CN109597993A - Sentence analysis processing method, device, equipment and computer readable storage medium - Google Patents
Sentence analysis processing method, device, equipment and computer readable storage medium Download PDFInfo
- Publication number
- CN109597993A CN109597993A CN201811464437.5A CN201811464437A CN109597993A CN 109597993 A CN109597993 A CN 109597993A CN 201811464437 A CN201811464437 A CN 201811464437A CN 109597993 A CN109597993 A CN 109597993A
- Authority
- CN
- China
- Prior art keywords
- point
- slot
- similitude
- intention
- word slot
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- 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
- G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Machine Translation (AREA)
Abstract
The invention discloses a kind of Sentence analysis processing method and processing device, equipment and storage medium, this method comprises: obtain the pre-training model on source domain big-sample data collection, and by pre-training model transfer learning to target domain；In target domain, each sentence feature that question sentence is preset in pre-training model is obtained, and semantic analysis is carried out to each sentence feature, to determine the default corresponding variant intention of question sentence；Intention similitude point of each intention in pre-training model is obtained, and determines that highest is intended to similitude point in each intention similitude point；Each word slot in pre-training model is obtained, determines word slot similitude of each word slot in pre-training model point, and determine highest word slot similitude point in each word slot similitude point；It obtains and exports highest and be intended to the corresponding final intention of similitude point and the corresponding final word slot of highest word slot similitude point.Having reached model and moving to frontier also Fast Learning and can execute the technical effect of speech understanding task.
Description
Technical field
The present invention relates to transfer learning technical field more particularly to a kind of Sentence analysis processing method, device, equipment and
Computer readable storage medium.
Background technique
Speech understanding model in artificial intelligence dialogue robot can play the pass for helping robot to understand that user is intended to
Keyness effect.As artificial intelligence dialogue robot is widely used, for example, the Alexa of Amazon, the small ice maker device of Microsoft
People and the siri of apple.The speech understanding ability of robot is particularly important, and needs not only to enough understand that user's is common
Demand scene, it is also necessary to the understandability of continuous expanding machinery people to new user demand scene.For new user demand
The support of scene generally requires collection and labeled data, and currently used technical solution is usually that rule match either increases
Training data.This process is not only time-consuming but also consumes money, and needs the mark team of profession.Therefore, there is mass data at some
After having learnt speech understanding model under scene, for new scene field, because only that a small amount of sample or zero sample and
It is unable to Fast Learning and executes speech understanding task as a technical problem to be solved urgently.
Summary of the invention
The main purpose of the present invention is to provide a kind of fill method of laser marking, laser mark printing device, equipment and meters
Calculation machine storage medium, it is intended to solve after model moves to frontier, because only that a small amount of sample or zero sample and cannot be fast
The technical issues of speed learns and executes speech understanding task.
To achieve the above object, the present invention provides a kind of Sentence analysis processing method, the Sentence analysis processing method packet
Include following steps:
The pre-training model on source domain big-sample data collection is obtained, and by the pre-training model transfer learning to target
Field；
In the target domain, each sentence feature that question sentence is preset in the pre-training model is obtained, and to each described
Sentence feature carries out semantic analysis, with the corresponding variant intention of the determination default question sentence；
Intention similitude point of each intention in pre-training model is obtained, and in each intention similitude point really
Determine highest and is intended to similitude point；
Each word slot in the pre-training model is obtained, determines word slot similitude of each institute's predicate slot in pre-training model
Point, and highest word slot similitude point is determined in each institute's predicate slot similitude point；
It obtains the corresponding final intention of the highest intention similitude point and the highest word slot similitude point is corresponding most
Whole word slot, and export the highest and be intended to and the final word slot.
Optionally, the intention similitude point for obtaining each intention in pre-training model, and in each intention
Determine that highest is intended to the step of similitude is divided in similitude point, comprising:
Obtain the first state vector in the pre-training model；
The corresponding intention name semantic vector of each intention is obtained, and calculates each intention name semantic vector and the first shape
Intention similitude point between state vector；
Each intention similitude point is compared, it is similar to obtain each highest intention being intended in similitude point
Property point.
Optionally, the step of acquisition each intention corresponding intention name semantic vector, comprising:
Each sentence information in the intention is obtained, and determines the corresponding statement semantics vector of each sentence information；
Obtain the average vector value of each sentence vector, and using the average vector value as the intention name semanteme to
Amount.
Optionally, each word slot obtained in the pre-training model, determines each institute's predicate slot in pre-training model
Word slot similitude the step of dividing, comprising:
Obtain each word slot in the pre-training model；
The word slot name and entirety word slot value of institute's predicate slot are obtained, and determines first similarity point and the institute of institute's predicate slot name
State the second similarity point of whole word slot value；
And it is similar with the determining word slot of institute's predicate slot is worth with the second similarity point according to the first similarity point
Property point.
Optionally, the second similarity of the first similarity of determining institute's predicate slot name point and the whole word slot value
The step of dividing, comprising:
Obtain the current position state in the pre-training model, and determine the second state of the current position state to
Amount；
The corresponding word slot name semantic vector of institute's predicate slot name is obtained, and determines institute's predicate slot name semantic vector and described second
First similarity point between state vector；
The corresponding value semantic vector of the whole word slot value is obtained, and determines the value semantic vector and described the
Second similarity point between two-state vector.
Optionally, described the step of obtaining the whole word slot value corresponding value semantic vector, comprising:
Obtain each sub- word slot value in institute's predicate slot, and determine the corresponding sub- value semanteme of each sub- word slot value to
Amount；
The third similarity point between the second state vector described in the sub- value vector sum is calculated, and obtains the third
Vector product of the similitude point between the sub- value vector；
The corresponding vector product of each sub- word slot value is obtained, and each vector product is added described whole to obtain
The corresponding value semantic vector of pronouns, general term for nouns, numerals and measure words slot value.
Optionally, the step of each word slot obtained in the pre-training model, comprising:
Obtain the default question sentence in the pre-training model；
Semantic analysis is carried out to the default question sentence in the target domain, with each in the determination pre-training model
Word slot.
In addition, to achieve the above object, the present invention also provides a kind of Sentence analysis processing unit, the Sentence analysis processing
Device includes:
Transferring module, for obtaining the pre-training model on source domain big-sample data collection, and by the pre-training model
Transfer learning is to target domain；
Determining module, in the target domain, obtaining each sentence spy for presetting question sentence in the pre-training model
Sign, and semantic analysis is carried out to each sentence feature, with the corresponding variant intention of the determination default question sentence；
First obtains module, for obtaining intention similitude point of each intention in pre-training model, and in each institute
It states and is intended to determine that highest is intended to similitude point in similitude point；
Second acquisition module determines each institute's predicate slot in pre-training for obtaining each word slot in the pre-training model
Word slot similitude point in model, and highest word slot similitude point is determined in each institute's predicate slot similitude point；
Output module, it is similar with the highest word slot for obtaining the corresponding final intention of the highest intention similitude point
Property point corresponding final word slot, and export the highest and be intended to and the final word slot.
