entity mark of sample word for belonging to target domain in each sample text region
Label calculate the sample areas label vector in each sample text region.
6. according to the method described in claim 5, it is characterized in that, described be based on each institute by the region vector aligned layer
The term vector of each sample word in sample text region is stated, the sample areas vector in each sample text region is calculated,
Include:
For each sample text region, by the region vector aligned layer by the sample text region middle position
Sample word determines the term vector of the core word and the context term vector matrix of the core word as core word;
By the context term vector matrix and each sample word in text filed is abutted by the region vector aligned layer
Term vector carry out the first dot product operation, and maximum pond is carried out to the result of first dot product operation, it is literary to obtain the sample
The region vector of one's respective area.
7. according to the method described in claim 5, it is characterized in that, described be based on each institute by the region vector aligned layer
The phonetic term vector of each sample word in sample text region is stated, the sample areas for calculating each sample text region is spelled
Sound vector, comprising:
For each sample text region, by the region vector aligned layer by the sample text region middle position
Sample word determines the phonetic term vector of the core word and the context phonetic term vector of the core word as core word
Matrix;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in it is each
The phonetic term vector of sample word carries out the second dot product operation, and carries out maximum pond to the result of second dot product operation,
Obtain the region phonetic vector in the sample text region.
8. according to the method described in claim 5, it is characterized in that, described be based on belonging to mesh in each sample text region
The corresponding entity tag of sample word in mark field calculates the sample areas label vector in each sample text region, packet
It includes:
For each sample text region, each reality in the sample text region is determined by the region vector aligned layer
The context term vector matrix of body label;
It is by the region vector aligned layer that the context term vector matrix of each entity tag and adjoining is text filed
In the term vector of each sample word carry out the operation of third dot product, and maximum pond is carried out to the result of third dot product operation
Change, obtains the corresponding maximum pond result of each entity tag;
The average value that the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer, obtains
The area label term vector in the sample text region.
9. a kind of intension recognizing method, which is characterized in that the described method includes:
According to the input information of user, statement text to be sorted is determined;
By the statement text input intention assessment model to be sorted, by the intention assessment model to the statement to be sorted
Text carries out intention assessment, obtains the intention classification of the statement text to be sorted;Wherein, the intention assessment model is to utilize
What method described in any item of the claim 1 to 8 generated.
10. according to the method described in claim 9, it is characterized in that, it is described by the intention assessment model to it is described to point
Class states text and carries out intention assessment, obtains the intention classification of the statement text to be sorted, comprising:
The statement text to be sorted is divided by the intention assessment model, obtains the statement text pair to be sorted
At least one answered is text filed;
Pass through the corresponding each text filed region vector of statement to be sorted text, area described in the intention assessment model extraction
Domain phonetic vector and area label vector;
It is marked by the intention assessment model based on each text filed region vector, region phonetic vector and region
Vector is signed, determines the intention classification of the statement text to be sorted.
11. according to the method described in claim 10, it is characterized in that, the intention assessment model includes the alignment of region vector
Layer;
It is described by the corresponding each text filed region of statement to be sorted text described in the intention assessment model extraction to
Amount, region phonetic vector and area label vector, comprising:
By the region vector aligned layer based on it is each it is described it is text filed in each word term vector, calculate each described
Text filed region vector;
By the region vector aligned layer based on it is each it is described it is text filed in each word phonetic term vector, calculate each
The text filed region phonetic vector;
By the region vector aligned layer determine it is each it is described it is text filed in belong to target domain domanial words it is corresponding
Entity tag, and based on it is each it is described it is text filed in the corresponding entity tag of each domanial words, calculate each text
The area label vector in region.
