CN109063035A - A kind of man-machine more wheel dialogue methods towards trip field - Google Patents
A kind of man-machine more wheel dialogue methods towards trip field Download PDFInfo
- Publication number
- CN109063035A CN109063035A CN201810779805.9A CN201810779805A CN109063035A CN 109063035 A CN109063035 A CN 109063035A CN 201810779805 A CN201810779805 A CN 201810779805A CN 109063035 A CN109063035 A CN 109063035A
- Authority
- CN
- China
- Prior art keywords
- question sentence
- current question
- current
- slot position
- model
- 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
- G06F40/35—Discourse or dialogue representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/20—Natural language analysis
- G06F40/205—Parsing
- G06F40/211—Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- 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
- 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
- G06F40/295—Named entity recognition
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Probability & Statistics with Applications (AREA)
- Algebra (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Machine Translation (AREA)
Abstract
A kind of man-machine more wheel dialogue methods towards trip field, it is used for information technology field.The present invention solves the problems, such as that current more wheel conversational systems have difficulties to the intent information of user's question sentence and slot position information extraction.The present invention carries out standardization processing to short text question sentence, and using the intent information of the short text question sentence after DAN, CNN or BLSTM model extraction standardization processing, and the model based on BLSTM achieves the ideal effect that Micro-F1 value is 93.47%；Using the slot position information of the short text question sentence after the BLSTM-CRF model extraction standardization processing for introducing intent features word, and model achieves the ideal effect that F1 value is 89.47%；The slot position information of history and the slot position information of current question sentence are determined into current dialog state information as input, and the intent information of the current question sentence of combination determines the reply strategy of next step；User is replied to according to the determining corresponding template of reply policy selection.Present invention could apply to information technology field use.
Description
Technical field
The invention belongs to interactive fields, and in particular to a kind of man-machine more wheel dialogue sides towards trip field
Method.
Background technique
Interactive system is the man-machine bi-directional exchanges of information system that machine is considered as to a cognitive subject, is that realization is man-machine
A kind of interactive mode.In conversational system, user is with specific purpose, it is desirable to obtain the letter for meeting specific restrictive condition
Breath or service, such as: booking tickets, make a reservation, finding commodity etc..User demand is taken turns and is stated usually needs point more, by user right
The demand of oneself is improved during words.
At present both at home and abroad for field of going on a journey more wheel conversational systems research, mainly include by its pipeline (be divided into from
Right three language understanding, dialogue management, spatial term parts) and using based on the building of neural network model end to end
Task conversational system.Wherein, it is the semantic expressiveness of structuring that natural language understanding, which is the text conversion that will be inputted, judges question sentence
Intention and slot position information；Dialogue management determines current dialogue state, and determines the reply strategy of next step；Natural language is raw
It is replied at according to the obtained corresponding template of reply policy selection；Although at present both at home and abroad for more wheels pair in field of going on a journey
The research of telephone system makes some progress, but it must be more complete that the premise of research method application, which is user's question sentence,
One question sentence.
And in actual conversational system, user's question sentence is usually to be presented in the form of short text, and short text again can not
The meeting avoided causes current more wheel conversational systems to the intent information and slot of user's question sentence there are some omissions and reference phenomenon
There are certain difficulties for position information extraction.
Summary of the invention
The purpose of the present invention is the intent informations and slot position information for the current more wheel conversational systems of solution to user's question sentence
The problem of extraction has difficulties.
The technical solution adopted by the present invention to solve the above technical problem is:
A kind of man-machine more wheel dialogue methods towards trip field, the specific steps of this method are as follows:
Step 1: carrying out standardization processing to the current question sentence of user, then pronoun is explicitly indicated to existing in current question sentence
Or the case where lacking sentence minor structure, according to the slot position information involved in being interacted before user, to the instruction in current question sentence
The sentence minor structure of pronoun and missing is substituted or is filled；The current question sentence that obtains that treated；
Step 2: obtaining step 1 treated the intent information of current question sentence using DAN, CNN or BLSTM model: will
Treated that current question sentence inputs DAN, CNN or BLSTM model for step 1；The output of DAN, CNN or BLSTM model is passed through
Softmax operates to obtain the intention probability of current question sentence, the intent information by the intention of maximum probability as current question sentence；
Step 3: the BLSTM-CRF model of step 1 treated current question sentence input introduces intent features vocabulary is obtained
To the slot position information of current question sentence；
Step 4: determining current dialogue shape using the slot position information of history and the slot position information of current question sentence as input
State information, and the intent information of the current question sentence of combination determines the reply strategy of next step；
Step 5: replying to user according to the corresponding template of reply policy selection that step 4 determines.
The beneficial effects of the present invention are: the present invention provides a kind of more wheel dialogue methods towards trip field, the present invention
Solve the standardization processing, intent information and slot position in more wheel conversational systems towards trip field to short text user's question sentence
Information extraction problem, in terms of being intended to information extraction, the model based on BLSTM that the present invention uses achieve Micro-F1 value for
93.47% ideal effect, in terms of slot position information extraction, the BLSTM-CRF for the introducing intent features vocabulary that the present invention uses
Model achieves the ideal effect that F1 value is 89.47%, and the present invention considers the influence of short text in more wheel dialogues, overcomes existing
There is the limitation of technology.
Detailed description of the invention
Fig. 1 is a kind of flow chart of man-machine more wheel dialogue methods towards trip field of the present invention；
Fig. 2 is DAN model schematic of the present invention；
Fig. 3 is CNN model schematic of the present invention；
Fig. 4 is BLSTM model schematic of the present invention；
Fig. 5 is the BLSTM-CRF model schematic of the present invention for introducing intent features vocabulary；
Fig. 6 is the screenshot that conversational system described in the embodiment of the present invention demonstrates interface.
Specific embodiment
Further description of the technical solution of the present invention with reference to the accompanying drawing, and however, it is not limited to this, all to this
Inventive technique scheme is modified or replaced equivalently, and without departing from the spirit and scope of the technical solution of the present invention, should all be covered
Within the protection scope of the present invention.
Specific embodiment 1: embodiment is described with reference to Fig. 1.One kind described in present embodiment is towards trip field
Man-machine take turns dialogue methods, the specific steps of this method more are as follows:
Step 1: carrying out standardization processing to the current question sentence of user, then pronoun is explicitly indicated to existing in current question sentence
Or the case where lacking sentence minor structure, according to the slot position information involved in being interacted before user, to the instruction in current question sentence
The sentence minor structure of pronoun and missing is substituted or is filled；The current question sentence that obtains that treated；
Step 2: obtaining step 1 treated the intent information of current question sentence using DAN, CNN or BLSTM model: will
Treated that current question sentence inputs DAN, CNN or BLSTM model for step 1；The output of DAN, CNN or BLSTM model is passed through
Softmax operates to obtain the intention probability of current question sentence, the intent information by the intention of maximum probability as current question sentence；
Step 3: the BLSTM-CRF model of step 1 treated current question sentence input introduces intent features vocabulary is obtained
To the slot position information of current question sentence；
Step 4: determining current dialogue shape using the slot position information of history and the slot position information of current question sentence as input
State information, and the intent information of the current question sentence of combination determines the reply strategy of next step；
Step 5: replying to user according to the corresponding template of reply policy selection that step 4 determines.
Specific embodiment 2: embodiment is described with reference to Fig. 2.Present embodiment is to one kind described in embodiment one
Man-machine more wheel dialogue methods towards trip field are further limited；Step 1 processing is obtained using the DAN model
The detailed process of the intent information of current question sentence afterwards are as follows:
Before being tested using deep learning, the present invention use first from Google increase income word2vec training word to
Amount, Chinese web page text data of the data used from wikipedia in April, 2018, using CBOW model, term vector dimension
Degree is 300 dimensions.Subsequent deep learning model be all be subject to the training completion term vector；
The question sentence data R in artificial constructed trip field1Item crawls the question sentence number in trip field using customized query
According to R2Item obtains the question sentence data R of chat in SMP2017 evaluation and test task class and vertical class3Item；By whole R1+R2+R3Question sentence carries out
It is randomly ordered, the question sentence of a part is therefrom randomly selected after randomly ordered as training set, the question sentence of rest part is as test
Collection；
The DAN model includes input layer (embedding), average term vector layer (Average Embedding) and entirely
Articulamentum；
The training process of DAN model is as follows:
The parameter setting of DAN model: term vector dimension is 300, and full articulamentum size is that 128, dropout is set as 0.5,
Mini-batch is dimensioned to 64, the number of iterations B1, patience is set as 5, i.e., trained after 5 tests are without promotion
Stop, using Adam learning rate update method, learning rate is set as 0.001；It is non-linear in full articulamentum and softmax layers of progress
Before converting (activation primitive), standardized operation is carried out to data using batch-normalization method.