In addition, to achieve the above object, the present invention also provides a kind of mobile terminals；
The mobile terminal includes: memory, processor and is stored on the memory and can be on the processor
The computer program of operation, in which:
The computer program realizes the step of Sentence analysis processing method as described above when being executed by the processor.
In addition, to achieve the above object, the present invention also provides a kind of storage mediums；
It is stored with computer program on the storage medium, realizes when the computer program is executed by processor as above-mentioned
Sentence analysis processing method the step of.
In the present embodiment, by obtaining the pre-training model on source domain big-sample data collection, and by the pre-training
Model transfer learning is to target domain；In the target domain, each sentence that question sentence is preset in the pre-training model is obtained
Feature, and semantic analysis is carried out to each sentence feature, with the corresponding variant intention of the determination default question sentence；It obtains each
The intention similitude being intended in pre-training model point, and determine that highest intention is similar in each intention similitude point
Property point；Each word slot in the pre-training model is obtained, determines word slot similitude of each institute's predicate slot in pre-training model point,
And highest word slot similitude point is determined in each institute's predicate slot similitude point；It is corresponding most to obtain the highest intention similitude point
It is intended to eventually and the highest word slot similitude divides corresponding final word slot, and exports the highest intention and the final word slot.
It, can come the simple classification model in principle of substitution model by way of the similitude point for calculating the similitude being intended to point and word slot
The problem of moving to target domain from source domain with very good solution, and after model moves to target domain from source domain,
Do not need user redesign planning, have scalability, do not need to increase training data again yet, thus saved manually at
This, solves after model moves to frontier, because only that a small amount of sample or zero sample and be unable to Fast Learning and execute
The technical issues of speech understanding task.
Detailed description of the invention
Fig. 1 be the hardware running environment that the embodiment of the present invention is related to terminal apparatus structure schematic diagram；
Fig. 2 is the flow diagram of Sentence analysis processing method first embodiment of the present invention；
Fig. 3 is the flow diagram of Sentence analysis processing method second embodiment of the present invention；
Fig. 4 is the functional block diagram of Sentence analysis processing unit of the present invention；
Fig. 5 is the prototype network structure chart of Sentence analysis processing method of the present invention.
The object of the invention is realized, the embodiments will be further described with reference to the accompanying drawings for functional characteristics and advantage.
Specific embodiment
It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not intended to limit the present invention.
As shown in Figure 1, Fig. 1 is the terminal structure schematic diagram for the hardware running environment that the embodiment of the present invention is related to.
The terminal of that embodiment of the invention is Sentence analysis processing equipment.
As shown in Figure 1, the terminal may include: processor 1001, such as CPU, network interface 1004, user interface
1003, memory 1005, communication bus 1002.Wherein, communication bus 1002 is for realizing the connection communication between these components.
User interface 1003 may include display screen (Display), input unit such as keyboard (Keyboard), optional user interface
1003 can also include standard wireline interface and wireless interface.Network interface 1004 optionally may include that the wired of standard connects
Mouth, wireless interface (such as WI-FI interface).Memory 1005 can be high speed RAM memory, be also possible to stable memory
(non-volatile memory), such as magnetic disk storage.Memory 1005 optionally can also be independently of aforementioned processor
1001 storage device.
Optionally, terminal can also include camera, RF (Radio Frequency, radio frequency) circuit, sensor, audio
Circuit, WiFi module etc..Wherein, sensor such as optical sensor, motion sensor and other sensors.Specifically, light
Sensor may include ambient light sensor and proximity sensor, wherein ambient light sensor can according to the light and shade of ambient light come
The brightness of display screen is adjusted, proximity sensor can close display screen and/or backlight when terminal device is moved in one's ear.Certainly,
Terminal device can also configure the other sensors such as gyroscope, barometer, hygrometer, thermometer, infrared sensor, herein no longer
It repeats.
It will be understood by those skilled in the art that the restriction of the not structure paired terminal of terminal structure shown in Fig. 1, can wrap
It includes than illustrating more or fewer components, perhaps combines certain components or different component layouts.
As shown in Figure 1, as may include that operating system, network are logical in a kind of memory 1005 of computer storage medium
Believe module, Subscriber Interface Module SIM and Sentence analysis processing routine.
In terminal shown in Fig. 1, network interface 1004 is mainly used for connecting background server, carries out with background server
Data communication；User interface 1003 is mainly used for connecting client (user terminal), carries out data communication with client；And processor
1001 can be used for calling the Sentence analysis processing routine stored in memory 1005, and execute following operation:
The pre-training model on source domain big-sample data collection is obtained, and the pre-training model is moved into target neck
Domain；
Each sentence feature in default question sentence is obtained in the pre-training model, and language is carried out to each sentence feature
Justice analysis, with the corresponding variant intention of the determination default question sentence；
Intention similitude point of each intention in pre-training model is obtained, and in each intention similitude point really
Determine highest and is intended to similitude point；
Each word slot in the pre-training model is obtained, determines word slot similitude of each institute's predicate slot in pre-training model
Point, and highest word slot similitude point is determined in each institute's predicate slot similitude point；
It obtains the corresponding highest intention of the highest intention similitude point and the highest word slot similitude point is corresponding most
High word slot, and export the highest and be intended to and the highest word slot.
The present invention provides a kind of Sentence analysis processing method, in one embodiment of Sentence analysis processing method, Sentence analysis
Processing method the following steps are included:
Step S10 obtains the pre-training model on source domain big-sample data collection, and the pre-training model is migrated and is learned
Practise target domain；
Source domain can be mature application scenarios, and there is a large amount of labeled data to be used to train each model.Target neck
Domain can be new application scenarios, only exist a small amount of or at all without labeled data.Transfer learning is that handle has been instructed in former field
The model parameter perfected the model of new target domain is shared with by certain mode come help new model training, it is substantially former
There are correlations for the data or task that reason is source domain and target domain.It is preset on source domain big-sample data collection
The model training of quantity, and from selecting one to show most excellent model as pre-training mould on the data set in these models
Then type again moves to this pre-training model in target domain small sample scene, and under target domain small sample scene, search
Collect certain customers' question sentence, further according to user's question sentence design idea/word slot frame, establishment officer is according to frame labeled data.It needs
Illustrate, under different scenes, the pre-training model framework used be it is the same, only the model of pre-training is being marked
Small Sample Database on be finetune (adjustment).It wherein, is that the parameter of large sample model is whole during finetune
The parameter for bringing initialization small sample model, then does trained fine tuning on new scene small sample labeled data.Also, when in mesh
After training successfully to default training pattern under the small sample scene of mark field and get small sample model, the small sample mould that will obtain
Type interaction gives actual user to use, and question sentence can be constantly collected in user's use process, and expand training set, then with widened number
This small sample model is promoted according to collection.