12. according to the method for claim 11, which is characterized in that the intention assessment model further includes that region vector pays attention to
Power layer, full articulamentum and softmax layers;
It is described by the intention assessment model based on each text filed region vector, region phonetic vector and area
Domain label vector determines the intention classification of the statement text to be sorted, comprising:
Each text filed region vector, region phonetic vector and region are marked by the region vector aligned layer
Label vector is spliced, and each text filed final area vector is obtained;
The weight of each text filed final area vector is determined by the region vector attention layer;
By the full articulamentum based on each text filed the final area vector and its weight, determine described to be sorted
State the class vector of text;
The intention classification of the statement text to be sorted is determined based on the class vector by described softmax layers.
13. according to the method for claim 11, which is characterized in that it is described by the intention assessment model to it is described to point
Class statement text is divided, and obtaining the statement text to be sorted, corresponding at least one is text filed, comprising:
Participle operation is carried out to the statement text to be sorted by the region vector aligned layer, obtains the statement to be sorted
The m word that text includes;
I-th to the N+i word in the m word is divided into a text area by the region vector aligned layer
Domain, obtaining the statement text to be sorted, corresponding at least one is text filed;
Wherein, the i=1...m-N, the N and the m are the positive integer greater than 1.
14. according to the method for claim 11, which is characterized in that described to be based on each by the region vector aligned layer
It is described it is text filed in each word term vector, calculate each text filed region vector, comprising:
For each described text filed, by the region vector aligned layer using the word in this article one's respective area middle position as
Core word determines the term vector of the core word and the context term vector matrix of the core word;
By the region vector aligned layer by the context term vector matrix and it is adjacent it is text filed in each word word
Vector carries out the operation of the 4th dot product, and carries out maximum pond to the result of the 4th dot product operation, obtains this article one's respective area
Region vector.
15. according to the method for claim 11, which is characterized in that described to be based on each by the region vector aligned layer
It is described it is text filed in each word phonetic term vector, calculate each text filed region phonetic vector, comprising:
For each described text filed, by the region vector aligned layer using the word in this article one's respective area middle position as
Core word determines the phonetic term vector of the core word and the context pinyin word vector matrix of the core word;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in it is each
The phonetic term vector of word carries out the operation of the 5th dot product, and carries out maximum pond to the result of the 5th dot product operation, obtains
The region phonetic vector of this article one's respective area.
16. according to the method for claim 11, which is characterized in that described to be determined each by the region vector aligned layer
It is described it is text filed in belong to the corresponding entity tag of domanial words of target domain, comprising:
For each described text filed, preset target domain dictionary is based on by the region vector aligned layer, by this article
The word for belonging to the target domain dictionary in the word that one's respective area includes is determined as domanial words;
The corresponding label of each domanial words is searched in the target domain dictionary by the region vector aligned layer, is obtained
The entity tag of each domanial words.
17. according to the method for claim 11, which is characterized in that it is described based on it is each it is described it is text filed in each field
The corresponding entity tag of word calculates each text filed area label vector, comprising:
For each described text filed, each entity tag in this article one's respective area is determined by the region vector aligned layer
Context term vector matrix;
It is by the region vector aligned layer that the context term vector matrix of each entity tag and adjoining is text filed
In each term vector carry out the operation of the 6th dot product, and maximum pond is carried out to the result of the 6th dot product operation, obtained each
The corresponding maximum pond result of the entity tag;
The average value that the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer, obtains
The area label term vector of this article one's respective area.
18. according to the method described in claim 9, it is characterized in that, the input information according to user, determines table to be sorted
State text, comprising:
When the input information is voice messaging, the voice messaging is converted into text, and by the text and formerly
The corresponding text of input information is determined as statement text to be sorted;
Wherein, formerly input information is received in the preset time period before receiving the input information.
19. a kind of intention assessment model generating means, which is characterized in that described device includes:
Input module, each sample text for concentrating training sample input initial intention assessment model;
First determining module, for according to the sample areas in the corresponding each sample text region of each sample text to
Amount, sample areas phonetic vector and sample areas label vector, and the initial intention assessment model is utilized, determine each institute
State the prediction intention value of sample text;
Computing module, the true intention for prediction intention value and each sample text based on each sample text
Value calculates the penalty values of the initial intention assessment model;
Second determining module, for working as the penalty values within a preset range, using the initial intention assessment model as intention
Identification model.