By the input layer of step 1 treated current question sentence inputs trained DAN model, input layer is by current question sentence
In each word be mapped as corresponding term vector, it is assumed that the dimension of each term vector be q, then obtain dimension be n*q term vector
Matrix；Wherein: n is the length of current question sentence, if the curtailment n of current question sentence, is filled by zero-padding method；?
Input layer, using above-mentioned trained term vector in advance, obtained term vector matrix can be static state, that is, train for experiment
It immobilizes in journey；
The term vector matrix that input layer obtains is averaged by average term vector layer according to dimension；And will it is average after word to
Moment matrix is input to full articulamentum；The output of full articulamentum operates to obtain current question sentence to belong to by softmax is intended to the general of i
Rate；Wherein, the formula of softmax operation is as follows:
Wherein: y0Represent current question sentence, p (y0=i) it is that current question sentence belongs to the probability for being intended to i, i=1,2 ..., m, m generation
The total number for figure classification of expressing the meaning, eiIt is intended to the e index of i.
Deep learning experiment of the invention is carried out in the environment of tensorflow1.2.
Specific embodiment 3: embodiment is described with reference to Fig. 3 present embodiment is to one kind described in embodiment one
Man-machine more wheel dialogue methods towards trip field are further limited；Step 1 processing is obtained using the CNN model
The detailed process of the intent information of current question sentence afterwards are as follows:
The question sentence data R in artificial constructed trip field1Item crawls the question sentence number in trip field using customized query
According to R2Item obtains the question sentence data R of chat in SMP2017 evaluation and test task class and vertical class3Item；By whole R1+R2+R3Question sentence carries out
It is randomly ordered, the question sentence of a part is therefrom randomly selected after randomly ordered as training set, the question sentence of rest part is as test
Collection；CNN model includes input layer, convolutional layer, pond layer and full articulamentum；
The training process of CNN model is as follows:
The parameter setting of CNN model: term vector dimension is 300, and convolution kernel size is 256, dropout 0.5, mini-
Batch is dimensioned to 64, the number of iterations B2, patience is set as 5, i.e., and it is trained after 5 tests are without promotion to stop,
Using Adam learning rate update method, learning rate is set as 0.001；
By the input layer of step 1 treated current question sentence inputs trained CNN model, input layer is by current question sentence
In each word be mapped as corresponding term vector, it is assumed that the dimension of each term vector be q, then obtain dimension be n*q term vector
Matrix；Wherein: n is the length of current question sentence, if the curtailment n of current question sentence, passes through zero padding (zero-padding)
Method is filled；In input layer, using above-mentioned trained term vector in advance, obtained term vector matrix can be quiet for experiment
State, i.e., it immobilizes in the training process.
Use the convolution of different convolution kernel size h*k (size of the window of word of the h for institute's convolution, k are term vector dimension)
Layer carries out convolution operation to the term vector matrix of input layer output, respectively obtains the corresponding spy of convolutional layer of each size convolution kernel
Sign, experiment use multiple convolution and carry out convolution operation, we can extract the key message of similar n-gram in sentence, obtain
Richer feature representation.Feature (feature map) size obtained due to the convolution kernel of multiple and different sizes is also different
Sample, therefore we are input to pond layer to by each feature, by maximizing pondization operation, make the dimension phase of each feature
Together；
The identical feature of the dimension that pond layer is obtained is passed through as the input of full articulamentum, the output of full articulamentum
Softmax operates to obtain the probability that current question sentence belongs to intention i；Wherein, the formula of softmax operation is as follows:
y0Represent current question sentence, p (y0=i) it is that current question sentence belongs to the probability for being intended to i, i=1,2 ..., m, m, which is represented, to be intended to
The total number of classification, eiIt is intended to the e index of i.
Specific embodiment 4: embodiment is described with reference to Fig. 4.Present embodiment is to one kind described in embodiment one
Man-machine more wheel dialogue methods towards trip field are further limited, and are obtained at step 1 using the BLSTM model
The detailed process of the intent information of current question sentence after reason are as follows:
The question sentence data R in artificial constructed trip field1Item crawls the question sentence number in trip field using customized query
According to R2Item obtains the question sentence data R of chat in SMP2017 evaluation and test task class and vertical class3Item；By whole R1+R2+R3Question sentence carries out
It is randomly ordered, the question sentence of a part is therefrom randomly selected after randomly ordered as training set, the question sentence of rest part is as test
Collection；BLSTM model includes input layer, LSTM layers of forward direction, backward LSTM layers and full articulamentum；
The training process of BLSTM model is as follows:
The parameter setting of BLSTM model: term vector dimension is 300, and LSTM layers of forward direction and backward LSTM layers of hidden layer are big
Small is 128, dropout 0.5, and mini-batch is dimensioned to 64, the number of iterations B3, patience is set as 5, that is, passes through
It crosses 5 tests to stop without training after promotion, using Adam learning rate update method, learning rate is set as 0.001；
By the input layer of step 1 treated current question sentence inputs trained BLSTM model, input layer will be asked currently
Sentence in each word be mapped as corresponding term vector, it is assumed that the dimension of each term vector be q, then obtain dimension be n*q word to
Moment matrix；Wherein: n is the length of current question sentence, if the curtailment n of current question sentence, passes through zero padding (zero-
Padding) method is filled；
The term vector matrix that input layer is obtained operates before carrying out to LSTM and backward LSTM, and will be preceding to LSTM and backward
The hidden layer that LSTM is operated is spliced according to dimension, obtains the hiding layer state h at each moment1,h2,…,hT, wherein
h1、h2And hTRespectively represent the hiding layer state at the 1st moment, the 2nd moment and T moment；
Using obtained hiding layer state as input, the output of full articulamentum operates to obtain full articulamentum by softmax
Current question sentence belongs to the probability for being intended to i；Wherein, the formula of softmax operation is as follows:
y0Represent current question sentence, p (y0=i) it is that current question sentence belongs to the probability for being intended to i, i=1,2 ..., m, m, which is represented, to be intended to
The total number of classification, eiIt is intended to the e index of i.
Specific embodiment 5: embodiment is described with reference to Fig. 5.Present embodiment is to embodiment one, two, three or four
A kind of man-machine more wheel dialogue methods towards trip field are further limited, by step 1 in the step 3
Treated, and current question sentence input introduces the BLSTM-CRF model of intent features vocabulary, obtains the slot position information of current question sentence
Detailed process are as follows:
The word vector that this section experiment uses trains word vector, the data used by the word2vec that Google increases income to use
Chinese web page text data from wikipedia in April, 2018, is different from term vector training process, in training process
In text is not segmented in advance, use SKIP-GRAM model, word vector dimension is set as 50 dimensions.
The question sentence data R in artificial constructed trip field1Item crawls the question sentence number in trip field using customized query
According to R2Item obtains the question sentence data R of chat in SMP2017 evaluation and test task class and vertical class3Item；By whole R1+R2+R3Question sentence carries out
It is randomly ordered, the question sentence of a part is therefrom randomly selected after randomly ordered as training set, the question sentence of rest part is as test
Collection；BLSTM-CRF model includes input layer, LSTM layers of forward direction, backward LSTM layers and CRF output layer；
The training process of BLSTM-CRF model is as follows:
The parameter setting of BLSTM-CRF model, word vector dimension are 50 dimensions, and LSTM layer of forward direction and backward LSTM's layers is hiding
Layer dimension is 128, it is intended that characteristic dimension 10, dropout 0.5, mini-batch are dimensioned to 64, and the number of iterations is
B4, patience is set as 5, i.e., and it is trained after 5 tests are without promotion to stop, using Adam learning rate update method, study
Rate is set as 0.001；
By the input layer of step 1 treated current question sentence inputs trained BLSTM-CRF model, input layer output
The word vector of current question sentence is expressed as X=(X1,X2,…,XN), X1For the word vector of the first character of current question sentence, N is current
The total number of word in question sentence；
Construct intent features vocabulary, it is intended that feature vocabulary includes corresponding Feature Words in each intention；Such as inquire train meaning
Train website, inquiry hotel under figure are intended to trade name etc. of going with wine.According to intent features word pair included in intent features vocabulary
Treated that current question sentence is labeled for step 1, the arrangement set Y=(y after being marked1,y2,…,yN′), y1,y2,…,
yN′The sequence obtained after respectively marking, N ' is the sequence total number in set Y；If there are intent features in current question sentence
Word then marks rule using IOBES and is labeled to the intent features word in current question sentence；If a word is that a name is real
The beginning of body is then labeled as B-label, if it is name entity inside but be not that first character is then labeled as I-label,
Ending if it is name entity is so labeled as E-label, is then labeled as S- if it is the information of single name entity
Label is labeled as O if other situations；
Arrangement set Y=(y after mark1,y2,…,yN′) by the input layer of trained BLSTM-CRF model
(embedding) vector that operation obtains each sequence indicates, respectively by the vector of each sequence indicate with the word of current question sentence to
Amount indicates to be spliced, using spliced result as LSTM layers of the forward direction of BLSTM-CRF model and backward LSTM layers of input；
To LSTM layers and backward LSTM layers of output before splicing, and using random initializtion to preceding to LSTM layers and backward
LSTM layers of parameter matrix is initialized, and the hiding layer state H at each moment is obtained1,H2,…,HT, wherein H1、H2And HTPoint
The hiding layer state at the 1st moment, the 2nd moment and T moment is not represented；
The splicing result of LSTM layers and backward LSTM layers output of forward direction is Q, and the size of splicing result Q is 2*N*h, and h is hidden
The dimension of layer is hidden, splicing result Q obtains state matrix P by CRF output layer；What state matrix P was calculated is hidden state to sight
The emission probability of survey state；
The output layer of model is not to directly adopt softmax layers, but pass through CRF layers and final label is calculated.