Step S20 obtains each sentence feature that question sentence is preset in the pre-training model in the target domain, and
Semantic analysis is carried out to each sentence feature, with the corresponding variant intention of the determination default question sentence；
Intention refers to that we identify this expression of user is specifically intended to what does, and are specifically intended that a classifier,
User demand is divided into some type.Such as: " I will determine Beijing to the air ticket in Shanghai " the words is the need that user expresses him
It asks, this can be defined as " informing " intention；" air ticket has several points? " the words indicates that user is inquiring ticket information,
This can be defined as " request " intention.It is default when being got from pre-training model under target domain small sample scene
After question sentence, it is also necessary to obtain the sentence word for forming default question sentence, or Chinese phrase etc..Then in pre-training model
The sentence word of input is substituted for corresponding word embedding (insertion word) in Embeddings layers, then by instructing in advance
It is each to extract to practice the two-way LSTM network architecture in representation layers of the common (common characteristics extract layer) in model
A sentence feature, then semantic analysis is carried out to these sentence features, so that it is determined that each different intention, it should be noted that
It is each be intended to be stated by several words in practical application, such as " confirmation purchase ".Wherein, LSTM (Long
Short-Term Memory) it is shot and long term memory network, it is a kind of time recurrent neural network, when being suitable for handling and predicting
Between be spaced and postpone relatively long critical event in sequence.
Step S30 obtains intention similitude point of each intention in pre-training model, and similar in each intention
Property point in determine that a highest is intended to a similitude point；
In Intent task (be intended to task) layer in pre-training model, using two-way LSTM layers by common
Representation layers of obtained feature are further abstracted, then again by the two-way each direction LSTM the last one
State is stitched together, and is denoted as hintent.By each statement word for being intended to name (intent name) inside our pre-training model
Language is converted into the semantic vector of the similar the same regular length of embedding by semantic network, then takes the semanteme
Vector and hintentBilinear operation is done, to obtain the intention similitude point of the intention, since each intention is using identical
Method, which is got, is intended to corresponding intention similitude point, therefore, can carry out size comparison by dividing each intention similitude,
It is intended to similitude point to obtain the highest highest of score value.
Supplemented by assistant solve the framework and bilinear operation of semantic network of the invention, carry out illustrating below
It is bright.
For example, it is assumed that intentional map title sni=(w1, w2...wn), each word is first substituted for by Semantic network
Corresponding word embedding:E (wi).Then one layer of DNN (Deep Neural Network, deep neural network) is used
Network is by E (wi) do Nonlinear Mapping and obtain the semantic vector of the word, finally the semantic vector of all n words is done average
To the semantic vector of the intention name.Bilinear operation is by two input vector V1And V2Do following matrix operation: score=
vT 1Wv2, obtain the similarity score of two vectors.
Step S40 obtains each word slot in the pre-training model, determines word of each institute's predicate slot in pre-training model
Slot similitude point, and highest word slot similitude point is determined in each institute's predicate slot similitude point；
Word slot is the definition for key message in user's expression, such as in the expression for ordering air ticket, our slot position has
" departure time ", " starting point ", " destination ", these three key messages needs are identified.In obtaining pre-training model
Each word slot, and determine the corresponding word slot similitude timesharing of each word slot, need the Slot task (word first in pre-training model
Slot task) layer determines the state of current location, it is specifically exactly on each input position by common
The state of representation layers of two-way LSTM and Intent task layers of two-way LSTM is stitched together as current location
State, remember t moment state be ht slot.Agree to that the map title is the same, we are by the statement word of each word slot name (slot name)
Also semantic vector r is converted into using semantic networki slotname.I-th of word slot may have multiple values simultaneously, each
Value again may be by semantic network and be converted into semantic vector, and the semantic vector of j-th of value of note is
rI, j slotvalue.It should be noted that the marking of all values adds after doing normalized with the semantic vector for corresponding to value
Weight average obtains the semantic vector r of entire word slot valuei slotvalue.R is used againi slotvalueWith ht slotQuadratic linear operation is done, is obtained
To the similarity score of the value of the word slot.The similarity score of word slot name is added this with the similarity score of word slot value
The state h of word slot and current locationt slotTotal similarity score, i.e. word slot similitude point.Then in each word slot similitude point
Determine highest word slot similitude point.
Step S50 obtains the highest and is intended to the corresponding final intention of similitude point and the highest word slot similitude point
Corresponding final word slot, and export the final intention and the final word slot.
In pre-training model, highest is intended to the corresponding intention of similitude point as final and is intended to, by highest word slot phase
Like the corresponding word slot of property point as final word slot, this final word slot and final intention are then exported again.
Supplemented by assistant solve pre-training model structure process of the invention, be exemplified below.
For example, as shown in figure 5, the model is divided into Embeddings layers, Representation layers of Common, Intent
Task layers and Slot task layers.Wherein, the sentence word of input is substituted for corresponding word Embeddings layers
Embedding, such as W0, Wt, WT+1Deng.And Representation layers of Common, Task layers and Slot task layers of Intent
It is using the two-way LSTM network architecture.Using two-way LSTM layers by common in Task layers of Intent
Representation layers of obtained feature are further abstracted, then again by the two-way each direction LSTM the last one
State is stitched together, and is denoted as hintent, then by hintentIt is carried out with each be intended to such as (Intent1, Intent2, Intent3)
Semantic Similarity (similarity system design), obtain the maximum value of similitude, i.e. Softmax, then again by similitude most
Big intention is exported the τ in i.e. figure.And exporting final word slot is also first to determine current location using identical method
State remembers that the state of t moment is ht slot, by Slot Value 1, Slot Value 2, until Slot Value n and
ht slotSimilarity system design is carried out, i.e., Semantic Similarity, Attention (attention) in figure are needed to all values
The semantic vector that does after normalized with corresponding value of similarity score be weighted and averaged, obtain entire word slot value
Semantic vector ri slotvalue.R is used againi slotvalueWith ht slotQuadratic linear operation is done, the similitude for obtaining the value of the word slot is beaten
Point.It is also required at the same time by each slot name and ht slotSimilarity system design is carried out, is beaten with obtaining the similitude of word slot name
Point.The similarity score of word slot name is added to obtain the state h of the word slot and current location with the similarity score of word slot valuet slot
Total similarity score.Then determine that highest word slot similitude is divided and exported i.e. into figure in each word slot similitude point
St。
In the present embodiment, by obtaining the pre-training model on source domain big-sample data collection, and by the pre-training
Model transfer learning is to target domain；In the target domain, each sentence that question sentence is preset in the pre-training model is obtained
Feature, and semantic analysis is carried out to each sentence feature, with the corresponding variant intention of the determination default question sentence；It obtains each
The intention similitude being intended in pre-training model point, and determine that highest intention is similar in each intention similitude point
Property point；Each word slot in the pre-training model is obtained, determines word slot similitude of each institute's predicate slot in pre-training model point,
And highest word slot similitude point is determined in each institute's predicate slot similitude point；It is corresponding most to obtain the highest intention similitude point
It is intended to eventually and the highest word slot similitude divides corresponding final word slot, and exports the highest intention and the final word slot.