20. device according to claim 19, which is characterized in that described device further include:
Module is adjusted, for working as the penalty values not in the preset range, adjusts the ginseng of the initial intention assessment model
Number, and initial intention assessment model adjusted is continued to train based on the training sample set.
21. device according to claim 19, which is characterized in that described device further include:
Initialization module is initialized for the parameter to each layer in initial intention assessment model;Wherein, the initial intention
Identification model includes region vector aligned layer, region vector attention layer, full articulamentum and maximum value softmax layers soft;
The parameter of the region vector aligned layer includes at least: term vector, the phonetic of each sample word in each sample text
The context term vector matrix of term vector and context term vector matrix and each entity tag.
22. device according to claim 21, which is characterized in that first determining module, comprising:
First divides submodule, for which being inputted the region vector and is aligned for each sample text
Layer, divides the sample text based on the region vector aligned layer, obtains at least one corresponding sample of the sample text
This is text filed;
First extracting sub-module, for extracting the corresponding each sample text of the sample text for each sample text
Sample areas vector, sample areas phonetic vector and the sample areas label vector in region;
First determines submodule, for the sample area based on the corresponding each sample text region of each sample text
Domain vector, sample areas phonetic vector and sample areas label vector determine the prediction intention value of each sample text.
23. device according to claim 22, which is characterized in that first extracting sub-module, comprising:
First computing unit, for being based on each sample in each sample text region by the region vector aligned layer
The term vector of word calculates the sample areas vector in each sample text region;
Second computing unit, for being based on each sample in each sample text region by the region vector aligned layer
The phonetic term vector of word calculates the sample areas phonetic vector in each sample text region;
Third computing unit belongs to target for determining in each sample text region by the region vector aligned layer
The corresponding entity tag of sample word in field, and based on the sample word for belonging to target domain in each sample text region
The corresponding entity tag of language calculates the sample areas label vector in each sample text region.
24. device according to claim 23, which is characterized in that first computing unit is used for:
For each sample text region, by the region vector aligned layer by the sample text region middle position
Sample word determines the term vector of the core word and the context term vector matrix of the core word as core word;
By the context term vector matrix and each sample word in text filed is abutted by the region vector aligned layer
Term vector carry out the first dot product operation, and maximum pond is carried out to the result of first dot product operation, it is literary to obtain the sample
The region vector of one's respective area.
25. device according to claim 23, which is characterized in that second computing unit is used for:
For each sample text region, by the region vector aligned layer by the sample text region middle position
Sample word determines the phonetic term vector of the core word and the context phonetic term vector of the core word as core word
Matrix;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in it is each
The phonetic term vector of sample word carries out the second dot product operation, and carries out maximum pond to the result of second dot product operation,
Obtain the region phonetic vector in the sample text region.
26. device according to claim 23, which is characterized in that the third computing unit is used for:
For each sample text region, each reality in the sample text region is determined by the region vector aligned layer
The context term vector matrix of body label;
It is by the region vector aligned layer that the context term vector matrix of each entity tag and adjoining is text filed
In the term vector of each sample word carry out the operation of third dot product, and maximum pond is carried out to the result of third dot product operation
Change, obtains the corresponding maximum pond result of each entity tag;
The average value that the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer, obtains
The area label term vector in the sample text region.
27. a kind of intention assessment device, which is characterized in that described device includes:
Third determining module determines statement text to be sorted for the input information according to user;
Identification module, for passing through the intention assessment model pair for the statement text input intention assessment model to be sorted
The statement text to be sorted carries out intention assessment, obtains the intention classification of the statement text to be sorted;Wherein, the intention
Identification model is generated using device described in any one of claim 19 to 26.