If directly passing through softmax in output obtains optimal label, this method does not account for the label of context, model it is defeated
Outgoing label is unable to fully use adjacent label information.Global optimization characteristic based on CRF, be utilized adjacent output label it
Between related information predict current label, achieved the effect that global optimization；
By state matrix P, each sequences y in the arrangement set Y after calculating markkProbability score s (X, yk), k=
1,2,…,N′；
State matrix P operates each sequences y in the arrangement set Y after being marked by softmaxkProbability P (yk|
X), wherein；
Wherein: e represents e index.
Each sequences y in sequence of calculation set YkLog probability log (p (yk| X)):
By the maximum sequences y of log probabilitykSlot position information as current question sentence, in which: log probability refers to is with 10
Bottom takes probability P (yk| X) logarithm.
By state matrix P, each sequences y in the arrangement set Y after calculating markkProbability score s (X, yk) tool
Body process are as follows:
A indicates transfer matrix, and calculating is the transition probability size hidden between layer state,It indicates from label yk
It is transferred to mark yk+1Probability；List entries X is represented to observation state ykEmission probability；y0And yNIndicate current sentence
Beginning (start) and terminate (eos) label, therefore the size of A is (k+2) * (k+2) (k is hidden state number), all
On possible flag sequence.
By constructing intent features vocabulary, current question sentence is labeled according to intent features vocabulary, slot can be improved
The accuracy rate of position information extraction.
Specific embodiment 6: present embodiment is to a kind of man-machine more wheels towards trip field described in embodiment one
Dialogue method is further limited, the detailed process of step 4 described in present embodiment are as follows:
Dialogue state tracking is using the slot position information of history and the slot position information of current question sentence as input, in history sentence
There are the slot positions of vacancy in son, then current dialogue state is other than the slot position information of Historical heritage sentence, also by current question sentence
The correspondence slot position information of appearance, which is filled into, to be come；The slot position information and history slot position information that occur in current question sentence are rushed
It is prominent, then it needs to carry out confirmation operation to conflict slot position；With the current dialog state information of determination, determined in conjunction with step 2 current
The intent information of question sentence, to determine the reply strategy of next step.
Determine that strategy is replied in the dialogue of next step according to current dialogue states: dialogue state, which tracks, believes the slot position of history
The slot position information conduct input of breath, current question sentence, finally exports current status information.In this stage, system maintenance slot
The state of position, slot position state table are as shown in table 1:
1 slot position state table of table
For the slot position of the vacancy in historical information, if there is the slot position information in current question sentence, this is by the slot position
Information is dissolved into current state.In the event of the slot position of conflict, being as previously mentioned " departure place " is " Beijing ", such as current
Refer to that " departure place " is " Foochow " in question sentence, then with regard to needing to carry out confirmation operation to conflict slot position.
Then, dialog strategy optimization module determines current system mode and next step according to current dialogue states
Reply strategy.The system mode of setting is as shown in table 2:
2 system state table of table
After dialogue management module generates final movement, dialogue responder module is transferred to specifically to execute.Talk with responder module according to
The final movement generated according to dialogue management generates corresponding language as agent speech transmission to webpage front-end and is presented to use
Family, this completes the complete dialogues of a wheel.
Talking with used in responder module is the method based on template.Such as the sentence of final query result is generated, it is first
Some revert statement templates are manually first formulated, then the corresponding value of slot position is substituted into template sentence.Such as order the sentence of air ticket
Subtemplate is " time sets out, and the train from from_loc to to_loc is as follows: ", if active user wishes that the departure time is
" tomorrow ", departure place are " Harbin ", and destination is " Beijing ", then the final revert statement generated after replacement is " tomorrow
It sets out, as follows from Harbin to Pekinese's train: ".In order to improve user experience, avoiding the sentence generated every time all is identical sentence
Son, I has formulated multiple template to each clause, randomly chooses one from template when generating and replying and is replied.
Specific embodiment 7: present embodiment is to a kind of man-machine more wheels towards trip field described in embodiment five
Dialogue method is further limited, the detailed process of step 4 described in present embodiment are as follows:
Dialogue state tracking is using the slot position information of history and the slot position information of current question sentence as input, in history sentence
There are the slot positions of vacancy in son, then current dialogue state is other than the slot position information of Historical heritage sentence, also by current question sentence
The correspondence slot position information of appearance, which is filled into, to be come；The slot position information and history slot position information that occur in current question sentence are rushed
It is prominent, then it needs to carry out confirmation operation to conflict slot position；With the current dialog state information of determination, determined in conjunction with step 2 current
The intent information of question sentence, to determine the reply strategy of next step.
Specific embodiment 8: present embodiment is to a kind of man-machine more wheels towards trip field described in embodiment one
Dialogue method is further limited, and carries out standardization processing to the current question sentence of user in step 1 described in present embodiment
Process are as follows: delete expression present in current question sentence and Chinese incorrect codes, correct wrong word existing for current question sentence, and set phase
The synonym table answered replaces with correct word to the near synonym for including in current question sentence.
Embodiment
The method mentioned according to the present invention, we have built the Task conversational system towards trip field, such as Fig. 6
It is shown；
Entire more wheel conversational systems are distributed according to three levels of front end, Intermediate Control Layer, background system.It mainly bears front end
Duty receives the input sentence of user, is sent into dialogue Understanding Module, while system is generated corresponding reply and shows user, realizes user
With more wheels interaction of machine.Middle layer is responsible for connection front end and backstage, according to the input of front end and semaphore control backstage
System, while the operation result for receiving backstage feeds back to front-end interface.The system on backstage is mainly appointing towards trip field
Business type conversational system.Front-end interface is that the form of webpage mainly uses the technologies such as html, ajax and javascript to realize, is led here
Introduce the realization of the Task conversational system towards trip field on backstage.
Firstly, conversation sentence often will appear demonstrative pronoun and sentence lacks part minor structure in daily dialogue
Situation, and in the conversational system using text as form, in fact it could happen that situations such as messy code, wrong word, therefore the present invention attempts base
Standardization processing is carried out to question sentence in single-wheel and more wheels, in the case where more wheels, the slot position of task based access control type conversational system setting
Mechanism makes phase to current sentence in conjunction with global information (slot position information above) and current information (this slot position information)
The standardization processing answered.
Understand followed by dialogue.After user's input problem, we identify the intention of question sentence and the slot position letter of carrying
Breath.Intention assessment is based on LSTM model training and obtains, and slot position is identified by trained using BLSTM and conditional random field models
It arrives, the input for talking with understanding is the input text of user, and output is the current intention demand of user, setting out of including in text
The key slot position information such as ground, destination, time.The attitude (positive or negative) of user is also judged simultaneously, it is assumed that current round
System the slot position information extracted is confirmed, if dialogue understand without extraction negative attitude, default use
Family confirms that slot information is correct.
Conversational system towards trip field includes the slot positions such as departure place, destination, time, hotel's name, each slot altogether
With slot value be sky, slot value is to be confirmed and slot value has confirmed that three kinds of situations.Slot position information above and current question sentence are combined first
Slot position information, determine current dialogue state.
Next the reply strategy of next step conversational system is determined.Slot position acquisition of information is completed or history slot position information and is worked as
Before the slot position information extracted when clashing, need to confirm whether slot position information collects correctly；It is not collected in some slot position values
To when, it is necessary to inquired to user；Such as when slot position corresponding under intention is all own acknowledgement state, then selection terminates
This conversation tasks.
Finally, selecting suitable sentence to show from the data of json format according to the reply strategy that dialogue management exports
To user.