It, can come the simple classification model in principle of substitution model by way of the similitude point for calculating the similitude being intended to point and word slot
The problem of moving to target domain from source domain with very good solution, and after model moves to target domain from source domain,
Do not need user redesign planning, have scalability, do not need to increase training data again yet, thus saved manually at
This, solves after model moves to frontier, because only that a small amount of sample or zero sample and be unable to Fast Learning and execute
The technical issues of speech understanding task.
Further, on the basis of first embodiment of the invention, the of Sentence analysis processing method of the present invention is proposed
Two embodiments, the present embodiment are the step S30 of first embodiment of the invention, obtain meaning of each intention in pre-training model
Figure similitude point, and the refinement for dividing middle determining highest to be intended to the step of similitude is divided in each intention similitude, reference Fig. 3,
Include:
Step S31 obtains the first state vector in the pre-training model；
Step S32 obtains the corresponding intention name semantic vector of each intention, and calculates each intention name semantic vector
Intention similitude point between first state vector；
First state vector can be task layers of Intent in a model, using two-way LSTM layers by common
Representation layers of obtained feature are further abstracted, then again by the two-way each direction LSTM the last one
State be stitched together after state vector.Be intended to name be intended to name, it is intended that statement word.When getting pre-training model
In first state vector after, it is also necessary to obtain again it is each be intended to corresponding intentions name semantic vector, then again to intention name
Semantic vector and first state vector do quadratic linear operation, to obtain the intention similitude point.And due to each intention
There is intention similitude corresponding with the intention point, and the method obtained is also essentially identical, it is therefore possible to use above-mentioned
Method is divided to obtain the corresponding intention similitude of each intention.
Step S33 is compared each intention similitude point, to obtain each highest being intended in similitude point
It is intended to similitude point.
When the intention similitude timesharing for getting each intention, it is also necessary to carry out size ratio to each intention similitude point
Compared with, it has been determined that the highest intention similitude of score point, and it is intended to similitude point as highest.It should be noted that each
It is intended to similitude point to require to be compared with other similitudes point that are intended to.
In the present embodiment, by each similitude being intended between name semantic vector and first state vector point of determination come
The similitude point highest which is intended to determined, to ensure that the accuracy that determining user is intended to, improve user uses body
Test sense.
Specifically, the step of obtaining each intention corresponding intention name semantic vector, comprising:
Step S321 obtains each sentence information in the intention, and determines the corresponding sentence language of each sentence information
Adopted vector；
It obtains the corresponding intention name semantic vector of intention and needs first to obtain all sentence information in the intention, and determine each
The corresponding statement semantics vector of a sentence information.Intentional map title sn is assumed for example, it is assumed that havingi=(w1, w2...wn),Semantic
Each word is first substituted for corresponding word embedding:E (w by networki).Then one layer of DNN (Deep is used
Neural Network, deep neural network) network is by E (wi) do Nonlinear Mapping and obtain the semantic vector of the word.
Step S322 obtains the average vector value of each sentence vector, and using the average vector value as the meaning
Map title semantic vector.
After obtaining each sentence vector in a model, it is also necessary to determine the average value of each sentence vector, i.e., it is average to
Magnitude, and using this average vector value as intention name semantic vector.
In the present embodiment, by determining the corresponding statement semantics vector of sentence information all in intention, and take it flat
Mean value, to improve the accuracy that detection is intended to similitude, has ensured the usage experience of user as name semantic vector is intended to
Sense.
Further, first embodiment of the invention to second embodiment any one on the basis of, propose the present invention
The 3rd embodiment of Sentence analysis processing method, the present embodiment are the step S40 of first embodiment of the invention, obtain the pre- instruction
Practice each word slot in model, determine the refinement for the step of word slot similitude of each institute's predicate slot in pre-training model is divided, comprising:
Step S41 obtains each word slot in the pre-training model；
Step S42 obtains the word slot name and entirety word slot value of institute's predicate slot, and the first of determining institute's predicate slot name is similar
Property point and the whole word slot value second similarity point；
First similarity point can be the similitude between word slot name and current position state point.Second similarity point can be with
It is the similitude point between whole word slot value and current position state.In practical application, word slot be usually by one or
The statement of multiple words, such as " food ", and general each word slot can have some possible values, for example, " food " this
Word slot can be easily obtained the value being likely to occur: " cake ", " apple ", " roast leg of lamb " etc..Lead in pre-training model
It crosses and default question sentence is analyzed, to determine each word slot being likely to occur, then determine the word slot name and whole word slot of word slot
Value, and determine the corresponding word slot name semantic vector of word slot name value semantic vector corresponding with whole word slot value, and
By representation layers of common of two-way LSTM and Intent on each input position in task layers of Intent
The state that the state of task layers of two-way LSTM is stitched together as current location, i.e. state vector, then in word slot name language
Adopted vector sum state vector does quadratic linear operation, obtains word slot name corresponding first similarity point, with value semantic vector
Quadratic linear operation is done with state vector, obtains the corresponding second similarity point of whole word slot value.For example, when having three in word slot
When a word slot vector A1, A2, A3, these three vectors do operation with current state vector respectively and respectively obtain a score value, then
Becomes C1, C2, C3 after the normalization of three score values, then A1*C1+A2*C2+A3*C3 be exactly entire word slot value it is semantic to
Amount.Wherein, word slot name is the name of slot position, the statement word of slot position.Whole word slot value can be and each word slot value
It is worth a related word slot value.
Step S43, and that is divided according to the first similarity point and the second similarity determines institute's predicate slot with value
Word slot similitude point.
After getting first similarity point and second similarity point, it is also necessary to by the corresponding first similarity point of word slot name
Second similarity split-phase corresponding with whole word slot value is obtained itself and value, and regard itself and value as the word slot and present bit
The word slot similitude set point.
In the present embodiment, by determining the first similarity of word slot name and the second similarity of whole word slot value, come
The word slot similitude of word slot is determined, to improve the accuracy of determining word slot similitude.
Specifically, it is determined that the second similarity point of the first similarity of institute's predicate slot name point and the whole word slot value
Step, comprising:
Step S421 obtains the current position state in the pre-training model, and determines the current position state
Second state vector；
By common on each input position in task layers of Intent in pre-training model
The state of representation layers of two-way LSTM and Intent task layers of two-way LSTM is stitched together as current location
State, i.e. the second state vector.
Step S422 obtains the corresponding word slot name semantic vector of institute's predicate slot name, and determines institute's predicate slot name semantic vector
First similarity point between second state vector；
The word slot name semantic vector of word slot name can be done word slot name by one layer of DNN network in the preset model
Nonlinear operation obtains the word slot name semantic vector, and word slot name semantic vector and the second state vector are then done secondary line again
Property operation obtain first similarity point.
Step S423 obtains the corresponding value semantic vector of the whole word slot value, and determine the value semanteme to
Second similarity point between amount and second state vector.
Obtain the corresponding semantic vector of whole word slot value can first calculate each word slot value in word slot it is semantic to
Amount, then determine the similitude point of these semantic vectors, and normalized is done later with corresponding word slot to these similitudes point
The semantic vector of value is weighted and averaged, to obtain the corresponding value semantic vector of whole word slot value, then value is semantic
The second state vector of vector sum does quadratic linear operation to obtain second similarity point.