28. device according to claim 27, which is characterized in that the identification module, comprising:
Second division submodule is obtained for being divided by the intention assessment model to the statement text to be sorted
The statement text to be sorted is corresponding, and at least one is text filed;
Second extracting sub-module, for passing through the corresponding each text of statement text to be sorted described in the intention assessment model extraction
Region vector, region phonetic vector and the area label vector of one's respective area;
Second determines submodule, for passing through the intention assessment model based on each text filed region vector, area
Domain phonetic vector and area label vector determine the intention classification of the statement text to be sorted.
29. device according to claim 28, which is characterized in that the intention assessment model includes the alignment of region vector
Layer;
Second extracting sub-module, comprising:
4th computing unit, for by the region vector aligned layer based on it is each it is described it is text filed in each word word
Vector calculates each text filed region vector;
5th computing unit, for by the region vector aligned layer based on it is each it is described it is text filed in each word spelling
Sound term vector calculates each text filed region phonetic vector;
6th computing unit, for by the region vector aligned layer determine it is each it is described it is text filed in belong to target domain
The corresponding entity tag of domanial words, and based on it is each it is described it is text filed in the corresponding entity tag of each domanial words,
Calculate each text filed area label vector.
30. device according to claim 29, which is characterized in that the intention assessment model further includes that region vector pays attention to
Power layer, full articulamentum and softmax layers;
Described second determines submodule, is used for:
Each text filed region vector, region phonetic vector and region are marked by the region vector aligned layer
Label vector is spliced, and each text filed final area vector is obtained;
The weight of each text filed final area vector is determined by the region vector attention layer;
By the full articulamentum based on each text filed the final area vector and its weight, determine described to be sorted
State the class vector of text;
The intention classification of the statement text to be sorted is determined based on the class vector by described softmax layers.
31. device according to claim 29, which is characterized in that described second divides submodule, is used for:
Participle operation is carried out to the statement text to be sorted by the region vector aligned layer, obtains the statement to be sorted
The m word that text includes;
I-th to the N+i word in the m word is divided into a text area by the region vector aligned layer
Domain, obtaining the statement text to be sorted, corresponding at least one is text filed;
Wherein, the i=1...m-N, the N and the m are the positive integer greater than 1.
32. device according to claim 29, which is characterized in that the 4th computing unit is used for:
For each described text filed, by the region vector aligned layer using the word in this article one's respective area middle position as
Core word determines the term vector of the core word and the context term vector matrix of the core word;
By the region vector aligned layer by the context term vector matrix and it is adjacent it is text filed in each word word
Vector carries out the operation of the 4th dot product, and carries out maximum pond to the result of the 4th dot product operation, obtains this article one's respective area
Region vector.
33. device according to claim 29, which is characterized in that the 5th computing unit is used for:
For each described text filed, by the region vector aligned layer using the word in this article one's respective area middle position as
Core word determines the phonetic term vector of the core word and the context pinyin word vector matrix of the core word;
By the region vector aligned layer by the context pinyin word vector matrix and the adjoining it is text filed in it is each
The phonetic term vector of word carries out the operation of the 5th dot product, and carries out maximum pond to the result of the 5th dot product operation, obtains
The region phonetic vector of this article one's respective area.
34. device according to claim 29, which is characterized in that the 6th computing unit is used for:
For each described text filed, preset target domain dictionary is based on by the region vector aligned layer, by this article
The word for belonging to the target domain dictionary in the word that one's respective area includes is determined as domanial words;
The corresponding label of each domanial words is searched in the target domain dictionary by the region vector aligned layer, is obtained
The entity tag of each domanial words.