The research contents of this paper is more wheel conversational systems towards trip field, and intention assessment and slot position identification are dialogues
The key that query understands takes following several method to solve the problems, such as to lack in the corresponding corpus in trip field herein
It carries out the collection of corpus: the form of manual compiling being taken to obtain the higher a collection of specific subject corpus of quality, artificial constructed number
It is 3000 according to collection size；Know from Baidu and crawled the training set that size is 18048 using customized query, right
The mass data obtained through crawler carries out later period screening and is left more than 10651 datas after arranging.In addition, SMP2017 evaluation and test is appointed
It include 3069 corpus of chat class and vertical class (including totally 30 vertical fields such as flight, train) in business, I am by flight, train
Relevant corpus is put into accordingly vertical class, and remaining hang down class and chat class corpus arrangement are put into " other " class, such as 3 institute of table
Show.
Table 3
In conjunction with above-mentioned three kinds of approach, arranges obtain 13852 query corpus altogether.It, will to the data that manual sorting is collected into
Data carry out it is randomly ordered after, randomly select wherein 5% data as test set, it is remaining as training set.And it will
The training set and test set of the dialogue evaluation and test data of smp2017 are put into corresponding training set and test set, it is intended that identification number
It is shown in Table 4 according to the statistical conditions of collection, wherein training set shares 12503 problems, and development set has 1349 problems.
Each theme corpus situation after table 4 manually marks
For two classification problems, more commonly used index includes accuracy rate precision, recall rate recall and F1
Value.It due to 30 vertical classes in SMP and chats in training set corpus, sample is 62 under " flight " classification, under " train " classification
Sample is 70, and sample is 2167 under " other " classification, and the sample of " flight " and " train " class is very few, therefore this project not needle
Model is trained to the training corpus of SMP and is tested on testing material, but by the training set of SMP and artificially collects mark
Corpus combine and be trained, and compare experimental analysis to the test set after integration.
Table 5 is the intention assessment experimental result of DAN model.The recall of " flight " classification compares pair of the other three classification
It should be worth relatively low, illustrate that classifier tends to will to belong under the question sentence misclassification to other classifications of " flight " classification.
5 DAN intention assessment experimental result of table
Table 6 is the intention assessment experimental result of CNN model.Intention assessment model based on CNN all table in each classification
Now preferably.
6 CNN intention assessment experimental result of table
Table 7 is the intention assessment experimental result of BLSTM model, the intention assessment modelling effect based on BLSTM compared with other three
A model ideal is a bit.
7 BLSTM intention assessment experimental result of table
Table 8 is the intention assessment Comparative result of DAN model, CNN model and BLSTM model on trip FIELD Data,
Middle first three items precision, recall, F1 value is calculated in a manner of weighted, that is, is calculating overall assessment index
When use lower sample number of all categories as weight.It can be seen that CNN model compares biography with LSTM model in intention assessment problem
System method effect is good, and the micro-F1 value of LSTM model compares other three classifier height with macro-F1 value.Therefore this project
Have finally chosen intention assessment model of the model based on LSTM as conversational system.
The comparison of 8 intention assessment experimental result of table
The main research of this paper be towards trip field more wheel conversational systems, therefore herein to user query into
Artificial mark is gone.
This subject study scalability and diversity of slot position, to can be used regular expression or rule carries out piece
Obtained slot position is lifted, is extracted using the mode based on regular expression, such as " train with seats type, train type, train vehicle
Secondary code, commercial air flights code, Hotel Star " etc. then uses sequence labelling mould to be difficult to be extracted with regular expression
Type is identified.Specific mark system is as shown in table 9, such as " departure place " and " destination ", for flight ticket booking and fire
It is required information for vehicle ticket booking task；" time " element can distinguish user and specifically it is expected departure times, and " hotel " is right
User is answered to expect the associated hotel moved in.
9 slot position explanation of table
The slot position composition quantity of each classification included in training set, development set is as shown in table 10.Wherein, departure place
Quantity with two ingredients in destination be it is most, the two ingredients are the key messages in trip field, in train, flight, wine
Three, shop etc. refers in being intended to.And the minimum number that hotel occurs, this is because hotel's element only has in an intention
It is involved.
10 slot position quantity of table
In addition, mainly passing through internet channel (including 12306 official websites, major airline official website, ctrip.com station etc.) herein
It is collected intent features vocabulary, is illustrated as shown in table 11:
11 intent features vocabulary of table
There are mainly two types of the labelling schemes for naming Entity recognition, and one is IOB formats, if a word is that a name is real
The beginning of body is then labeled as B-label, if it is name entity inside but be not that first character is then labeled as I-label,
Otherwise being labeled as O.Another scheme is IOBES, it is a kind of variant of IOB scheme, is marked than the former more E-label
Remember the ending of a name entity.It determines to use IOBES labelling schemes herein, it encodes the single information for naming entity, and (S is simultaneously
And the ending (E) of name entity is clearly marked.Using this scheme, a word mark can be had later at I-label
High confidence level the selectional restriction of subsequent word in I-label or E-label, and IOB scheme be merely capable of determining it is subsequent
Word not the inside of another entity tag (i.e. the latter word can only be novel entities beginning B-label perhaps O or continue
A upper entity I-label, but it is unlikely to be the I-label of a novel entities).
The common evaluation index of sequence labelling includes accuracy rate accuracy, rate of precision precision, recall rate recall
With F1 value (F1-score).
Wherein, it is predicted as positive example, practical is also positive example, we are known as true positive (TP), second situation, in advance
Surveying is positive example, and be actually negative example, we are known as false positive (FP), the third situation, predicts the example that is negative, is actually positive
Example, referred to as false negative (FN), a kind of last situation predict the example that is negative, practical be also negative example, referred to as true
negative(TN)。
Recall indicates that the sample for being actually positive example, model are classified into the specific gravity of positive example, and recall is higher,
Illustrate that model successfully recalls more positive examples；Precision expression is judged as in the sample of positive example that actually it is to model
The specific gravity of positive example, precision is higher, illustrates that model is stronger to the separating capacity of positive example sample.F1-score is the comprehensive of the two
It closes.F1-score is higher, illustrates that disaggregated model is more steady, effect is more ideal.
The CRF tool that the present invention uses is CRF++0.58 version.It is named Entity recognition using CRF++, mainly includes
Following steps: template file is determined；Handle training data and test data；Model training and parameter adjustment.
The feature templates of this paper CRF are as follows:
#Unigram
U00:% × [- 2,0]
U01:% × [- 1,0]
U02:% × [0,0]
U03:% × [1,0]
U04:% × [2,0]
U05:% × [- 1,0]/% × [0,0]
U06:% × [0,0]/% × [1,0]
U07:% × [- 2,0]/% × [- 1,0]/% × [0,0]
U08:% × [- 1,0]/% × [0,0]/% × [1,0]
U09:% × [0,0]/% × [1,0]/% × [2,0]
#Bigram
B
The feature templates of CRF include two kinds of forms:
The first is Unigram template: first character is U, this is for describing unigram feature
Template.The %x [#, #] of every a line generates the state probability function in a CRFs: f (s, o), and wherein s is the label of t moment
(output), o is the example function in context such as the CRF++ supporting paper of t moment:
Func1=if (output=B and feature=" U02: that ") return 1else return 0
It is the state probability function generated by U02:%x [0,0] in the first row of input file.By the of input file
A line " substitution " is into function, function return 1, meanwhile, if certain a line of input file is also " that " in the 1st column, and it
Output (the 2nd column) be also equally B, then this function also returns to 1 in this line.
Second is Bigram template: first character is B, and every a line %x [#, #] generates in a CRFs
Side (Edge) function: f (s', s, o), wherein s' is the label at t -1 moment.That is, Bigram type and Unigram be substantially
Machine is same, only it is contemplated that the label at t -1 moment.If only writing a B, default generates f (s', s), it means that preceding
One output token and current token would be combined into bigram features.
In CRF++, the parameter of CRF model mainly includes following several:
12 CRF parameter list of table
Model is adjusted mainly for-f parameter and-c parameter herein.
This project is shown in Table 13 using the BLSTM-CRF model for introducing intent features vocabulary, specific experimental result.
The BLSTM-CRF experimental result of the introducing intent features vocabulary of table 13
As can be seen that being 10 dimensions and the BLSTM-CRF model for introducing intent features vocabulary being used to imitate being intended to characteristic dimension
It is highly desirable on fruit.
Table 14 is slot position identification experimental result comparison.From the point of view of the experimental result of each element, for " departure place " and
The recognition effect of " destination " two elements is preferable, the reason is that being directed to departure place and destination, the context of question sentence has bright
True preposition instruction, such as " from Beijing to Harbin ", " from Beijing to Foochow " etc.；It is poor for the recognition effect of hotel's name, it is former
Because may be that training corpus is limited, the study for hotel's name fails to learn completely, and wants in test data for " hotel's name "
There is the noun not occurred in training data much in element, and it is bad that this shows model in " hotel's name " element, and
BLSTM-CRF model in the case where introducing intent features vocabulary, to the recognition effect in " hotel name " compared with other models promoted compared with
Greatly, this also illustrates the models to have the ability of certain identification unregistered word.