In the present embodiment, by determining the current position state in pre-training model, to determine the first phase of word slot name
Like the second similarity of property and whole word slot value, to ensure that whether the word slot in system is to improve required for user
The usage experience sense of user.
Specifically, the step of obtaining the whole word slot value corresponding value semantic vector, comprising:
Step A10 obtains each sub- word slot value in institute's predicate slot, and determines that the corresponding son of each sub- word slot value takes
It is worth semantic vector；
Sub- word slot value can be any one word slot value in word slot.All sub- word slot values in word slot are obtained,
And it is corresponding to obtain sub- word slot value by one layer of DNN network in the preset model word slot value to be done nonlinear operation
Sub- value semantic vector.
Step A11 calculates the third similarity point between the second state vector described in the sub- value vector sum, and obtains
Vector product of the third similarity point between the sub- value vector；
Third similarity point can be the similitude between any one word slot value and current position state point.Pass through two
Sublinear operation divides to calculate the third similarity between sub- value vector sum state vector, then determines third similarity point and son
Vector product between value vector.
Step A12 obtains the corresponding vector product of each sub- word slot value, and each vector product is added to obtain
Take the corresponding value semantic vector of the whole word slot value.
Obtain the corresponding vector product of each sub- word slot value, then all vector products are added again with obtain itself and
Value, finally will value semantic vector corresponding with word slot value as a whole is worth.
In the present embodiment, by determined according to all sub- word slot values the corresponding value semanteme of whole word slot value to
Amount, to ensure that value semantic vector and all word slot values in word slot are all related, ensure that the standard of value semantic vector
True property, improves the experience sense of user.
Specifically, the step of obtaining each word slot in the pre-training model, comprising:
Step S411 obtains the default question sentence in the pre-training model；
Step S412 carries out semantic analysis to the default question sentence in the target domain, with the determination pre-training
Each word slot in model.
In pre-training model, since each default question sentence needs the word slot used to be different from, it is therefore desirable to obtain pre-
Default question sentence in training pattern, and preset question sentence to this and carry out semantic analysis, thus each in pre-training model to determine
Word slot.For example, when discovery needs thing relevant to food, word slot name is at this time when carrying out semantic analysis to default question sentence
It can be food, and each word slot in word slot can be then cake, apple, roast leg of lamb etc..
In the present embodiment, by determining each word slot in pre-training model according to the default question sentence under target domain,
To ensure that each word slot is related to default question sentence, avoids unrelated word slot and occupy word slot space, saved resource, improved
The usage experience sense of user.
In addition, the embodiment of the present invention also proposes a kind of Sentence analysis processing unit, the Sentence analysis processing referring to Fig. 4
Device includes:
Transferring module, for obtaining the pre-training model on source domain big-sample data collection, and by the pre-training model
Transfer learning is to target domain；
Determining module, in the target domain, obtaining each sentence spy for presetting question sentence in the pre-training model
Sign, and semantic analysis is carried out to each sentence feature, with the corresponding variant intention of the determination default question sentence；
First obtains module, for obtaining intention similitude point of each intention in pre-training model, and in each institute
It states and is intended to determine that highest is intended to similitude point in similitude point；
Second acquisition module determines each institute's predicate slot in pre-training for obtaining each word slot in the pre-training model
Word slot similitude point in model, and highest word slot similitude point is determined in each institute's predicate slot similitude point；
Output module, it is similar with the highest word slot for obtaining the corresponding final intention of the highest intention similitude point
Property point corresponding final word slot, and export the highest and be intended to and the final word slot.
Optionally, described first module is obtained, is also used to:
Obtain the first state vector in the pre-training model；
The corresponding intention name semantic vector of each intention is obtained, and calculates each intention name semantic vector and the first shape
Intention similitude point between state vector；
Each intention similitude point is compared, it is similar to obtain each highest intention being intended in similitude point
Property point.
Optionally, described first module is obtained, is also used to:
Each sentence information in the intention is obtained, and determines the corresponding statement semantics vector of each sentence information；
Obtain the average vector value of each sentence vector, and using the average vector value as the intention name semanteme to
Amount.
Optionally, described second module is obtained, is also used to:
Obtain each word slot in the pre-training model；
The word slot name and entirety word slot value of institute's predicate slot are obtained, and determines first similarity point and the institute of institute's predicate slot name
State the second similarity point of whole word slot value；
And it is similar with the determining word slot of institute's predicate slot is worth with the second similarity point according to the first similarity point
Property point.
Optionally, described second module is obtained, is also used to:
Obtain the current position state in the pre-training model, and determine the second state of the current position state to
Amount；
The corresponding word slot name semantic vector of institute's predicate slot name is obtained, and determines institute's predicate slot name semantic vector and described second
First similarity point between state vector；
The corresponding value semantic vector of the whole word slot value is obtained, and determines the value semantic vector and described the
Second similarity point between two-state vector.
Optionally, described second module is obtained, is also used to:
Obtain each sub- word slot value in institute's predicate slot, and determine the corresponding sub- value semanteme of each sub- word slot value to
Amount；
The third similarity point between the second state vector described in the sub- value vector sum is calculated, and obtains the third
Vector product of the similitude point between the sub- value vector；
The corresponding vector product of each sub- word slot value is obtained, and each vector product is added described whole to obtain
The corresponding value semantic vector of pronouns, general term for nouns, numerals and measure words slot value.
Optionally, described second module is obtained, is also used to:
Obtain the default question sentence in the pre-training model；
Semantic analysis is carried out to the default question sentence in the target domain, with each in the determination pre-training model
Word slot.
Wherein, the step of each Implement of Function Module of Sentence analysis processing unit can refer to Sentence analysis processing of the present invention
Each embodiment of method, details are not described herein again.
The present invention also provides a kind of terminal, the terminal includes: memory, processor, communication bus and is stored in institute
State the Sentence analysis processing routine on memory:
The communication bus is for realizing the connection communication between processor and memory；
The processor is for executing the Sentence analysis processing routine, to realize above-mentioned each reality of Sentence analysis processing method
The step of applying.
The present invention also provides a kind of storage medium, the storage medium is stored with one or more than one program, institute
Stating one or more than one program can also be executed by one or more than one processor for realizing above-mentioned sentence point
The step of analysing each embodiment of processing method.
Storage medium specific embodiment of the present invention and above-mentioned each embodiment of Sentence analysis processing method are essentially identical, herein
It repeats no more.
It should be noted that, in this document, the terms "include", "comprise" or its any other variant are intended to non-row
His property includes, so that the process, method, article or the system that include a series of elements not only include those elements, and
And further include other elements that are not explicitly listed, or further include for this process, method, article or system institute it is intrinsic
Element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that including being somebody's turn to do
There is also other identical elements in the process, method of element, article or system.
The serial number of the above embodiments of the invention is only for description, does not represent the advantages or disadvantages of the embodiments.
Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment side
Method can be realized by means of software and necessary general hardware platform, naturally it is also possible to by hardware, but in many cases
The former is more preferably embodiment.Based on this understanding, technical solution of the present invention substantially in other words does the prior art
The part contributed out can be embodied in the form of software products, which is stored in one as described above
In storage medium (such as ROM/RAM, magnetic disk, CD), including some instructions are used so that terminal device (it can be mobile phone,
Computer, server, air conditioner or network equipment etc.) execute method described in each embodiment of the present invention.
The above is only a preferred embodiment of the present invention, is not intended to limit the scope of the invention, all to utilize this hair
Equivalent structure or equivalent flow shift made by bright specification and accompanying drawing content is applied directly or indirectly in other relevant skills
Art field, is included within the scope of the present invention.
Claims (10)
1. a kind of Sentence analysis processing method, which is characterized in that the Sentence analysis processing method the following steps are included:
The pre-training model on source domain big-sample data collection is obtained, and the pre-training model transfer learning to target is led
Domain；
In the target domain, each sentence feature that question sentence is preset in the pre-training model is obtained, and to each sentence
Feature carries out semantic analysis, with the corresponding variant intention of the determination default question sentence；
Intention similitude point of each intention in pre-training model is obtained, and is determined most in each intention similitude point
Height is intended to similitude point；
Each word slot in the pre-training model is obtained, determines word slot similitude of each institute's predicate slot in training pattern point, and
Highest word slot similitude point is determined in each institute's predicate slot similitude point；
It obtains the highest and is intended to the corresponding final intention of similitude point and the corresponding final word of the highest word slot similitude point
Slot, and export the highest and be intended to and the final word slot.
2. Sentence analysis processing method as described in claim 1, which is characterized in that described to obtain each intention in pre-training
Intention similitude point in model, and determine that highest is intended to the step of similitude is divided in each intention similitude point, comprising:
Obtain the first state vector in the pre-training model；
Obtain it is each it is described be intended to corresponding intention name semantic vector, and calculate each intention name semantic vector and first state to
Intention similitude point between amount；
Each intention similitude point is compared, is intended to similitude to obtain each highest being intended in similitude point
Point.
3. Sentence analysis processing method as claimed in claim 2, which is characterized in that described to obtain the corresponding meaning of each intention
The step of map title semantic vector, comprising:
Each sentence information in the intention is obtained, and determines the corresponding statement semantics vector of each sentence information；
The average vector value of each sentence vector is obtained, and using the average vector value as the intention name semantic vector.
4. Sentence analysis processing method as described in claim 1, which is characterized in that described to obtain in the pre-training model
Each word slot determines the step of word slot similitude of each institute's predicate slot in training pattern is divided, comprising:
Obtain each word slot in the pre-training model；
Obtain institute's predicate slot word slot name and whole word slot value, and determine the first similarity point of institute's predicate slot name and described whole
The second similarity of pronouns, general term for nouns, numerals and measure words slot value point；
According to the word slot similitude point for determining institute's predicate slot with value of the first similarity point and the second similarity point.
5. Sentence analysis processing method as claimed in claim 4, which is characterized in that the first phase of determining institute's predicate slot name
The step of dividing like property point and the second similarity of the whole word slot value, comprising:
The current position state in the pre-training model is obtained, and determines the second state vector of the current position state；
The corresponding word slot name semantic vector of institute's predicate slot name is obtained, and determines institute's predicate slot name semantic vector and second state
First similarity point between vector；
The corresponding value semantic vector of the whole word slot value is obtained, and determines the value semantic vector and second shape
Second similarity point between state vector.
6. Sentence analysis processing method as claimed in claim 5, which is characterized in that described to obtain the whole word slot value pair
The step of value semantic vector answered, comprising:
Each sub- word slot value in institute's predicate slot is obtained, and determines the corresponding sub- value semantic vector of each sub- word slot value；
The third similarity point between the second state vector described in the sub- value vector sum is calculated, and it is similar to obtain the third
Property vector product point between the sub- value vector；
The corresponding vector product of each sub- word slot value is obtained, and each vector product is added to obtain the whole word
The corresponding value semantic vector of slot value.
7. Sentence analysis processing method as claimed in claim 4, which is characterized in that described to obtain in the pre-training model
The step of each word slot, comprising:
Obtain the default question sentence in the pre-training model；
Semantic analysis is carried out to the default question sentence in the target domain, with each word in the determination pre-training model
Slot.
8. a kind of Sentence analysis processing unit, which is characterized in that the Sentence analysis processing unit includes:
Transferring module is migrated for obtaining the pre-training model on source domain big-sample data collection, and by the pre-training model
Learn to target domain；
Determining module, in the target domain, obtaining each sentence feature for presetting question sentence in the pre-training model, and
Semantic analysis is carried out to each sentence feature, with the corresponding variant intention of the determination default question sentence；
First obtains module, for obtaining intention similitude point of each intention in pre-training model, and in each meaning
Determine that highest is intended to similitude point in figure similitude point；
Second acquisition module determines each institute's predicate slot in pre-training model for obtaining each word slot in the pre-training model
In word slot similitude point, and highest word slot similitude point is determined in each institute's predicate slot similitude point；
Output module is intended to the corresponding final intention of similitude point and the highest word slot similitude point for obtaining the highest
Corresponding final word slot, and export the highest and be intended to and the final word slot.
9. a kind of Sentence analysis processing equipment, which is characterized in that the Sentence analysis processing equipment includes: memory, processor
And it is stored in the Sentence analysis processing routine that can be run on the memory and on the processor, the Sentence analysis processing
The step of Sentence analysis processing method as described in any one of claims 1 to 7 is realized when program is executed by the processor.