35. device according to claim 29, which is characterized in that the 6th computing unit is used for:
For each described text filed, each entity tag in this article one's respective area is determined by the region vector aligned layer
Context term vector matrix;
It is by the region vector aligned layer that the context term vector matrix of each entity tag and adjoining is text filed
In each term vector carry out the operation of the 6th dot product, and maximum pond is carried out to the result of the 6th dot product operation, obtained each
The corresponding maximum pond result of the entity tag;
The average value that the corresponding maximum pond result of each entity tag is calculated by the region vector aligned layer, obtains
The area label term vector of this article one's respective area.
36. device according to claim 27, which is characterized in that the third determining module is used for:
When the input information is voice messaging, the voice messaging is converted into text, and by the text and formerly
The corresponding text of input information is determined as statement text to be sorted;
Wherein, formerly input information is received in the preset time period before receiving the input information.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811290240.4A CN109657229A (en) | 2018-10-31 | 2018-10-31 | A kind of intention assessment model generating method, intension recognizing method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811290240.4A CN109657229A (en) | 2018-10-31 | 2018-10-31 | A kind of intention assessment model generating method, intension recognizing method and device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN109657229A true CN109657229A (en) | 2019-04-19 |
Family
ID=66110323
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811290240.4A Pending CN109657229A (en) | 2018-10-31 | 2018-10-31 | A kind of intention assessment model generating method, intension recognizing method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109657229A (en) |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110232439A (en) * | 2019-08-07 | 2019-09-13 | 南京硅基智能科技有限公司 | A kind of intension recognizing method based on deep learning network |
CN110287285A (en) * | 2019-05-31 | 2019-09-27 | 平安科技(深圳)有限公司 | A kind of problem intension recognizing method, device, computer equipment and storage medium |
CN110544470A (en) * | 2019-09-11 | 2019-12-06 | 拉扎斯网络科技(上海)有限公司 | voice recognition method and device, readable storage medium and electronic equipment |
CN110598214A (en) * | 2019-09-10 | 2019-12-20 | 四川长虹电器股份有限公司 | Intention recognition result error correction method |
CN110689878A (en) * | 2019-10-11 | 2020-01-14 | 浙江百应科技有限公司 | XLNET-based intelligent voice conversation intention recognition method |
CN110889469A (en) * | 2019-09-19 | 2020-03-17 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
CN111414481A (en) * | 2020-03-19 | 2020-07-14 | 哈尔滨理工大学 | Chinese semantic matching method based on pinyin and BERT embedding |
CN111737962A (en) * | 2020-06-24 | 2020-10-02 | 平安科技(深圳)有限公司 | Entity revision method, entity revision device, computer equipment and readable storage medium |
US10916242B1 (en) | 2019-08-07 | 2021-02-09 | Nanjing Silicon Intelligence Technology Co., Ltd. | Intent recognition method based on deep learning network |
CN112507704A (en) * | 2020-12-15 | 2021-03-16 | 中国联合网络通信集团有限公司 | Multi-intention recognition method, device, equipment and storage medium |
WO2021073298A1 (en) * | 2019-10-18 | 2021-04-22 | 珠海格力电器股份有限公司 | Speech information processing method and apparatus, and intelligent terminal and storage medium |
CN113051930A (en) * | 2021-03-25 | 2021-06-29 | 润联软件***(深圳)有限公司 | Intent recognition method and device based on Bert model and related equipment |
CN113326351A (en) * | 2021-06-17 | 2021-08-31 | 湖北亿咖通科技有限公司 | User intention determining method and device |
CN114330358A (en) * | 2021-11-25 | 2022-04-12 | 深圳市普渡科技有限公司 | Intention recognition method and device, computer equipment and storage medium |
WO2023131207A1 (en) * | 2022-01-07 | 2023-07-13 | Huawei Technologies Co., Ltd. | Methods and systems for streamable multimodal language understanding |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8688453B1 (en) * | 2011-02-28 | 2014-04-01 | Nuance Communications, Inc. | Intent mining via analysis of utterances |
CN107943860A (en) * | 2017-11-08 | 2018-04-20 | 北京奇艺世纪科技有限公司 | The recognition methods and device that the training method of model, text are intended to |
CN108268444A (en) * | 2018-01-10 | 2018-07-10 | 南京邮电大学 | A kind of Chinese word cutting method based on two-way LSTM, CNN and CRF |
CN108536669A (en) * | 2018-02-27 | 2018-09-14 | 北京达佳互联信息技术有限公司 | Literal information processing method, device and terminal |
CN108549637A (en) * | 2018-04-19 | 2018-09-18 | 京东方科技集团股份有限公司 | Method for recognizing semantics, device based on phonetic and interactive system |
-
2018
- 2018-10-31 CN CN201811290240.4A patent/CN109657229A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8688453B1 (en) * | 2011-02-28 | 2014-04-01 | Nuance Communications, Inc. | Intent mining via analysis of utterances |
CN107943860A (en) * | 2017-11-08 | 2018-04-20 | 北京奇艺世纪科技有限公司 | The recognition methods and device that the training method of model, text are intended to |
CN108268444A (en) * | 2018-01-10 | 2018-07-10 | 南京邮电大学 | A kind of Chinese word cutting method based on two-way LSTM, CNN and CRF |
CN108536669A (en) * | 2018-02-27 | 2018-09-14 | 北京达佳互联信息技术有限公司 | Literal information processing method, device and terminal |
CN108549637A (en) * | 2018-04-19 | 2018-09-18 | 京东方科技集团股份有限公司 | Method for recognizing semantics, device based on phonetic and interactive system |
Non-Patent Citations (1)
Title |
---|
吴慧玲等: "基于不良文本信息过滤预处理方法的研究", 《网络安全技术与应用》 * |
Cited By (24)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110287285B (en) * | 2019-05-31 | 2023-06-16 | 平安科技(深圳)有限公司 | Method and device for identifying problem intention, computer equipment and storage medium |
CN110287285A (en) * | 2019-05-31 | 2019-09-27 | 平安科技(深圳)有限公司 | A kind of problem intension recognizing method, device, computer equipment and storage medium |
US10916242B1 (en) | 2019-08-07 | 2021-02-09 | Nanjing Silicon Intelligence Technology Co., Ltd. | Intent recognition method based on deep learning network |
CN110232439A (en) * | 2019-08-07 | 2019-09-13 | 南京硅基智能科技有限公司 | A kind of intension recognizing method based on deep learning network |
EP3869411A4 (en) * | 2019-08-07 | 2022-02-09 | Nanjing Silicon Intelligence Technology Co., Ltd. | Intent identification method based on deep learning network |
WO2021022816A1 (en) * | 2019-08-07 | 2021-02-11 | 南京硅基智能科技有限公司 | Intent identification method based on deep learning network |
CN110598214A (en) * | 2019-09-10 | 2019-12-20 | 四川长虹电器股份有限公司 | Intention recognition result error correction method |
CN110544470A (en) * | 2019-09-11 | 2019-12-06 | 拉扎斯网络科技(上海)有限公司 | voice recognition method and device, readable storage medium and electronic equipment |
CN110544470B (en) * | 2019-09-11 | 2022-03-29 | 拉扎斯网络科技(上海)有限公司 | Voice recognition method and device, readable storage medium and electronic equipment |
CN110889469A (en) * | 2019-09-19 | 2020-03-17 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
CN110889469B (en) * | 2019-09-19 | 2023-07-21 | 北京市商汤科技开发有限公司 | Image processing method and device, electronic equipment and storage medium |
CN110689878A (en) * | 2019-10-11 | 2020-01-14 | 浙江百应科技有限公司 | XLNET-based intelligent voice conversation intention recognition method |
CN110689878B (en) * | 2019-10-11 | 2020-07-28 | 浙江百应科技有限公司 | Intelligent voice conversation intention recognition method based on X L Net |
WO2021073298A1 (en) * | 2019-10-18 | 2021-04-22 | 珠海格力电器股份有限公司 | Speech information processing method and apparatus, and intelligent terminal and storage medium |
CN111414481A (en) * | 2020-03-19 | 2020-07-14 | 哈尔滨理工大学 | Chinese semantic matching method based on pinyin and BERT embedding |
CN111414481B (en) * | 2020-03-19 | 2023-09-26 | 哈尔滨理工大学 | Chinese semantic matching method based on pinyin and BERT embedding |
WO2021114880A1 (en) * | 2020-06-24 | 2021-06-17 | 平安科技(深圳)有限公司 | Method and apparatus for entity revision, computer device, and readable storage medium |
CN111737962A (en) * | 2020-06-24 | 2020-10-02 | 平安科技(深圳)有限公司 | Entity revision method, entity revision device, computer equipment and readable storage medium |
CN112507704A (en) * | 2020-12-15 | 2021-03-16 | 中国联合网络通信集团有限公司 | Multi-intention recognition method, device, equipment and storage medium |
CN112507704B (en) * | 2020-12-15 | 2023-10-03 | 中国联合网络通信集团有限公司 | Multi-intention recognition method, device, equipment and storage medium |
CN113051930A (en) * | 2021-03-25 | 2021-06-29 | 润联软件***(深圳)有限公司 | Intent recognition method and device based on Bert model and related equipment |
CN113326351A (en) * | 2021-06-17 | 2021-08-31 | 湖北亿咖通科技有限公司 | User intention determining method and device |
CN114330358A (en) * | 2021-11-25 | 2022-04-12 | 深圳市普渡科技有限公司 | Intention recognition method and device, computer equipment and storage medium |
WO2023131207A1 (en) * | 2022-01-07 | 2023-07-13 | Huawei Technologies Co., Ltd. | Methods and systems for streamable multimodal language understanding |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109657229A (en) | A kind of intention assessment model generating method, intension recognizing method and device | |
CN109947912B (en) | Model method based on intra-paragraph reasoning and joint question answer matching | |
Lin et al. | SCN: Switchable context network for semantic segmentation of RGB-D images | |
CN110490946B (en) | Text image generation method based on cross-modal similarity and antagonism network generation | |
Liu et al. | Very deep convolutional neural network based image classification using small training sample size | |
CN106960206A (en) | Character identifying method and character recognition system | |
CN110309287B (en) | Retrieval type chatting dialogue scoring method for modeling dialogue turn information | |
CN108960407B (en) | Recurrent neural network language model training method, device, equipment and medium | |
Othman et al. | A new deep learning application based on movidius ncs for embedded object detection and recognition | |
CN110706692B (en) | Training method and system of child voice recognition model | |
US20160372118A1 (en) | Context-dependent modeling of phonemes | |
CN110148400A (en) | The pronunciation recognition methods of type, the training method of model, device and equipment | |
CN109543722A (en) | A kind of emotion trend forecasting method based on sentiment analysis model | |
CN111581361A (en) | Intention identification method and device | |
CN108228576B (en) | Text translation method and device | |
CN107871497A (en) | Audio recognition method and device | |
CN107329949A (en) | A kind of semantic matching method and system | |
CN107529650A (en) | Network model construction and closed loop detection method, corresponding device and computer equipment | |
CN113361251B (en) | Text generation image method and system based on multi-stage generation countermeasure network | |
CN107679462A (en) | A kind of depth multiple features fusion sorting technique based on small echo | |
CN110059220A (en) | A kind of film recommended method based on deep learning Yu Bayesian probability matrix decomposition | |
CN109448706A (en) | Neural network language model compression method and system | |
CN114998601A (en) | Online update target tracking method and system based on Transformer | |
CN113822434A (en) | Model selection learning for knowledge distillation | |
CN114548274A (en) | Multi-modal interaction-based rumor detection method and system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190419 |