From the point of view of the comparison of the experimental result of model, the recognition effect of BLSTM is worst compared with other models, and CRF is as statistical machine
Classic sequence dimensioning algorithm in device study, has had reached comparatively ideal effect in small data set.BLSTM-CRF models coupling
BLSTM is to the understandability of contextual information and global optimization's ability of CRF, and effect is compared with CRF model and BLSTM model
It is good；And BLSTM-CRF (introduced feature vocabulary) model is obviously more outstanding than other models in the recognition capability to unregistered word, because
This effect is most ideal.Combination learning model based on BLSTM-CRF is taken it as a whole, and recognition capability is not so good as BLSTM-CRF mould
Type.
14 slot position of table identifies experimental result comparison
Claims (8)
1. a kind of man-machine more wheel dialogue methods towards trip field, which is characterized in that the specific steps of this method are as follows:
Step 1: carry out standardization processing to the current question sentence of user, then pronoun is explicitly indicated or lacks to existing in current question sentence
The case where weary sentence minor structure, according to the slot position information involved in being interacted before user, to the demonstrative pronoun in current question sentence
It is substituted or is filled with the sentence minor structure of missing；The current question sentence that obtains that treated；
Step 2: obtaining step 1 treated the intent information of current question sentence using DAN, CNN or BLSTM model: by step
One treated that current question sentence inputs DAN, CNN or BLSTM model；Softmax is passed through in the output of DAN, CNN or BLSTM model
Operation obtains the intention probability of current question sentence, the intent information by the intention of maximum probability as current question sentence；
Step 3: the BLSTM-CRF model of step 1 treated current question sentence input introduces intent features vocabulary is worked as
The slot position information of preceding question sentence；
Step 4: determining current dialogue state letter using the slot position information of history and the slot position information of current question sentence as input
Breath, and the intent information of the current question sentence of combination determines the reply strategy of next step；
Step 5: replying to user according to the corresponding template of reply policy selection that step 4 determines.
2. a kind of man-machine more wheel dialogue methods towards trip field according to claim 1, which is characterized in that use institute
State the detailed process that DAN model obtains step 1 treated the intent information of current question sentence are as follows:
The DAN model includes input layer, average term vector layer and full articulamentum；The training process of DAN model is as follows:
The parameter setting of DAN model: term vector dimension is 300, and full articulamentum size is that 128, dropout is set as 0.5,
Mini-batch is dimensioned to 64, the number of iterations B1, patience is set as 5, i.e., by 5 test accuracy rates without mentioning
When rising, deconditioning；Using Adam learning rate update method, learning rate is set as 0.001；
By the input layer of step 1 treated current question sentence inputs trained DAN model, input layer will be in current question sentence
Each word is mapped as corresponding term vector, it is assumed that the dimension of each term vector is q, then obtains the term vector matrix that dimension is n*q；
Wherein: n is the length of current question sentence, if the curtailment n of current question sentence, is filled by zero-padding method；
The term vector matrix that input layer obtains is averaged by average term vector layer according to dimension；And the term vector square after being averaged
Battle array is input to full articulamentum；The output of full articulamentum passes through softmax and operates to obtain the probability that current question sentence belongs to intention i；Its
In, the formula of softmax operation is as follows:
Wherein: y0Represent current question sentence, p (y0=i) it is that current question sentence belongs to the probability for being intended to i, i=1,2 ..., m, m represents meaning
The total number of figure classification, eiIt is intended to the e index of i.
3. a kind of man-machine more wheel dialogue methods towards trip field according to claim 1, which is characterized in that use institute
State the detailed process that CNN model obtains step 1 treated the intent information of current question sentence are as follows:
The CNN model includes input layer, convolutional layer, pond layer and full articulamentum；The training process of CNN model is as follows:
The parameter setting of CNN model: term vector dimension is 300, and convolution kernel size is 256, dropout 0.5, mini-batch
64 are dimensioned to, the number of iterations B2, patience is set as 5, i.e., by 5 test accuracy rates without being promoted when, stop
Training；Using Adam learning rate update method, learning rate is set as 0.001；
By the input layer of step 1 treated current question sentence inputs trained CNN model, input layer will be in current question sentence
Each word is mapped as corresponding term vector, it is assumed that the dimension of each term vector is q, then obtains the term vector matrix that dimension is n*q；
Wherein: n is the length of current question sentence, if the curtailment n of current question sentence, is filled by zero-padding method；
Convolution operation is carried out to the term vector matrix that input layer exports using the convolutional layer of different convolution kernel sizes, is respectively obtained every
The corresponding feature of convolutional layer of a size convolution kernel, is input to pond layer for each feature, by maximizing pondization operation, makes every
The dimension of a feature is identical；
The identical feature of the dimension that pond layer is obtained passes through softmax as the input of full articulamentum, the output of full articulamentum
Operation obtains current question sentence and belongs to the probability for being intended to i；Wherein, the formula of softmax operation is as follows:
y0Represent current question sentence, p (y0=i) it is that current question sentence belongs to the probability for being intended to i, i=1,2 ..., m, m, which is represented, is intended to classification
Total number, eiIt is intended to the e index of i.
4. a kind of man-machine more wheel dialogue methods towards trip field according to claim 1, which is characterized in that use institute
State the detailed process that BLSTM model obtains step 1 treated the intent information of current question sentence are as follows:
The BLSTM model includes input layer, LSTM layers of forward direction, backward LSTM layers and full articulamentum；BLSTM model was trained
Journey is as follows:
The parameter setting of BLSTM model: term vector dimension is 300, and the hidden layer size that LSTM layers and backward LSTM layers of forward direction is
128, dropout 0.5, mini-batch are dimensioned to 64, the number of iterations B3, patience is set as 5, i.e., by 5
Secondary test accuracy rate without promoted when, deconditioning；Using Adam learning rate update method, learning rate is set as 0.001；
By the input layer of step 1 treated current question sentence inputs trained BLSTM model, input layer will be in current question sentence
Each word be mapped as corresponding term vector, it is assumed that the dimension of each term vector be q, then obtain dimension be n*q term vector square
Battle array；Wherein: n is the length of current question sentence, if the curtailment n of current question sentence, is filled by zero-padding method；
The term vector matrix that input layer is obtained operates before carrying out to LSTM and backward LSTM, and by preceding to LSTM and backward LSTM
It operates obtained hidden layer to be spliced according to dimension, obtains the hiding layer state h at each moment1,h2,…,hT, wherein h1、h2
And hTRespectively represent the hiding layer state at the 1st moment, the 2nd moment and T moment；
Using obtained hiding layer state as input, the output of full articulamentum operates to obtain current full articulamentum by softmax
Question sentence belongs to the probability for being intended to i；Wherein, the formula of softmax operation is as follows:
y0Represent current question sentence, p (y0=i) it is that current question sentence belongs to the probability for being intended to i, i=1,2 ..., m, m, which is represented, is intended to classification
Total number, eiIt is intended to the e index of i.
5. a kind of man-machine more wheel dialogue methods towards trip field according to claim 1,2,3 or 4, feature exist
In obtaining by the BLSTM-CRF model of step 1 treated current question sentence input introduces intent features vocabulary in the step 3
To the detailed process of the slot position information of current question sentence are as follows:
The BLSTM-CRF model includes input layer, LSTM layers of forward direction, backward LSTM layers and CRF output layer；BLSTM-CRF mould
The training process of type is as follows:
The parameter setting of BLSTM-CRF model, word vector dimension are 50 dimensions, and LSTM layers of forward direction and backward LSTM layers of hidden layer are tieed up
Degree is 128, it is intended that characteristic dimension 10, dropout 0.5, mini-batch are dimensioned to 64, the number of iterations B4,
Patience is set as 5, i.e., by 5 test accuracy rates without promoted when, deconditioning；Using Adam learning rate update side
Method, learning rate are set as 0.001；
By the input layer of step 1 treated current question sentence inputs trained BLSTM-CRF model, input layer output is current
The word vector of question sentence is expressed as X=(X1,X2,…,XN), X1For the word vector of the first character of current question sentence, N is current question sentence
The total number of middle word；
Construct intent features vocabulary, it is intended that feature vocabulary includes corresponding Feature Words in each intention；According to intent features vocabulary
In included intent features word to step 1, treated that current question sentence is labeled, the arrangement set Y=after being marked
(y1,y2,…,yN′), y1,y2,…,yN′The sequence obtained after respectively marking, N ' is the sequence total number in set Y；
Arrangement set Y=(y after mark1,y2,…,yN′) obtained by the input layer operation of trained BLSTM-CRF model
The vector of each sequence indicates, indicates to splice with the expression of the word vector of current question sentence by the vector of each sequence respectively, will spell
Result after connecing is as LSTM layers of the forward direction of trained BLSTM-CRF model and backward LSTM layers of input；
To LSTM layers and backward LSTM layers of output before splicing, and using random initializtion to preceding to LSTM layers and LSTM layers backward
Parameter matrix initialized, obtain the hiding layer state H at each moment1,H2,…,HT, wherein H1、H2And HTIt respectively represents
The hiding layer state at the 1st moment, the 2nd moment and T moment；
The splicing result of LSTM layers and backward LSTM layers output of forward direction is Q, and the size of splicing result Q is 2*N*h, and h is hidden layer
Dimension, splicing result Q obtains state matrix P by CRF output layer；
By state matrix P, each sequences y in the arrangement set Y after calculating markkProbability score s (X, yk), k=1,
2,…,N′；
State matrix P operates each sequences y in the arrangement set Y after being marked by softmaxkProbability P (yk| X),
Wherein；
Each sequences y in sequence of calculation set YkLog probability log (p (yk| X)):
By the maximum sequences y of log probabilitykSlot position information as current question sentence.
6. a kind of man-machine more wheel dialogue methods towards trip field according to claim 1, which is characterized in that the step
Rapid four detailed process are as follows:
Dialogue state tracking is using the slot position information of history and the slot position information of current question sentence as input, in history sentence
There are the slot positions of vacancy, then current dialogue state also current question sentence occurs other than the slot position information of Historical heritage sentence
Correspondence slot position information be filled into come；Conflict for the slot position information occurred in current question sentence with history slot position information, then
It needs to carry out confirmation operation to conflict slot position；With the current dialog state information of determination, the current question sentence determined in conjunction with step 2
Intent information, to determine the reply strategy of next step.
7. a kind of man-machine more wheel dialogue methods towards trip field according to claim 5, which is characterized in that the step
Rapid four detailed process are as follows:
Dialogue state tracking is using the slot position information of history and the slot position information of current question sentence as input, in history sentence
There are the slot positions of vacancy, then current dialogue state also current question sentence occurs other than the slot position information of Historical heritage sentence
Correspondence slot position information be filled into come；Conflict for the slot position information occurred in current question sentence with history slot position information, then
It needs to carry out confirmation operation to conflict slot position；With the current dialog state information of determination, the current question sentence determined in conjunction with step 2
Intent information, to determine the reply strategy of next step.
8. a kind of man-machine more wheel dialogue methods towards trip field according to claim 1, which is characterized in that step 1
In the process of standardization processing is carried out to the current question sentence of user are as follows: delete expression present in current question sentence and Chinese incorrect codes,
Wrong word existing for current question sentence is corrected, and sets corresponding synonym table, the near synonym for including in current question sentence are replaced with
Correct word.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810779805.9A CN109063035B (en) | 2018-07-16 | 2018-07-16 | Man-machine multi-turn dialogue method for trip field |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810779805.9A CN109063035B (en) | 2018-07-16 | 2018-07-16 | Man-machine multi-turn dialogue method for trip field |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109063035A true CN109063035A (en) | 2018-12-21 |
| CN109063035B CN109063035B (en) | 2021-11-09 |
Family
ID=64816886
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810779805.9A Active CN109063035B (en) | 2018-07-16 | 2018-07-16 | Man-machine multi-turn dialogue method for trip field |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109063035B (en) |
Cited By (45)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109686360A (en) * | 2019-01-08 | 2019-04-26 | 哈尔滨理工大学 | A kind of voice is made a reservation robot |
| CN109783819A (en) * | 2019-01-18 | 2019-05-21 | 广东小天才科技有限公司 | A kind of generation method and system of regular expression |
| CN109885668A (en) * | 2019-01-25 | 2019-06-14 | 中译语通科技股份有限公司 | A kind of expansible field interactive system status tracking method and apparatus |
| CN109979457A (en) * | 2019-05-29 | 2019-07-05 | 南京硅基智能科技有限公司 | A method of thousand people, thousand face applied to Intelligent dialogue robot |
| CN110019688A (en) * | 2019-01-23 | 2019-07-16 | 艾肯特公司 | The method that robot is trained |
| CN110059193A (en) * | 2019-06-21 | 2019-07-26 | 南京擎盾信息科技有限公司 | Legal advice system based on law semanteme part and document big data statistical analysis |
| CN110110057A (en) * | 2019-04-22 | 2019-08-09 | 南京硅基智能科技有限公司 | A kind of dynamic words art group system and its application in Intelligent dialogue robot |
| CN110110846A (en) * | 2019-04-24 | 2019-08-09 | 重庆邮电大学 | Auxiliary driver's vehicle exchange method based on convolutional neural networks |
| CN110111787A (en) * | 2019-04-30 | 2019-08-09 | 华为技术有限公司 | A kind of semanteme analytic method and server |
| CN110209561A (en) * | 2019-05-09 | 2019-09-06 | 北京百度网讯科技有限公司 | Evaluating method and evaluating apparatus for dialogue platform |
| CN110413756A (en) * | 2019-07-29 | 2019-11-05 | 北京小米智能科技有限公司 | The method, device and equipment of natural language processing |
| CN110442676A (en) * | 2019-07-02 | 2019-11-12 | 北京邮电大学 | Patent retrieval method and device based on more wheel dialogues |
| CN110502608A (en) * | 2019-07-05 | 2019-11-26 | 平安科技（深圳）有限公司 | The interactive method and human-computer dialogue device of knowledge based map |
| CN110532374A (en) * | 2019-09-04 | 2019-12-03 | 深圳市慧择时代科技有限公司 | The processing method and processing device of insurance information |
| CN110569344A (en) * | 2019-08-22 | 2019-12-13 | 阿里巴巴集团控股有限公司 | Method and device for determining standard question sentence corresponding to dialog text |
| CN110717027A (en) * | 2019-10-18 | 2020-01-21 | 易小博(武汉)科技有限公司 | Multi-round intelligent question-answering method, system, controller and medium |
| CN110765252A (en) * | 2019-10-18 | 2020-02-07 | 北京邮电大学 | Knowledge-driven task-oriented dialogue management method and system easy to configure |
| CN110825949A (en) * | 2019-09-19 | 2020-02-21 | 平安科技（深圳）有限公司 | Information retrieval method based on convolutional neural network and related equipment thereof |
| CN110928997A (en) * | 2019-12-04 | 2020-03-27 | 北京文思海辉金信软件有限公司 | Intention recognition method and device, electronic equipment and readable storage medium |
| CN111177341A (en) * | 2019-12-11 | 2020-05-19 | 江苏艾佳家居用品有限公司 | End-to-end ID + SF model-based user conversation demand extraction method and system |
| CN111259668A (en) * | 2020-05-07 | 2020-06-09 | 腾讯科技（深圳）有限公司 | Reading task processing method, model training device and computer equipment |
| CN111353035A (en) * | 2020-03-11 | 2020-06-30 | 镁佳(北京)科技有限公司 | Man-machine conversation method and device, readable storage medium and electronic equipment |
| CN111353029A (en) * | 2020-02-22 | 2020-06-30 | 杭州电子科技大学 | Semantic matching-based multi-turn spoken language understanding method |
| CN111382238A (en) * | 2018-12-27 | 2020-07-07 | 上海智臻智能网络科技股份有限公司 | Interactive method and device of multi-element interactive process |
| CN111414195A (en) * | 2019-02-03 | 2020-07-14 | 北京邮电大学 | Universal configurable, multi-model-compatible and domain-transportable multi-round dialogue system |
| WO2020143381A1 (en) * | 2019-01-10 | 2020-07-16 | 华为技术有限公司 | Dialogue processing method and device |
| CN111524533A (en) * | 2020-03-02 | 2020-08-11 | 北京声智科技有限公司 | Voice operation method and device, electronic equipment and storage medium |
| CN111723574A (en) * | 2020-07-09 | 2020-09-29 | 腾讯科技（深圳）有限公司 | Information processing method and device and computer readable storage medium |
| CN111813899A (en) * | 2020-08-31 | 2020-10-23 | 腾讯科技（深圳）有限公司 | Intention identification method and device based on multiple rounds of conversations |
| CN111813900A (en) * | 2019-04-10 | 2020-10-23 | 北京猎户星空科技有限公司 | Multi-turn conversation processing method and device, electronic equipment and storage medium |
| CN112000787A (en) * | 2020-08-17 | 2020-11-27 | 上海小鹏汽车科技有限公司 | Voice interaction method, server and voice interaction system |
| WO2020244065A1 (en) * | 2019-06-04 | 2020-12-10 | 平安科技（深圳）有限公司 | Character vector definition method, apparatus and device based on artificial intelligence, and storage medium |
| CN112100349A (en) * | 2020-09-03 | 2020-12-18 | 深圳数联天下智能科技有限公司 | Multi-turn dialogue method and device, electronic equipment and storage medium |
| CN112131885A (en) * | 2019-06-24 | 2020-12-25 | 北京猎户星空科技有限公司 | Semantic recognition method and device, electronic equipment and storage medium |
| CN112364147A (en) * | 2020-12-01 | 2021-02-12 | 四川长虹电器股份有限公司 | Cross-domain multi-turn dialogue method based on knowledge graph and implementation system |
| CN112380332A (en) * | 2020-11-17 | 2021-02-19 | 深圳追一科技有限公司 | Interactive knowledge feedback method, device and computer storage medium |
| CN112445888A (en) * | 2019-09-05 | 2021-03-05 | 腾讯科技（深圳）有限公司 | Information extraction method and related equipment |
| CN112632254A (en) * | 2020-12-28 | 2021-04-09 | 平安科技（深圳）有限公司 | Conversation state determining method, terminal device and storage medium |
| CN112800737A (en) * | 2019-10-29 | 2021-05-14 | 京东数字科技控股有限公司 | Natural language text generation method and device and dialogue system |
| CN112988960A (en) * | 2021-02-09 | 2021-06-18 | 中国科学院自动化研究所 | Dialog state tracking method, device, equipment and storage medium |
| WO2021135603A1 (en) * | 2019-12-31 | 2021-07-08 | 华为技术有限公司 | Intention recognition method, server and storage medium |
| CN113254610A (en) * | 2021-05-14 | 2021-08-13 | 廖伟智 | Multi-round conversation generation method for patent consultation |
| CN113282736A (en) * | 2021-07-08 | 2021-08-20 | 北京百度网讯科技有限公司 | Dialogue understanding and model training method, device, 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 |
| CN113569918A (en) * | 2021-07-05 | 2021-10-29 | 北京淇瑀信息科技有限公司 | Classification temperature adjusting method, classification temperature adjusting device, electronic equipment and medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130158980A1 (en) * | 2011-12-15 | 2013-06-20 | Microsoft Corporation | Suggesting intent frame(s) for user request(s) |
| CN106156003A (en) * | 2016-06-30 | 2016-11-23 | 北京大学 | A kind of question sentence understanding method in question answering system |
| CN106874441A (en) * | 2017-02-07 | 2017-06-20 | 腾讯科技（上海）有限公司 | Intelligent answer method and apparatus |
| CN107369443A (en) * | 2017-06-29 | 2017-11-21 | 北京百度网讯科技有限公司 | Dialogue management method and device based on artificial intelligence |
| CN107451115A (en) * | 2017-07-11 | 2017-12-08 | 中国科学院自动化研究所 | The construction method and system of Chinese Prosodic Hierarchy forecast model end to end |
-
2018
- 2018-07-16 CN CN201810779805.9A patent/CN109063035B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20130158980A1 (en) * | 2011-12-15 | 2013-06-20 | Microsoft Corporation | Suggesting intent frame(s) for user request(s) |
| CN106156003A (en) * | 2016-06-30 | 2016-11-23 | 北京大学 | A kind of question sentence understanding method in question answering system |
| CN106874441A (en) * | 2017-02-07 | 2017-06-20 | 腾讯科技（上海）有限公司 | Intelligent answer method and apparatus |
| CN107369443A (en) * | 2017-06-29 | 2017-11-21 | 北京百度网讯科技有限公司 | Dialogue management method and device based on artificial intelligence |
| CN107451115A (en) * | 2017-07-11 | 2017-12-08 | 中国科学院自动化研究所 | The construction method and system of Chinese Prosodic Hierarchy forecast model end to end |
Non-Patent Citations (1)
| Title |
|---|
| 王东升等: "面向限定领域问答***的自然语言理解方法综述", 《计算机科学》 * |
Cited By (69)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111382238A (en) * | 2018-12-27 | 2020-07-07 | 上海智臻智能网络科技股份有限公司 | Interactive method and device of multi-element interactive process |
| CN109686360A (en) * | 2019-01-08 | 2019-04-26 | 哈尔滨理工大学 | A kind of voice is made a reservation robot |
| US11977815B2 (en) | 2019-01-10 | 2024-05-07 | Huawei Technologies Co., Ltd. | Dialogue processing method and device |
| WO2020143381A1 (en) * | 2019-01-10 | 2020-07-16 | 华为技术有限公司 | Dialogue processing method and device |
| CN111427992A (en) * | 2019-01-10 | 2020-07-17 | 华为技术有限公司 | Conversation processing method and device |
| CN109783819A (en) * | 2019-01-18 | 2019-05-21 | 广东小天才科技有限公司 | A kind of generation method and system of regular expression |
| CN109783819B (en) * | 2019-01-18 | 2023-10-20 | 广东小天才科技有限公司 | Regular expression generation method and system |
| CN110019688A (en) * | 2019-01-23 | 2019-07-16 | 艾肯特公司 | The method that robot is trained |
| CN109885668A (en) * | 2019-01-25 | 2019-06-14 | 中译语通科技股份有限公司 | A kind of expansible field interactive system status tracking method and apparatus |
| WO2020151017A1 (en) * | 2019-01-25 | 2020-07-30 | 中译语通科技股份有限公司 | Scalable field human-machine dialogue system state tracking method and device |
| CN111414195B (en) * | 2019-02-03 | 2022-03-25 | 北京邮电大学 | Universal configurable, multi-model-compatible and domain-transportable multi-round dialogue system |
| CN111414195A (en) * | 2019-02-03 | 2020-07-14 | 北京邮电大学 | Universal configurable, multi-model-compatible and domain-transportable multi-round dialogue system |
| CN111813900B (en) * | 2019-04-10 | 2023-12-08 | 北京猎户星空科技有限公司 | Multi-round dialogue processing method and device, electronic equipment and storage medium |
| CN111813900A (en) * | 2019-04-10 | 2020-10-23 | 北京猎户星空科技有限公司 | Multi-turn conversation processing method and device, electronic equipment and storage medium |
| CN110110057A (en) * | 2019-04-22 | 2019-08-09 | 南京硅基智能科技有限公司 | A kind of dynamic words art group system and its application in Intelligent dialogue robot |
| CN110110846A (en) * | 2019-04-24 | 2019-08-09 | 重庆邮电大学 | Auxiliary driver's vehicle exchange method based on convolutional neural networks |
| CN110111787A (en) * | 2019-04-30 | 2019-08-09 | 华为技术有限公司 | A kind of semanteme analytic method and server |
| CN110111787B (en) * | 2019-04-30 | 2021-07-09 | 华为技术有限公司 | Semantic parsing method and server |
| US11900924B2 (en) | 2019-04-30 | 2024-02-13 | Huawei Technologies Co., Ltd. | Semantic parsing method and server |
| CN110209561A (en) * | 2019-05-09 | 2019-09-06 | 北京百度网讯科技有限公司 | Evaluating method and evaluating apparatus for dialogue platform |
| CN110209561B (en) * | 2019-05-09 | 2024-02-09 | 北京百度网讯科技有限公司 | Evaluation method and evaluation device for dialogue platform |
| CN109979457A (en) * | 2019-05-29 | 2019-07-05 | 南京硅基智能科技有限公司 | A method of thousand people, thousand face applied to Intelligent dialogue robot |
| WO2020244065A1 (en) * | 2019-06-04 | 2020-12-10 | 平安科技（深圳）有限公司 | Character vector definition method, apparatus and device based on artificial intelligence, and storage medium |
| CN110059193A (en) * | 2019-06-21 | 2019-07-26 | 南京擎盾信息科技有限公司 | Legal advice system based on law semanteme part and document big data statistical analysis |
| CN112131885A (en) * | 2019-06-24 | 2020-12-25 | 北京猎户星空科技有限公司 | Semantic recognition method and device, electronic equipment and storage medium |
| CN110442676A (en) * | 2019-07-02 | 2019-11-12 | 北京邮电大学 | Patent retrieval method and device based on more wheel dialogues |
| WO2021003819A1 (en) * | 2019-07-05 | 2021-01-14 | 平安科技（深圳）有限公司 | Man-machine dialog method and man-machine dialog apparatus based on knowledge graph |
| CN110502608B (en) * | 2019-07-05 | 2023-10-10 | 平安科技（深圳）有限公司 | Man-machine conversation method and man-machine conversation device based on knowledge graph |
| CN110502608A (en) * | 2019-07-05 | 2019-11-26 | 平安科技（深圳）有限公司 | The interactive method and human-computer dialogue device of knowledge based map |
| EP3772012A1 (en) * | 2019-07-29 | 2021-02-03 | Beijing Xiaomi Intelligent Technology Co., Ltd. | Method and device for natural language processing and storage medium |
| CN110413756B (en) * | 2019-07-29 | 2022-02-15 | 北京小米智能科技有限公司 | Method, device and equipment for processing natural language |
| KR102330061B1 (en) | 2019-07-29 | 2021-11-26 | 베이징 시아오미 인텔리전트 테크놀로지 컴퍼니 리미티드 | Natural language processing method, apparatus, device and computer readable storage medium |
| CN110413756A (en) * | 2019-07-29 | 2019-11-05 | 北京小米智能科技有限公司 | The method, device and equipment of natural language processing |
| KR20210016262A (en) * | 2019-07-29 | 2021-02-15 | 베이징 시아오미 인텔리전트 테크놀로지 컴퍼니 리미티드 | Methods, devices and devices for processing natural language |
| US11501078B2 (en) | 2019-07-29 | 2022-11-15 | Beijing Xiaomi Intelligent Technology Co., Ltd. | Method and device for performing reinforcement learning on natural language processing model and storage medium |
| WO2021017173A1 (en) * | 2019-07-29 | 2021-02-04 | 北京小米智能科技有限公司 | Natural language processing method, apparatus and device |
| CN110569344A (en) * | 2019-08-22 | 2019-12-13 | 阿里巴巴集团控股有限公司 | Method and device for determining standard question sentence corresponding to dialog text |
| CN110532374A (en) * | 2019-09-04 | 2019-12-03 | 深圳市慧择时代科技有限公司 | The processing method and processing device of insurance information |
| CN110532374B (en) * | 2019-09-04 | 2022-05-10 | 深圳市慧择时代科技有限公司 | Insurance information processing method and device |
| CN112445888A (en) * | 2019-09-05 | 2021-03-05 | 腾讯科技（深圳）有限公司 | Information extraction method and related equipment |
| CN110825949A (en) * | 2019-09-19 | 2020-02-21 | 平安科技（深圳）有限公司 | Information retrieval method based on convolutional neural network and related equipment thereof |
| CN110765252A (en) * | 2019-10-18 | 2020-02-07 | 北京邮电大学 | Knowledge-driven task-oriented dialogue management method and system easy to configure |
| CN110717027A (en) * | 2019-10-18 | 2020-01-21 | 易小博(武汉)科技有限公司 | Multi-round intelligent question-answering method, system, controller and medium |
| CN112800737A (en) * | 2019-10-29 | 2021-05-14 | 京东数字科技控股有限公司 | Natural language text generation method and device and dialogue system |
| CN110928997A (en) * | 2019-12-04 | 2020-03-27 | 北京文思海辉金信软件有限公司 | Intention recognition method and device, electronic equipment and readable storage medium |
| CN111177341A (en) * | 2019-12-11 | 2020-05-19 | 江苏艾佳家居用品有限公司 | End-to-end ID + SF model-based user conversation demand extraction method and system |
| CN111177341B (en) * | 2019-12-11 | 2022-07-08 | 江苏艾佳家居用品有限公司 | End-to-end ID + SF model-based user conversation demand extraction method and system |
| WO2021135603A1 (en) * | 2019-12-31 | 2021-07-08 | 华为技术有限公司 | Intention recognition method, server and storage medium |
| CN111353029A (en) * | 2020-02-22 | 2020-06-30 | 杭州电子科技大学 | Semantic matching-based multi-turn spoken language understanding method |
| CN111353029B (en) * | 2020-02-22 | 2020-09-22 | 杭州电子科技大学 | Semantic matching-based multi-turn spoken language understanding method |
| CN111524533B (en) * | 2020-03-02 | 2023-12-08 | 北京声智科技有限公司 | Voice operation method, device, electronic equipment and storage medium |
| CN111524533A (en) * | 2020-03-02 | 2020-08-11 | 北京声智科技有限公司 | Voice operation method and device, electronic equipment and storage medium |
| CN111353035A (en) * | 2020-03-11 | 2020-06-30 | 镁佳(北京)科技有限公司 | Man-machine conversation method and device, readable storage medium and electronic equipment |
| CN111259668A (en) * | 2020-05-07 | 2020-06-09 | 腾讯科技（深圳）有限公司 | Reading task processing method, model training device and computer equipment |
| CN111723574A (en) * | 2020-07-09 | 2020-09-29 | 腾讯科技（深圳）有限公司 | Information processing method and device and computer readable storage medium |
| CN112000787A (en) * | 2020-08-17 | 2020-11-27 | 上海小鹏汽车科技有限公司 | Voice interaction method, server and voice interaction system |
| CN112000787B (en) * | 2020-08-17 | 2021-05-14 | 上海小鹏汽车科技有限公司 | Voice interaction method, server and voice interaction system |
| CN111813899A (en) * | 2020-08-31 | 2020-10-23 | 腾讯科技（深圳）有限公司 | Intention identification method and device based on multiple rounds of conversations |
| CN112100349B (en) * | 2020-09-03 | 2024-03-19 | 深圳数联天下智能科技有限公司 | Multi-round dialogue method and device, electronic equipment and storage medium |
| CN112100349A (en) * | 2020-09-03 | 2020-12-18 | 深圳数联天下智能科技有限公司 | Multi-turn dialogue method and device, electronic equipment and storage medium |
| CN112380332A (en) * | 2020-11-17 | 2021-02-19 | 深圳追一科技有限公司 | Interactive knowledge feedback method, device and computer storage medium |
| CN112364147A (en) * | 2020-12-01 | 2021-02-12 | 四川长虹电器股份有限公司 | Cross-domain multi-turn dialogue method based on knowledge graph and implementation system |
| CN112632254A (en) * | 2020-12-28 | 2021-04-09 | 平安科技（深圳）有限公司 | Conversation state determining method, terminal device and storage medium |
| CN112632254B (en) * | 2020-12-28 | 2024-05-28 | 平安科技（深圳）有限公司 | Dialogue state determining method, terminal device and storage medium |
| CN112988960A (en) * | 2021-02-09 | 2021-06-18 | 中国科学院自动化研究所 | Dialog state tracking method, device, 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 |
| CN113254610A (en) * | 2021-05-14 | 2021-08-13 | 廖伟智 | Multi-round conversation generation method for patent consultation |
| CN113569918A (en) * | 2021-07-05 | 2021-10-29 | 北京淇瑀信息科技有限公司 | Classification temperature adjusting method, classification temperature adjusting device, electronic equipment and medium |
| CN113282736A (en) * | 2021-07-08 | 2021-08-20 | 北京百度网讯科技有限公司 | Dialogue understanding and model training method, device, equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109063035B (en) | 2021-11-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109063035A (en) | A kind of man-machine more wheel dialogue methods towards trip field | |
| CN110838288B (en) | Voice interaction method and system and dialogue equipment | |
| US11693894B2 (en) | Conversation oriented machine-user interaction | |
| CN107239444B (en) | A kind of term vector training method and system merging part of speech and location information | |
| CN105095182B (en) | A kind of return information recommendation method and device | |
| CN110110335B (en) | Named entity identification method based on stack model | |
| CN107203511A (en) | A kind of network text name entity recognition method based on neutral net probability disambiguation | |
| CN111081220B (en) | Vehicle-mounted voice interaction method, full-duplex dialogue system, server and storage medium | |
| CN108846063A (en) | Determine the method, apparatus, equipment and computer-readable medium of problem answers | |
| CN103593412B (en) | A kind of answer method and system based on tree structure problem | |
| CN114168749A (en) | Question generation system based on knowledge graph and question word drive | |
| CN109447266A (en) | A kind of agricultural science and technology service intelligent sorting method based on big data | |
| WO2022048194A1 (en) | Method, apparatus and device for optimizing event subject identification model, and readable storage medium | |
| CN109325780A (en) | A kind of exchange method of the intelligent customer service system in E-Governance Oriented field | |
| CN116562265B (en) | Information intelligent analysis method, system and storage medium | |
| CN110297889A (en) | A kind of enterprise's Sentiment orientation analysis method based on Fusion Features | |
| CN113901208A (en) | Method for analyzing emotion tendentiousness of intermediate-crossing language comments blended with theme characteristics | |
| CN116757652A (en) | Online recruitment recommendation system and method based on large language model | |
| CN116737922A (en) | Tourist online comment fine granularity emotion analysis method and system | |
| Joukhadar et al. | Arabic dialogue act recognition for textual chatbot systems | |
| Hongli | [Retracted] Design and Application of English Grammar Error Correction System Based on Deep Learning | |
| CN110689359A (en) | Method and device for dynamically updating model | |
| CN117632098A (en) | Intelligent building design system based on AIGC | |
| CN114298011B (en) | Neural network, training method, aspect emotion analysis method, device and storage medium | |
| CN111708870A (en) | Deep neural network-based question answering method and device and storage medium |
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 |