10. a kind of computer readable storage medium, which is characterized in that be stored with sentence point on the computer readable storage medium
Processing routine is analysed, is realized as described in any one of claims 1 to 7 when the Sentence analysis processing routine is executed by processor
The step of Sentence analysis processing method.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811464437.5A CN109597993B (en) | 2018-11-30 | 2018-11-30 | Statement analysis processing method, device, equipment and computer readable storage medium |
| PCT/CN2019/081282 WO2020107765A1 (en) | 2018-11-30 | 2019-04-03 | Statement analysis processing method, apparatus and device, and computer-readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811464437.5A CN109597993B (en) | 2018-11-30 | 2018-11-30 | Statement analysis processing method, device, equipment and computer readable storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109597993A true CN109597993A (en) | 2019-04-09 |
| CN109597993B CN109597993B (en) | 2021-11-05 |
Family
ID=65959469
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811464437.5A Active CN109597993B (en) | 2018-11-30 | 2018-11-30 | Statement analysis processing method, device, equipment and computer readable storage medium |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN109597993B (en) |
| WO (1) | WO2020107765A1 (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110309875A (en) * | 2019-06-28 | 2019-10-08 | 哈尔滨工程大学 | A kind of zero sample object classification method based on the synthesis of pseudo- sample characteristics |
| CN110399492A (en) * | 2019-07-22 | 2019-11-01 | 阿里巴巴集团控股有限公司 | The training method and device of disaggregated model aiming at the problem that user's question sentence |
| CN110674648A (en) * | 2019-09-29 | 2020-01-10 | 厦门大学 | Neural network machine translation model based on iterative bidirectional migration |
| CN110909541A (en) * | 2019-11-08 | 2020-03-24 | 杭州依图医疗技术有限公司 | Instruction generation method, system, device and medium |
| CN111460118A (en) * | 2020-03-26 | 2020-07-28 | 聚好看科技股份有限公司 | Artificial intelligence conflict semantic recognition method and device |
| CN111738016A (en) * | 2020-06-28 | 2020-10-02 | 中国平安财产保险股份有限公司 | Multi-intention recognition method and related equipment |
| CN111767377A (en) * | 2020-06-22 | 2020-10-13 | 湖北马斯特谱科技有限公司 | Low-resource environment-oriented high-efficiency spoken language understanding and recognizing method |
| CN111931512A (en) * | 2020-07-01 | 2020-11-13 | 联想(北京)有限公司 | Statement intention determining method and device and storage medium |
| WO2020238985A1 (en) * | 2019-05-31 | 2020-12-03 | 中国科学院深圳先进技术研究院 | Model training method, dialogue generation method, apparatus and device, and storage medium |
| CN112883180A (en) * | 2021-02-24 | 2021-06-01 | 挂号网（杭州）科技有限公司 | Model training method and device, electronic equipment and storage medium |
| CN113326360A (en) * | 2021-04-25 | 2021-08-31 | 哈尔滨工业大学 | Natural language understanding method in small sample scene |
| WO2021169745A1 (en) * | 2020-02-25 | 2021-09-02 | 升智信息科技(南京)有限公司 | User intention recognition method and apparatus based on statement context relationship prediction |
| CN114444462A (en) * | 2022-01-26 | 2022-05-06 | 北京百度网讯科技有限公司 | Model training method and man-machine interaction method and device |
| CN117709394A (en) * | 2024-02-06 | 2024-03-15 | 华侨大学 | Vehicle track prediction model training method, multi-model migration prediction method and device |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111859909B (en) * | 2020-07-10 | 2022-05-31 | 山西大学 | Semantic scene consistency recognition reading robot |
| CN112016300B (en) * | 2020-09-09 | 2022-10-14 | 平安科技（深圳）有限公司 | Pre-training model processing method, pre-training model processing device, downstream task processing device and storage medium |
| CN112214998B (en) * | 2020-11-16 | 2023-08-22 | 中国平安财产保险股份有限公司 | Method, device, equipment and storage medium for joint identification of intention and entity |
| CN112507712B (en) * | 2020-12-11 | 2024-01-26 | 北京百度网讯科技有限公司 | Method and device for establishing slot identification model and slot identification |
| CN112926313B (en) * | 2021-03-10 | 2023-08-15 | 新华智云科技有限公司 | Method and system for extracting slot position information |
| CN113139816A (en) * | 2021-04-26 | 2021-07-20 | 北京沃东天骏信息技术有限公司 | Information processing method, device, electronic equipment and storage medium |
| CN113378970B (en) * | 2021-06-28 | 2023-08-22 | 山东浪潮成方数字服务有限公司 | Sentence similarity detection method and device, electronic equipment and storage medium |
| CN117574878B (en) * | 2024-01-15 | 2024-05-17 | 西湖大学 | Component syntactic analysis method, device and medium for mixed field |
Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106156003A (en) * | 2016-06-30 | 2016-11-23 | 北京大学 | A kind of question sentence understanding method in question answering system |
| US20170169013A1 (en) * | 2015-12-11 | 2017-06-15 | Microsoft Technology Licensing, Llc | Personalizing Natural Language Understanding Systems |
| CN107330120A (en) * | 2017-07-14 | 2017-11-07 | 三角兽（北京）科技有限公司 | Inquire answer method, inquiry answering device and computer-readable recording medium |
| CN107341146A (en) * | 2017-06-23 | 2017-11-10 | 上海交通大学 | The semantic resolution system of transportable spoken language and its implementation based on semantic groove internal structure |
| CN107688614A (en) * | 2017-08-04 | 2018-02-13 | 平安科技（深圳）有限公司 | It is intended to acquisition methods, electronic installation and computer-readable recording medium |
| CN107832476A (en) * | 2017-12-01 | 2018-03-23 | 北京百度网讯科技有限公司 | A kind of understanding method of search sequence, device, equipment and storage medium |
| CN108021660A (en) * | 2017-12-04 | 2018-05-11 | 中国人民解放军国防科技大学 | Topic self-adaptive microblog emotion analysis method based on transfer learning |
| CN108182264A (en) * | 2018-01-09 | 2018-06-19 | 武汉大学 | A kind of ranking based on cross-cutting ranking recommended models recommends method |
| CN108197167A (en) * | 2017-12-18 | 2018-06-22 | 深圳前海微众银行股份有限公司 | Human-computer dialogue processing method, equipment and readable storage medium storing program for executing |
| CN108305612A (en) * | 2017-11-21 | 2018-07-20 | 腾讯科技（深圳）有限公司 | Text-processing, model training method, device, storage medium and computer equipment |
| CN108334496A (en) * | 2018-01-30 | 2018-07-27 | 中国科学院自动化研究所 | Human-computer dialogue understanding method and system and relevant device for specific area |
| CN108681585A (en) * | 2018-05-14 | 2018-10-19 | 浙江工业大学 | A kind of construction method of the multi-source transfer learning label popularity prediction model based on NetSim-TL |
| CN108874779A (en) * | 2018-06-21 | 2018-11-23 | 东北大学 | The control method that system is write the poem according to figure established based on K8s cluster |
-
2018
- 2018-11-30 CN CN201811464437.5A patent/CN109597993B/en active Active
-
2019
- 2019-04-03 WO PCT/CN2019/081282 patent/WO2020107765A1/en active Application Filing
Patent Citations (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170169013A1 (en) * | 2015-12-11 | 2017-06-15 | Microsoft Technology Licensing, Llc | Personalizing Natural Language Understanding Systems |
| CN106156003A (en) * | 2016-06-30 | 2016-11-23 | 北京大学 | A kind of question sentence understanding method in question answering system |
| CN107341146A (en) * | 2017-06-23 | 2017-11-10 | 上海交通大学 | The semantic resolution system of transportable spoken language and its implementation based on semantic groove internal structure |
| CN107330120A (en) * | 2017-07-14 | 2017-11-07 | 三角兽（北京）科技有限公司 | Inquire answer method, inquiry answering device and computer-readable recording medium |
| CN107688614A (en) * | 2017-08-04 | 2018-02-13 | 平安科技（深圳）有限公司 | It is intended to acquisition methods, electronic installation and computer-readable recording medium |
| CN108305612A (en) * | 2017-11-21 | 2018-07-20 | 腾讯科技（深圳）有限公司 | Text-processing, model training method, device, storage medium and computer equipment |
| CN107832476A (en) * | 2017-12-01 | 2018-03-23 | 北京百度网讯科技有限公司 | A kind of understanding method of search sequence, device, equipment and storage medium |
| CN108021660A (en) * | 2017-12-04 | 2018-05-11 | 中国人民解放军国防科技大学 | Topic self-adaptive microblog emotion analysis method based on transfer learning |
| CN108197167A (en) * | 2017-12-18 | 2018-06-22 | 深圳前海微众银行股份有限公司 | Human-computer dialogue processing method, equipment and readable storage medium storing program for executing |
| CN108182264A (en) * | 2018-01-09 | 2018-06-19 | 武汉大学 | A kind of ranking based on cross-cutting ranking recommended models recommends method |
| CN108334496A (en) * | 2018-01-30 | 2018-07-27 | 中国科学院自动化研究所 | Human-computer dialogue understanding method and system and relevant device for specific area |
| CN108681585A (en) * | 2018-05-14 | 2018-10-19 | 浙江工业大学 | A kind of construction method of the multi-source transfer learning label popularity prediction model based on NetSim-TL |
| CN108874779A (en) * | 2018-06-21 | 2018-11-23 | 东北大学 | The control method that system is write the poem according to figure established based on K8s cluster |
Non-Patent Citations (3)
| Title |
|---|
| JÉRÉMIE TAFFOREAU: "Joint Syntactic and Semantic Analysis with a Multitask Deep Learning", 《INTERSPEECH 2016》 * |
| 吉米哥: "语音交互（二）——自然语言理解", 《ZHUANLI.ZHIHU.COM/P/45656022?UTM_SOURCE=QQ》 * |
| 鲁强: "基于迁移学习的知识图谱问答语义匹配模型", 《计算机应用》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11875126B2 (en) | 2019-05-31 | 2024-01-16 | Shenzhen Institutes Of Advanced Technology, Chinese Academy Of Sciences | Method, apparatus, device, and storage medium for training model and generating dialog |
| WO2020238985A1 (en) * | 2019-05-31 | 2020-12-03 | 中国科学院深圳先进技术研究院 | Model training method, dialogue generation method, apparatus and device, and storage medium |
| CN110309875A (en) * | 2019-06-28 | 2019-10-08 | 哈尔滨工程大学 | A kind of zero sample object classification method based on the synthesis of pseudo- sample characteristics |
| CN110399492A (en) * | 2019-07-22 | 2019-11-01 | 阿里巴巴集团控股有限公司 | The training method and device of disaggregated model aiming at the problem that user's question sentence |
| CN110674648A (en) * | 2019-09-29 | 2020-01-10 | 厦门大学 | Neural network machine translation model based on iterative bidirectional migration |
| CN110674648B (en) * | 2019-09-29 | 2021-04-27 | 厦门大学 | Neural network machine translation model based on iterative bidirectional migration |
| CN110909541A (en) * | 2019-11-08 | 2020-03-24 | 杭州依图医疗技术有限公司 | Instruction generation method, system, device and medium |
| WO2021169745A1 (en) * | 2020-02-25 | 2021-09-02 | 升智信息科技(南京)有限公司 | User intention recognition method and apparatus based on statement context relationship prediction |
| CN111460118A (en) * | 2020-03-26 | 2020-07-28 | 聚好看科技股份有限公司 | Artificial intelligence conflict semantic recognition method and device |
| CN111460118B (en) * | 2020-03-26 | 2023-10-20 | 聚好看科技股份有限公司 | Artificial intelligence conflict semantic recognition method and device |
| CN111767377B (en) * | 2020-06-22 | 2024-05-28 | 湖北马斯特谱科技有限公司 | Efficient spoken language understanding and identifying method oriented to low-resource environment |
| CN111767377A (en) * | 2020-06-22 | 2020-10-13 | 湖北马斯特谱科技有限公司 | Low-resource environment-oriented high-efficiency spoken language understanding and recognizing method |
| CN111738016B (en) * | 2020-06-28 | 2023-09-05 | 中国平安财产保险股份有限公司 | Multi-intention recognition method and related equipment |
| CN111738016A (en) * | 2020-06-28 | 2020-10-02 | 中国平安财产保险股份有限公司 | Multi-intention recognition method and related equipment |
| CN111931512A (en) * | 2020-07-01 | 2020-11-13 | 联想(北京)有限公司 | Statement intention determining method and device and storage medium |
| CN112883180A (en) * | 2021-02-24 | 2021-06-01 | 挂号网（杭州）科技有限公司 | Model training method and device, electronic equipment and storage medium |
| CN113326360A (en) * | 2021-04-25 | 2021-08-31 | 哈尔滨工业大学 | Natural language understanding method in small sample scene |
| CN114444462A (en) * | 2022-01-26 | 2022-05-06 | 北京百度网讯科技有限公司 | Model training method and man-machine interaction method and device |
| CN114444462B (en) * | 2022-01-26 | 2022-11-29 | 北京百度网讯科技有限公司 | Model training method and man-machine interaction method and device |
| CN117709394A (en) * | 2024-02-06 | 2024-03-15 | 华侨大学 | Vehicle track prediction model training method, multi-model migration prediction method and device |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2020107765A1 (en) | 2020-06-04 |
| CN109597993B (en) | 2021-11-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109597993A (en) | Sentence analysis processing method, device, equipment and computer readable storage medium | |
| US20200412975A1 (en) | Content capture with audio input feedback | |
| CN110249304B (en) | Visual intelligent management of electronic devices | |
| CN113704388A (en) | Training method and device for multi-task pre-training model, electronic equipment and medium | |
| US11641403B2 (en) | Analyzing augmented reality content usage data | |
| US20200412864A1 (en) | Modular camera interface | |
| US11915400B2 (en) | Location mapping for large scale augmented-reality | |
| US11934643B2 (en) | Analyzing augmented reality content item usage data | |
| US11574005B2 (en) | Client application content classification and discovery | |
| US11853399B2 (en) | Multimodal sentiment classification | |
| CN111930964B (en) | Content processing method, device, equipment and storage medium | |
| US20230146865A1 (en) | Optimizer based prunner for neural networks | |
| CN113723378B (en) | Model training method and device, computer equipment and storage medium | |
| CN114722937A (en) | Abnormal data detection method and device, electronic equipment and storage medium | |
| US20220319082A1 (en) | Generating modified user content that includes additional text content | |
| US20230289560A1 (en) | Machine learning techniques to predict content actions | |
| CN111814812A (en) | Modeling method, modeling device, storage medium, electronic device and scene recognition method | |
| US20190197480A1 (en) | Recommending relevant positions | |
| CN116601961A (en) | Visual label reveal mode detection | |
| US11928167B2 (en) | Determining classification recommendations for user content | |
| US20240211119A1 (en) | Analyzing augmented reality content item usage data | |
| EP4315107A1 (en) | Generating modified user content that includes additional text content | |
| CN117216378A (en) | Data processing method and device | |
| CN116977756A (en) | Object recognition model training method, device, equipment and storage medium | |
| CN117370657A (en) | Method, device, equipment and storage medium for recommending articles |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |