CN112131890A - Method, device and equipment for constructing intelligent recognition model of conversation intention - Google Patents
Method, device and equipment for constructing intelligent recognition model of conversation intention Download PDFInfo
- Publication number
- CN112131890A CN112131890A CN202010968430.8A CN202010968430A CN112131890A CN 112131890 A CN112131890 A CN 112131890A CN 202010968430 A CN202010968430 A CN 202010968430A CN 112131890 A CN112131890 A CN 112131890A
- Authority
- CN
- China
- Prior art keywords
- corpus data
- intention
- corpus
- data
- recognition 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.)
- Pending
Links
Images
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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/332—Query formulation
- G06F16/3329—Natural language query formulation or dialogue systems
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/35—Clustering; Classification
- G06F16/353—Clustering; Classification into predefined classes
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention is suitable for the technical field of big data artificial intelligence, and provides a method, a device and equipment for constructing an intelligent recognition model of conversation intention, wherein the method comprises the following steps: obtaining corpus data, wherein the corpus data comprises labeled corpus data and unlabeled corpus data; processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate intention corpus data with label information; and performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model. According to the invention, a large amount of unmarked corpus data is utilized, semi-automatic marking of the training corpus is realized based on a preset semantic analysis algorithm, a large-scale corpus marking process can be completed only by a small amount of correction, and the corpus marking cost is reduced; in addition, the corpus labeling and model optimization problems are used as a unified task to be iterated, so that the problems of manual intervention minimization, time consumption of data labeling and difficulty in model training in the generation process of the intention recognition model are solved.
Description
Technical Field
The invention belongs to the technical field of big data artificial intelligence, and particularly relates to a method, a device and equipment for constructing an intelligent recognition model of conversation intention.
Background
With the continuous development of social informatization and intellectualization, intelligent conversation applications such as intelligent customer service, intelligent assistant and chat robot based on natural language understanding are widely used. The intention recognition model is a core component of the application, and the quality of the intention recognition model has decisive influence on the application intellectualization degree and the user experience level.
At present, an intention recognition model is mainly trained in a supervision mode and needs to be marked with a large amount of linguistic data, and the application generally faces to massive internet users, so that the user intentions are complex and various and are easy to change along with time. The existing intention recognition model construction process mainly comprises a corpus labeling process and a model training process, wherein a large amount of labor and time are needed for corpus labeling and intention modification, and how to quickly, efficiently and inexpensively construct the intention recognition model is always a key point in application construction of the type. The existing method for constructing the intention recognition model mainly comprises the following steps: marking the field text data set by adopting a manual marking mode, and then carrying out model training; or the intention discovery and the intention corpus collection are carried out by a clustering method, but large-scale business data can obtain different theme categories by understanding from different angles, the themes discovered by clustering are often not needed by the business, and the method is far from meeting the actual business requirements; or training samples are generated by using a large amount of user selection data through a model prediction result as guidance, but the method uses standard samples to train an initial prediction model, and the standard samples are acquired by manual labeling.
Therefore, the existing method for constructing the intention recognition model has the problems of large amount of manual intervention, long time consumption of data annotation and difficult model training.
Disclosure of Invention
The embodiment of the invention aims to provide a construction method of an intelligent recognition model of conversation intention, and aims to solve the problems that the existing construction method of the intention recognition model needs a large amount of manual intervention, consumes much time for data annotation and is difficult in model training.
The embodiment of the invention is realized in such a way that a method for constructing an intelligent recognition model of conversation intention comprises the following steps:
obtaining corpus data, wherein the corpus data comprises labeled corpus data and unlabeled corpus data;
processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate intention corpus data with label information;
performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model; the initial intention recognition model is generated by the labeled corpus data through neural network training;
judging whether the iterative training meets a preset iterative ending condition or not; if not, returning to the step of obtaining the corpus data; if yes, the iterative training is ended.
Another objective of an embodiment of the present invention is to provide a device for constructing an intelligent recognition model of conversation intention, including:
the corpus data acquiring unit is used for acquiring corpus data, and the corpus data comprises labeled corpus data and unlabeled corpus data;
the intention corpus data generation unit is used for processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate intention corpus data with label information;
the iterative training unit is used for performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model; the initial intention recognition model is generated by the labeled corpus data through neural network training; and
the judging unit is used for judging whether the iterative training meets a preset iterative ending condition; if not, returning to the step of obtaining the corpus data; if yes, the iterative training is ended.
It is a further object of an embodiment of the present invention to provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, causes the processor to execute the steps of the construction method of the conversation intention intelligent recognition model.
Another object of an embodiment of the present invention is a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the processor to execute the steps of the construction method of the conversation intention intelligent recognition model.
According to the construction method of the intelligent recognition model of the conversation intention, provided by the embodiment of the invention, the semi-automatic labeling of the training corpus is realized by utilizing a large amount of non-labeled corpus data and based on a preset semantic analysis algorithm, the large-scale corpus labeling process can be completed only by a small amount of correction, and the corpus labeling cost is reduced; in addition, the corpus labeling and model optimization problems are used as a unified task to be iterated, so that the problems of manual intervention minimization, time consumption of data labeling and difficulty in model training in the generation process of the intention recognition model are solved.
Drawings
Fig. 1 is a flowchart illustrating an implementation of a method for constructing an intelligent recognition model of session intention according to an embodiment of the present invention;
FIG. 2 is a flowchart illustrating an implementation of another method for constructing an intelligent recognition model of conversation intention according to an embodiment of the present invention;
FIG. 3 is a flowchart of an implementation of a method for constructing a session intention intelligent recognition model according to another embodiment of the present invention;
FIG. 4 is a flowchart illustrating an implementation of a method for constructing a session intention intelligent recognition model according to another embodiment of the present invention;
fig. 5 is a flowchart illustrating an implementation of a method for constructing an intelligent recognition model of session intention according to an embodiment of the present invention;
FIG. 6 is a flowchart of an implementation of a method for constructing an optimized intelligent recognition model of conversation intention according to an embodiment of the present invention;
FIG. 7 is a flowchart of an implementation of another method for constructing an optimized intelligent recognition model of conversation intention according to an embodiment of the present invention;
fig. 8 is a block diagram of a device for constructing an intelligent recognition model of conversation intention according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The terminology used in the embodiments of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in the examples of the invention and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.
It should be understood that although the terms first, second, etc. may be used to describe various information in embodiments of the present invention, the information should not be limited by these terms. These terms are only used to distinguish one type of information from another.
As shown in fig. 1, in an embodiment, a method for constructing an intelligent recognition model of conversation intention is provided, which may specifically include the following steps:
step S101, corpus data is obtained, wherein the corpus data comprises labeled corpus data and unlabeled corpus data.
In the embodiment of the invention, the corpus data is obtained based on the existing big data analysis, and comprises corpus data of a corresponding business field and text corpus data of other fields or general fields, wherein the labeled corpus data can be artificially labeled corpus data or standard corpus data labeled for machine learning in the prior art; the unlabeled corpus data comprises unlabeled text corpus data accumulated by the existing business system and a large amount of unlabeled text corpus data in other fields or general fields.
And S102, processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate the intention corpus data with the label information.
In the embodiment of the invention, the labeled corpus data carries corresponding label information, and the label information comprises intention categories and corpus core words; the intention category is determined according to specific services, and the corpus core words refer to words and phrases capable of representing key subject information of the corpus sample and the like; the preset semantic analysis algorithm may be a combination of one or more text analysis algorithms, and can automatically analyze large-scale corpus data, and provide support for automatically labeling and recommending corpus data, including but not limited to keyword complete matching, document vocabulary weight vector similarity calculation, document semantic model vector similarity calculation, and the like, specifically including a corpus core word extraction algorithm, a semantic compression algorithm, a semantic matching algorithm, and a topic automatic analysis algorithm. The corpus core word extraction algorithm comprises a TF-IDF algorithm, a TextRank algorithm, an LDA algorithm, a machine learning classification model and the like. The semantic compression algorithm is a semantic model for inputting words or phrases and outputting semantic vectors; the construction of the semantic model can be obtained based on the unmarked corpus training or the supervised model training; when the unsupervised model is adopted for training, a Bert model and the like can be adopted; when supervised model training is adopted, output pre-bottleneck vectors of the classification model and the like can be adopted. The semantic matching algorithm is a model for inputting two text samples and outputting semantic similarity of the two text samples, and can be obtained by obtaining dense vector representation of a text by using a semantic model and then calculating by methods such as Euclidean distance and cosine similarity; semantic similarity refers to the close relationship of two texts in semantics, and the greater the semantic similarity, the closer the text semantics are. The topic automatic analysis algorithm is used for automatically performing topic analysis on corpus data through a clustering algorithm such as KMeans or a topic analysis model such as LDA and PLSA to obtain candidate topic categories. And when clustering is carried out by using a KMeans equidistance algorithm, the corpus distance is calculated through a semantic matching model.
In the embodiment of the invention, based on a preset semantic analysis algorithm, semantic extraction and analysis are firstly carried out on labeled corpus data, and labeling or recommendation labeling of unlabeled corpus data is automatically carried out by combining semantic information of the labeled corpus data under various scenes. The method comprises the following steps of recommending key corpora for marking, wherein the step of recommending key corpora for marking refers to the step of combining the existing marking result of the corpora and marking the core corpora data which are least relevant to the existing marked corpora and the purpose fuzzy corpora data in the unmarked corpora by comprehensive recommendation; and automatically labeling or recommending and labeling the unlabeled corpus data related to the intention core words in combination with the intention core word modification. And the corpus relevance judgment is calculated based on a semantic matching algorithm. The core corpus refers to the most representative corpus in the multiple pieces of unlabeled corpus data. The most representative corpus data in the plurality of unlabeled corpus data may be a central corpus after all corpora are clustered, and the corpus with the smallest distance to the clustering center point. The meaning fuzzy corpus in the unmarked corpus means that the corpus meaning relevancy is calculated based on a semantic matching algorithm and is obtained by comprehensively calculating all model relevancy. In addition, the automatic labeling or recommendation of the part of the corpus means that the corpus data related to the labeled corpus data in the unlabeled corpus data is automatically labeled into a corresponding category or recommended into a corresponding category to wait for confirmation. And calculating the relevance between each unmarked corpus and the current marked corpus data through a semantic matching algorithm in the unmarked corpus data. The automatic recommendation of the possible intention categories refers to that in intention management, unmarked corpora are clustered based on the current marking result, and the core category which is farthest away from the existing marked corpora is recommended as a candidate intention. The recommending of the core words related to the intentions means that in the management of the intentions, based on the current intention category and related labeled corpora, extraction is performed through a corpus core word extraction algorithm, and the extraction result is used as a corpus candidate core word to be recommended. The recommending of the intention core linguistic data is that semantic calculation and sequencing are carried out on intention core words and a large number of unlabeled linguistic data based on a semantic matching model, and the sample with the highest semantic similarity is used as candidate intention core data to carry out labeling or labeling recommendation automatically.
In a preferred embodiment of the present invention, the semantic compression algorithm is a model that can input words, phrases and sentences, convert the words, phrases and sentences through a neural network, and output the words, phrases and sentences as 200-dimensional floating point vectors. For example, the word "Beijing" is input, the semantic compression model outputs the result of [0.1,0.2,0.3 … 0.8.8 ], and the vector dimension is 200 dimensions. The neural network is constructed by adopting an RNN network and two layers of fully-connected networks; the output of the first layer of fully-connected network is a 200-dimensional vector, and the vector is used as a semantic compression model output vector; and outputting the classification result by the second layer of fully-connected network.
In a preferred embodiment of the present invention, the semantic compression model is implemented based on the Bert model using domain corpora retraining. The using field corpus is obtained by training again, namely the model is obtained by fine tuning the using field corpus on the basis of the general model. The user domain corpus fine adjustment means that adjacent sentences in the corpus are used as training corpora, and if the user is rained today, the temperature is a little low. Two sentences of 'and' more wearing clothes when going out for a moment 'appear in the corpus and are in the adjacent relation of upper and lower sentences, so that the sentence' rains today and the temperature is a little low. "as model input, train the sentence" a while going out and want the multi-point clothes "as model output.
In a preferred embodiment of the invention, the semantic compression model is realized based on a word2vec model and is obtained by training a neural network by adopting a CBOW mode. The CBOW network is a shallow layer network, the input is a compressed representation of t vocabularies, the output is a compressed representation of 1 vocabulary, the network is trained through summing and average output compressed representation expectation values and through loss of difference values of the current vocabulary compressed representation and the output compressed expectation values. And a back propagation algorithm is adopted when the network is trained through the difference loss.
In a preferred embodiment of the present invention, the preset semantic analysis algorithm includes a core word extraction algorithm and a semantic matching algorithm; as shown in fig. 2, the step 102 includes:
step 201, according to the corpus data and a preset core word extraction algorithm, extracting corpus core words of the labeled corpus data and the unlabeled corpus data respectively.
Step 202, calculating the similarity of the corpus core words of the labeled corpus data and the unlabeled corpus data according to the preset semantic matching algorithm.
And 203, generating the intention corpus data with the label information according to the labeled corpus data and the unlabeled corpus data with the similarity smaller than the preset distance threshold.
In the embodiment of the invention, the automatic extraction of the corpus core words can be realized by using a TF-IDF algorithm, and the TF-IDF value is calculated for all vocabularies of all documents of the corpus at first; then, sequencing the vocabularies in each corpus according to TF-IDF values, and taking the first 10 vocabularies as document core vocabularies; and finally, counting all document core vocabularies, and taking the first N as corpus core vocabularies. Specifically, the automatic extraction of the corpus core words is realized by the following algorithm, and the weight of the vocabulary is weighted
Wherein
WcFor obtaining universal weight of vocabulary by using mass universal linguistic data calculation, WdAnd calculating to obtain the vocabulary service weight for using the service corpus. Wherein, WcAnd WdIs composed of
Wherein, WiTF-IDF values calculated for the vocabulary in the ith corpus. And selecting the vocabulary with the maximum vocabulary weight W as the corpus core words.
In the embodiment of the invention, the similarity of the corpus core words can be obtained by calculating the cosine similarity. Given two words X, Y, when calculating their semantic similarity, they are first converted into vector representations V using a semantic compression algorithmx＝[vx,1...vx,100]And Vy＝[vy,1...vy,100](ii) a Then calculating the vector similarity through the cosine similarity, wherein the formula is
The closer the cosine similarity is to 1, the closer the semantic relationship is, and when used as a distance representation, the conversion can be to dx,y＝1-sx,yIn this case, the closer the semantic relationship between the two is, the smaller the distance between them is.
In a preferred embodiment of the present invention, the preset semantic analysis algorithm further includes a semantic compression algorithm; as shown in fig. 3, the step 102 further includes:
step 301, in response to the deleting operation of the tag information of the first labeled corpus data and/or the first intention corpus data, calculating the vector representation of the first labeled corpus data and/or the first intention corpus data according to the semantic compression algorithm.
Step 302, obtaining second labeled corpus data and/or second intention corpus data, the similarity of which to the vector representation of the first labeled corpus data and/or the first intention corpus data conforms to a preset similarity threshold, and deleting the label information of the second labeled corpus data and/or the second intention corpus data.
In the embodiment of the invention, deleting the tag information of the part of the corpus data is to cancel the label or recommend the label information of the part of the corpus data; specifically, after any corpus document is deleted with marked intentions, vector representation D of the corpus is calculated through a semantic compression algorithm; and then, calculating vector representation for all documents in the original intention of the document by using a semantic compression algorithm, then calculating the similarity of the documents by using a semantic matching algorithm, and finally automatically canceling the annotation or canceling the recommendation of the annotation intention for 100 documents with the closest semantic distance.
In a preferred embodiment of the present invention, as shown in fig. 4, the step 102 includes:
step 401, obtaining the intention center representation of the labeled corpus data.
And 402, acquiring a matrix representation of the unlabeled corpus data.
Step 403, calculating a distance value between the matrix representation of the unlabeled corpus data and the representation of the intention center according to the preset semantic analysis algorithm, and processing the unlabeled corpus data meeting a preset distance threshold according to the representation of the intention center to generate intention corpus data with label information.
In the embodiment of the invention, the method for automatically labeling or recommending the most relevant linguistic data by combining the currently labeled linguistic data comprises the following steps: firstly, acquiring an intention center representation T of labeled corpus data; then acquiring matrix representation of all the unmarked corpus data; and then calculating the distance D between all the unmarked corpora and the center of each intention through a semantic analysis algorithm, and finally marking or recommending and marking 100 corpora which are closest to the intention as automatic marking results. Wherein, the intention center representation T is calculated by all intention core word representations of the intention. The intention core words are all converted into 200-dimensional vector matrixes through a semantic analysis algorithm. For example, it is intended that the ith core word be represented as TWi＝{twi,1...twi,200}. Then intention TjIs represented by the center of (1) as Tj＝{tj,1...tj,200}. Wherein, each dimension is obtained by calculating the corresponding dimension of all the intention core words, and the calculation formula is
Wherein n is the number of the intention core words.
In a preferred embodiment of the present invention, corpus annotation is performed according to the intent ambiguity of the un-annotated corpus dataPreferably, the set of intent diagrams is T ═ T1...tnD, the labeled corpus in the ith intentioni＝{d1...dmThe set of unlabeled corpora is DU ═ d1...du}. Arbitrary unlabeled corpus diAnd any labeled corpus djA distance of li,jThen the corpus diWith any intention TjA distance of
Then corpus diWith an intention ambiguity of
Wherein max (LDT) is corpus diThe maximum of all the intended distances.
In a preferred embodiment of the present invention, as shown in fig. 5, the step 102 includes:
step 501, calculating the vector representation of the un-labeled corpus data and the intention center representation of the labeled corpus data according to the semantic analysis algorithm.
Step 502, calculating the distance between the vector representation of the un-labeled corpus data and the intention center representation of the labeled corpus data, and using the un-labeled corpus data with the distance greater than a preset distance threshold value as a candidate corpus data set.
Step 503, clustering the corpus candidate data set to obtain a plurality of candidate intentions of the unlabeled corpus data within a preset center range, and recommending the candidate intentions and the corresponding unlabeled corpus data as the corpus intended data.
In the embodiment of the present invention, the intention acquisition may be clustered by the KMeans algorithm. Firstly, randomly initializing K clustering cores; then, calculating the distances to all clustering cores by using the compressed representation of all the corpora, and dividing the corpora into the core classes with the minimum distances; then, using the corpus of each clustering core category to calculate a mean value as a category core; repeating the steps of calculating the distance and updating the core for N times until the clustering core is not changed any more. The recommendation mode of the intention corpus data can be that the intention center representation T of all intentions is calculated, and the unlabeled corpus is clustered by adopting a KMeans algorithm to obtain K candidate centers; calculating the distances between all unmarked corpora and K candidate centers, and taking a corpus with the minimum distance from each candidate center as a core corpus to obtain K core corpuses; calculating the distance between each core corpus and T intention centers, and taking the minimum distance as the core corpus distance to obtain K core corpus distances; taking M corpuses with the maximum core corpus distance as recommended marking corpuses; or may be: processing all the unmarked documents by using a semantic compression algorithm to obtain a vector representation DS, then calculating the current intention center representation T, further calculating the distance between all the documents in the DS and all the intention center representation T, and taking the documents with the distance larger than 0.7 as a candidate document set; then clustering the candidate document set based on a Kmeans algorithm to obtain K centers; and finally recommending the K centers as candidate intentions.
In a preferred embodiment of the present invention, as shown in fig. 6, after the step 102, the method further includes:
step 601, receiving the adjustment operation of the user on the label information to obtain the intention core word.
Step 602, generating an intention core corpus according to the intention core words, and recommending the intention core corpus as intention corpus data.
In the embodiment of the invention, when the fact that a user deletes, modifies or adds the label information of a certain corpus is received, the processing such as labeling, recommending labeling, deleting or canceling labeling and the like is automatically carried out on other similar corpora based on a semantic analysis algorithm and in combination with an adjustment result. The user can intervene the whole process in a browser interaction mode, for example, a software system based on a B \ S architecture, and a Server end automatically recommends feedback and marks after the user triggers an action; or the user can intervene with the server interactively in real time through the client based on the software system of the C \ S architecture, wherein the client software can be visual interface software or a command line type instruction interface.
In the embodiment of the present invention, the functions implemented by the user instruction and the feedback information include: intention category management, corpus management, model management, and the like. The intention category management comprises functions of creating intention, deleting intention, adding intention core vocabulary and deleting intention core vocabulary. The corpus management comprises uploading unmarked corpus, uploading marked corpus and confirming marked results. The model management comprises starting model training, checking a model evaluation result and setting model training parameters.
The method for constructing the intelligent recognition model of the conversation intention comprehensively utilizes machine learning and natural language understanding technologies, lays a foundation for a subsequent iterative model generation process by introducing an interactive feedback intervention means, can support business experts to participate in the whole process of constructing the intention recognition model, and integrally solves the problems of corpus labeling and model training. In addition, a large amount of non-labeled corpus data is combined and utilized, semi-automatic labeling or recommended labeling of the training corpus is realized based on a preset semantic analysis algorithm, the interaction efficiency is improved, and the minimum intervention operation is realized.
And S103, performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model.
In the embodiment of the present invention, an initial intention recognition model is generated by training the labeled corpus data through a neural network, specifically, an intention recognition model is generated by training based on a machine learning model and labeled corpus data of a given intention category and corpus core word, wherein the machine learning model is a supervised learning method, including but not limited to SVM, LSTM network, Transformer, and the like; when an intention recognition model is constructed by adopting neural network models such as LSTM or Transformer, a loss function is used for calculating a difference value between a model prediction intention and an actual annotation intention, model parameters are adjusted through a back propagation algorithm such as ADAM, so that the loss is minimized, and a training process is completed; wherein, the loss function can adopt MSE, cross entropy loss and the like; in addition, the model may be an existing model. Specifically, the intention recognition model is obtained by combining a Bert model, an LSTM model and a fully-connected network, a vector output by the Bert model is used as an input vector of the LSTM model, an output vector of the LSTM model is used as an input of the fully-connected network, and the output of the fully-connected network is an intention category; wherein the loss function uses a cross entropy loss function; the gradient updating strategy is carried out by adopting an ADAM algorithm; the final output intention category of the fully-connected network is a probability value of 0-1, and the closer to 1, the higher the confidence level.
In embodiments of the invention, iterative training can optimize the intent recognition model training process for a particular scenario; the specific scene refers to the fact that large-scale unmarked service corpus data and a small amount of marked service corpus data exist during training of the intention recognition model; the iterative training refers to performing multiple iterations by taking model construction and corpus expansion as a whole until a termination condition is met.
In the embodiment of the invention, the existing model is utilized to automatically label a large amount of unlabeled data, and the corpus data with high confidence coefficient in the automatic labeling result is added into the training corpus through a specific method for repeated training, so that the robustness of the model can be enhanced and the identification precision of the model can be improved after continuous multiple iterations; the existing model is a model obtained after the last iterative training is finished; the corpus expansion means that the existing model is used for predicting the unmarked corpus data, and the recognition result conforming to the expansion rule is added into the training corpus; the expansion rules include, but are not limited to, selection based on confidence threshold, selection based on ranking result, and the like; termination conditions include, but are not limited to, maximum number of iterations, minimum model adjustment amplitude, etc.; the maximum iteration times are times which cannot be exceeded at most by the iterative training method; in the repeated training, corpus data with high confidence coefficient in the automatic labeling result is added into training corpus by a specific method, the specific method is to screen through confidence coefficient, the screening method can be to add all data with confidence coefficient larger than 0.95 into the training data, and the confidence value can be set by expert adjustment; the training data can also be added to the first N data with confidence degrees larger than 0.95, wherein N can be a numerical value of 100, 1000 and the like, and specific numerical values can be specified by experts.
Step S104, judging whether the iterative training meets a preset iterative ending condition; if not, returning to the step S101; if yes, the process proceeds to step S105.
In the embodiment of the present invention, the iteration ending condition, i.e. the iteration training ending condition, includes, but is not limited to, the maximum iteration number, the minimum model adjustment amplitude, and the like; the maximum iteration number refers to the number of times that the iterative training method cannot exceed at most.
And step S105, finishing the iterative training.
According to the construction method of the intelligent recognition model of the conversation intention, provided by the embodiment of the invention, the semi-automatic labeling of the training corpus is realized by utilizing a large amount of non-labeled corpus data and based on a preset semantic analysis algorithm, the large-scale corpus labeling process can be completed only by a small amount of correction, and the corpus labeling cost is reduced; in addition, the corpus labeling and model optimization problems are used as a unified task to be iterated, so that the problems of manual intervention minimization, time consumption of data labeling and difficulty in model training in the generation process of the intention recognition model are solved.
In an embodiment, as shown in fig. 7, step S104 may specifically include the following steps:
step S701, obtaining assessment corpus data carrying corpus test tag information.
In the embodiment of the present invention, the evaluation corpus data is used to evaluate the intention recognition effect of the target intention recognition model.
And step S701, determining model intention labeling result information according to the evaluation corpus data and the target intention identification model.
Step S702, calculating the loss difference of the corpus test label information and the model intention standard result.
Step S703, determining whether the corpus test tag information and the loss difference of the model intention standard result satisfy a preset condition.
Step S704, when the loss difference between the corpus test label information and the model intention standard result meets a preset condition, ending iterative training;
step S705, when the loss difference between the corpus test tag information and the model intention standard result does not satisfy the preset condition, modifying the tag information of the intention corpus data, and returning to the step S103.
In the embodiment of the present invention, the target intention recognition model obtained by iterative training may be evaluated by various methods, including but not limited to randomly extracting N corpora from the unmarked corpora, evaluating accuracy after recognition by the model, constructing test corpora, and comparing results after recognition by the model. And after the model effect is evaluated, performing targeted adjustment on the model, modifying the label information of the intention corpus data, and performing adaptive adjustment on the intention category, the intention core word and the intention corpus in sequence to form new intention corpus data, returning to the step of performing iterative training on a preset initial intention recognition model according to the intention corpus data, and constructing a target intention recognition model.
As shown in fig. 8, in an embodiment, an apparatus for constructing an intelligent recognition model of conversation intention is provided, which specifically includes:
the corpus data acquiring unit 810 is configured to acquire corpus data, where the corpus data includes labeled corpus data and unlabeled corpus data.
In the embodiment of the invention, the corpus data is obtained based on the existing big data analysis, and comprises corpus data of a corresponding business field and text corpus data of other fields or general fields, wherein the labeled corpus data can be artificially labeled corpus data or standard corpus data labeled for machine learning in the prior art; the unlabeled corpus data comprises unlabeled text corpus data accumulated by the existing business system and a large amount of unlabeled text corpus data in other fields or general fields.
And an intention corpus data generating unit 820, configured to process the unlabeled corpus data according to the labeled corpus data and a preset semantic analysis algorithm, and generate intention corpus data with label information.
In the embodiment of the invention, the labeled corpus data carries corresponding label information, and the label information comprises intention categories and corpus core words; the intention category is determined according to specific services, and the corpus core words refer to words and phrases capable of representing key subject information of the corpus sample and the like; the preset semantic analysis algorithm may be a combination of one or more text analysis algorithms, and can automatically analyze large-scale corpus data, and provide support for automatically labeling and recommending corpus data, including but not limited to keyword complete matching, document vocabulary weight vector similarity calculation, document semantic model vector similarity calculation, and the like, specifically including a corpus core word extraction algorithm, a semantic compression algorithm, a semantic matching algorithm, and a topic automatic analysis algorithm. The corpus core word extraction algorithm comprises a TF-IDF algorithm, a TextRank algorithm, an LDA algorithm, a machine learning classification model and the like. The semantic compression algorithm is a semantic model for inputting words or phrases and outputting semantic vectors; the construction of the semantic model can be obtained based on the unmarked corpus training or the supervised model training; when the unsupervised model is adopted for training, a Bert model and the like can be adopted; when supervised model training is adopted, output pre-bottleneck vectors of the classification model and the like can be adopted. The semantic matching algorithm is a model for inputting two text samples and outputting semantic similarity of the two text samples, and can be obtained by obtaining dense vector representation of a text by using a semantic model and then calculating by methods such as Euclidean distance and cosine similarity; semantic similarity refers to the close relationship of two texts in semantics, and the greater the semantic similarity, the closer the text semantics are. The topic automatic analysis algorithm is used for automatically performing topic analysis on corpus data through a clustering algorithm such as KMeans or a topic analysis model such as LDA and PLSA to obtain candidate topic categories. And when clustering is carried out by using a KMeans equidistance algorithm, the corpus distance is calculated through a semantic matching model.
In the embodiment of the invention, based on a preset semantic analysis algorithm, semantic extraction and analysis are firstly carried out on labeled corpus data, and labeling or recommendation labeling of unlabeled corpus data is automatically carried out by combining semantic information of the labeled corpus data under various scenes. The method comprises the following steps of recommending key corpora for marking, wherein the step of recommending key corpora for marking refers to the step of combining the existing marking result of the corpora and marking the core corpora data which are least relevant to the existing marked corpora and the purpose fuzzy corpora data in the unmarked corpora by comprehensive recommendation; and automatically labeling or recommending and labeling the unlabeled corpus data related to the intention core words in combination with the intention core word modification. And the corpus relevance judgment is calculated based on a semantic matching algorithm. The core corpus refers to the most representative corpus in the multiple pieces of unlabeled corpus data. The most representative corpus data in the plurality of unlabeled corpus data may be a central corpus after all corpora are clustered, and the corpus with the smallest distance to the clustering center point. The meaning fuzzy corpus in the unmarked corpus means that the corpus meaning relevancy is calculated based on a semantic matching algorithm and is obtained by comprehensively calculating all model relevancy. In addition, the automatic labeling or recommendation of the part of the corpus means that the corpus data related to the labeled corpus data in the unlabeled corpus data is automatically labeled into a corresponding category or recommended into a corresponding category to wait for confirmation. And calculating the relevance between each unmarked corpus and the current marked corpus data through a semantic matching algorithm in the unmarked corpus data. The automatic recommendation of the possible intention categories refers to that in intention management, unmarked corpora are clustered based on the current marking result, and the core category which is farthest away from the existing marked corpora is recommended as a candidate intention. The recommending of the core words related to the intentions means that in the management of the intentions, based on the current intention category and related labeled corpora, extraction is performed through a corpus core word extraction algorithm, and the extraction result is used as a corpus candidate core word to be recommended. The recommending of the intention core linguistic data is that semantic calculation and sequencing are carried out on intention core words and a large number of unlabeled linguistic data based on a semantic matching model, and the sample with the highest semantic similarity is used as candidate intention core data to carry out labeling or labeling recommendation automatically.
An iterative training unit 830, configured to perform iterative training on a preset initial intention recognition model according to the intention corpus data, and construct a target intention recognition model; the initial intention recognition model is generated through the labeled corpus data through neural network training.
In the embodiment of the present invention, an initial intention recognition model is generated by training the labeled corpus data through a neural network, specifically, an intention recognition model is generated by training based on a machine learning model and labeled corpus data of a given intention category and corpus core word, wherein the machine learning model is a supervised learning method, including but not limited to SVM, LSTM network, Transformer, and the like; when an intention recognition model is constructed by adopting neural network models such as LSTM or Transformer, a loss function is used for calculating a difference value between a model prediction intention and an actual annotation intention, model parameters are adjusted through a back propagation algorithm such as ADAM, so that the loss is minimized, and a training process is completed; wherein, the loss function can adopt MSE, cross entropy loss and the like; in addition, the model may be an existing model. Specifically, the intention recognition model is obtained by combining a Bert model, an LSTM model and a fully-connected network, a vector output by the Bert model is used as an input vector of the LSTM model, an output vector of the LSTM model is used as an input of the fully-connected network, and the output of the fully-connected network is an intention category; wherein the loss function uses a cross entropy loss function; the gradient updating strategy is carried out by adopting an ADAM algorithm; the final output intention category of the fully-connected network is a probability value of 0-1, and the closer to 1, the higher the confidence level.
In embodiments of the invention, iterative training can optimize the intent recognition model training process for a particular scenario; the specific scene refers to the fact that large-scale unmarked service corpus data and a small amount of marked service corpus data exist during training of the intention recognition model; the iterative training refers to performing multiple iterations by taking model construction and corpus expansion as a whole until a termination condition is met.
In the embodiment of the invention, the existing model is utilized to automatically label a large amount of unlabeled data, and the corpus data with high confidence coefficient in the automatic labeling result is added into the training corpus through a specific method for repeated training, so that the robustness of the model can be enhanced and the identification precision of the model can be improved after continuous multiple iterations; the existing model is a model obtained after the last iterative training is finished; the corpus expansion means that the existing model is used for predicting the unmarked corpus data, and the recognition result conforming to the expansion rule is added into the training corpus; the expansion rules include, but are not limited to, selection based on confidence threshold, selection based on ranking result, and the like; termination conditions include, but are not limited to, maximum number of iterations, minimum model adjustment amplitude, etc.; the maximum iteration times are times which cannot be exceeded at most by the iterative training method; in the repeated training, corpus data with high confidence coefficient in the automatic labeling result is added into training corpus by a specific method, the specific method is to screen through confidence coefficient, the screening method can be to add all data with confidence coefficient larger than 0.95 into the training data, and the confidence value can be set by expert adjustment; the training data can also be added to the first N data with confidence degrees larger than 0.95, wherein N can be a numerical value of 100, 1000 and the like, and specific numerical values can be specified by experts.
A judging unit 840, configured to judge whether the iterative training satisfies a preset iterative end condition; if not, returning to the step of obtaining the corpus data; if yes, the iterative training is ended.
In the embodiment of the present invention, the iteration ending condition, i.e. the iteration training ending condition, includes, but is not limited to, the maximum iteration number, the minimum model adjustment amplitude, and the like; the maximum iteration number refers to the number of times that the iterative training method cannot exceed at most.
According to the device for constructing the intelligent recognition model of the conversation intention, provided by the embodiment of the invention, the semi-automatic labeling of the training corpus is realized by utilizing a large amount of non-labeled corpus data and based on a preset semantic analysis algorithm, the large-scale corpus labeling process can be completed only by a small amount of correction, and the corpus labeling cost is reduced; in addition, the corpus labeling and model optimization problems are used as a unified task to be iterated, so that the problems of manual intervention minimization, time consumption of data labeling and difficulty in model training in the generation process of the intention recognition model are solved.
In one embodiment, a computer device is proposed, the computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the following steps when executing the computer program:
obtaining corpus data, wherein the corpus data comprises labeled corpus data and unlabeled corpus data;
processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate intention corpus data with label information;
performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model; the initial intention recognition model is generated by the labeled corpus data through neural network training;
judging whether the iterative training meets a preset iterative ending condition or not; if not, returning to the step of obtaining the corpus data; if yes, the iterative training is ended.
In one embodiment, a computer readable storage medium is provided, having a computer program stored thereon, which, when executed by a processor, causes the processor to perform the steps of:
obtaining corpus data, wherein the corpus data comprises labeled corpus data and unlabeled corpus data;
processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate intention corpus data with label information;
performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model; the initial intention recognition model is generated by the labeled corpus data through neural network training;
judging whether the iterative training meets a preset iterative ending condition or not; if not, returning to the step of obtaining the corpus data; if yes, the iterative training is ended.
It should be understood that, although the steps in the flowcharts of the embodiments of the present invention are shown in sequence as indicated by the arrows, the steps are not necessarily performed in sequence as indicated by the arrows. The steps are not performed in the exact order shown and described, and may be performed in other orders, unless explicitly stated otherwise. Moreover, at least a portion of the steps in various embodiments may include multiple sub-steps or multiple stages that are not necessarily performed at the same time, but may be performed at different times, and the order of performance of the sub-steps or stages is not necessarily sequential, but may be performed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.
It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by a computer program, which can be stored in a non-volatile computer-readable storage medium, and can include the processes of the embodiments of the methods described above when the program is executed. Any reference to memory, storage, database, or other medium used in the embodiments provided herein may include non-volatile and/or volatile memory, among others. Non-volatile memory can include read-only memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus Direct RAM (RDRAM), direct bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
The technical features of the embodiments described above may be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the embodiments described above are not described, but should be considered as being within the scope of the present specification as long as there is no contradiction between the combinations of the technical features.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.
Claims (10)
1. A construction method of a conversation intention intelligent recognition model is characterized by comprising the following steps:
obtaining corpus data, wherein the corpus data comprises labeled corpus data and unlabeled corpus data;
processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate intention corpus data with label information;
performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model; the initial intention recognition model is generated by the labeled corpus data through neural network training;
judging whether the iterative training meets a preset iterative ending condition or not; if not, returning to the step of obtaining the corpus data; if yes, the iterative training is ended.
2. The construction method of the intelligent conversation intention recognition model according to claim 1, wherein the preset semantic analysis algorithm comprises a core word extraction algorithm and a semantic matching algorithm;
the step of processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate the intention corpus data with the label information comprises the following steps:
according to the corpus data and a preset core word extraction algorithm, corpus core words of the labeled corpus data and the unlabeled corpus data are respectively extracted;
calculating the similarity of the corpus core words of the labeled corpus data and the unlabeled corpus data according to the preset semantic matching algorithm;
and generating the intention corpus data with the label information according to the labeled corpus data and the unlabeled corpus data with the similarity smaller than the preset distance threshold.
3. The construction method of the intelligent recognition model of conversation intention according to claim 2, characterized in that the preset semantic analysis algorithm further comprises a semantic compression algorithm;
the step of processing the unlabeled corpus data according to the labeled corpus data and a preset semantic analysis algorithm to generate the intention corpus data with the label information further includes:
responding to the deleting operation of the label information of the first labeled corpus data and/or the first intention corpus data, and calculating the vector representation of the first labeled corpus data and/or the first intention corpus data according to the semantic compression algorithm;
and acquiring second labeled corpus data and/or second intention corpus data, the similarity of which is expressed by the vector of the first labeled corpus data and/or the first intention corpus data and accords with a preset similarity threshold, and deleting the label information of the second labeled corpus data and/or the second intention corpus data.
4. The method for constructing a conversation intention intelligent recognition model according to claim 1, wherein the step of processing the unlabeled corpus data according to the labeled corpus data and a preset semantic analysis algorithm to generate the intention corpus data with labeled information comprises:
acquiring the intention center representation of the labeled corpus data;
acquiring matrix representation of the unmarked corpus data;
and calculating a distance value between the matrix representation of the unmarked corpus data and the expression of the intention center according to the preset semantic analysis algorithm, and processing the unmarked corpus data which accords with a preset distance threshold value according to the expression of the intention center to generate the intention corpus data with label information.
5. The method for constructing a conversation intention intelligent recognition model according to claim 1, wherein the step of processing the unlabeled corpus data according to the labeled corpus data and a preset semantic analysis algorithm to generate the intention corpus data with labeled information comprises:
calculating the vector representation of the unmarked corpus data and the intention center representation of the marked corpus data according to the semantic analysis algorithm;
calculating the distance between the vector representation of the unmarked corpus data and the distance represented by the intention center of the marked corpus data, and taking the unmarked corpus data of which the distance is greater than a preset distance threshold value as a candidate corpus data set;
clustering the candidate corpus data set to obtain a plurality of candidate intentions of the unlabeled corpus data in a preset center range, and recommending the candidate intentions and the corresponding unlabeled corpus data as intention corpus data.
6. The method for constructing a conversation intention intelligent recognition model according to claim 1, wherein after the step of processing the unlabeled corpus data according to the labeled corpus data and a preset semantic analysis algorithm to generate the intention corpus data with labeled information, the method further comprises:
receiving the adjustment operation of the user on the label information to obtain an intention core word;
and generating an intention core corpus according to the intention core words, and recommending the intention core corpus as intention corpus data.
7. The method for constructing the intelligent recognition model of conversation intention according to claim 1, wherein the judging whether the iterative training satisfies a preset iteration ending condition is performed; if not, returning to the step of obtaining the corpus data; if yes, the step of finishing the iterative training comprises the following steps:
obtaining assessment corpus data carrying corpus test tag information;
determining model intention labeling result information according to the evaluation corpus data and the target intention identification model;
calculating the loss difference of the corpus test label information and the model intention standard result;
judging whether the loss difference between the corpus test label information and the model intention standard result meets a preset condition or not;
when the loss difference between the corpus test label information and the model intention standard result meets a preset condition, ending iterative training;
and when the loss difference between the corpus test label information and the model intention standard result does not meet a preset condition, modifying the label information of the intention corpus data, returning to the step of performing iterative training on a preset initial intention recognition model according to the intention corpus data, and constructing a target intention recognition model.
8. An apparatus for constructing a conversation intention intelligent recognition model, comprising:
the corpus data acquiring unit is used for acquiring corpus data, and the corpus data comprises labeled corpus data and unlabeled corpus data;
the intention corpus data generation unit is used for processing the unmarked corpus data according to the marked corpus data and a preset semantic analysis algorithm to generate intention corpus data with label information;
the iterative training unit is used for performing iterative training on a preset initial intention recognition model according to the intention corpus data to construct a target intention recognition model; the initial intention recognition model is generated by the labeled corpus data through neural network training; and
the judging unit is used for judging whether the iterative training meets a preset iterative ending condition; if not, returning to the step of obtaining the corpus data; if yes, the iterative training is ended.
9. A computer device comprising a memory and a processor, the memory having stored thereon a computer program that, when executed by the processor, causes the processor to perform the steps of the method of constructing a conversational intent intelligent recognition model according to any one of claims 1 to 7.
10. A computer-readable storage medium, having stored thereon a computer program which, when executed by a processor, causes the processor to perform the steps of the method for constructing a conversation intention intelligent recognition model according to any one of claims 1 to 7.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010968430.8A CN112131890A (en) | 2020-09-15 | 2020-09-15 | Method, device and equipment for constructing intelligent recognition model of conversation intention |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010968430.8A CN112131890A (en) | 2020-09-15 | 2020-09-15 | Method, device and equipment for constructing intelligent recognition model of conversation intention |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112131890A true CN112131890A (en) | 2020-12-25 |
Family
ID=73845149
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010968430.8A Pending CN112131890A (en) | 2020-09-15 | 2020-09-15 | Method, device and equipment for constructing intelligent recognition model of conversation intention |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112131890A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112667811A (en) * | 2020-12-29 | 2021-04-16 | 中国平安人寿保险股份有限公司 | Corpus annotation correction method and device, terminal equipment and medium |
| CN113160800A (en) * | 2021-02-04 | 2021-07-23 | 吉林科讯信息科技有限公司 | Voice data labeling method, electronic equipment and storage device |
| CN113343709A (en) * | 2021-06-22 | 2021-09-03 | 北京三快在线科技有限公司 | Method for training intention recognition model, method, device and equipment for intention recognition |
| CN113792540A (en) * | 2021-09-18 | 2021-12-14 | 平安科技（深圳）有限公司 | Intention recognition model updating method and related equipment |
| CN114372446A (en) * | 2021-12-13 | 2022-04-19 | 北京五八信息技术有限公司 | Vehicle attribute labeling method, device and storage medium |
| CN114637848A (en) * | 2022-03-15 | 2022-06-17 | 美的集团(上海)有限公司 | Semantic classification method and device |
| WO2022142006A1 (en) * | 2020-12-30 | 2022-07-07 | 平安科技（深圳）有限公司 | Semantic recognition-based verbal skill recommendation method and apparatus, device, and storage medium |
| CN114818703A (en) * | 2022-06-28 | 2022-07-29 | 珠海金智维信息科技有限公司 | Multi-intention recognition method and system based on BERT language model and TextCNN model |
| WO2022213864A1 (en) * | 2021-04-06 | 2022-10-13 | 华为云计算技术有限公司 | Corpus annotation method and apparatus, and related device |
| CN116485450A (en) * | 2023-06-21 | 2023-07-25 | 江苏网进科技股份有限公司 | Automatic demand identification method |
| CN117392983A (en) * | 2023-12-08 | 2024-01-12 | 四川霍尼尔电气技术有限公司 | Intelligent voice interaction control system and control method for drilling machine |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107832425A (en) * | 2017-11-13 | 2018-03-23 | 北京神州泰岳软件股份有限公司 | A kind of corpus labeling method, the apparatus and system of more wheel iteration |
| CN111563208A (en) * | 2019-01-29 | 2020-08-21 | 株式会社理光 | Intention identification method and device and computer readable storage medium |
| CN111597336A (en) * | 2020-05-14 | 2020-08-28 | 腾讯科技（深圳）有限公司 | Processing method and device of training text, electronic equipment and readable storage medium |
-
2020
- 2020-09-15 CN CN202010968430.8A patent/CN112131890A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107832425A (en) * | 2017-11-13 | 2018-03-23 | 北京神州泰岳软件股份有限公司 | A kind of corpus labeling method, the apparatus and system of more wheel iteration |
| CN111563208A (en) * | 2019-01-29 | 2020-08-21 | 株式会社理光 | Intention identification method and device and computer readable storage medium |
| CN111597336A (en) * | 2020-05-14 | 2020-08-28 | 腾讯科技（深圳）有限公司 | Processing method and device of training text, electronic equipment and readable storage medium |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112667811A (en) * | 2020-12-29 | 2021-04-16 | 中国平安人寿保险股份有限公司 | Corpus annotation correction method and device, terminal equipment and medium |
| CN112667811B (en) * | 2020-12-29 | 2024-03-08 | 中国平安人寿保险股份有限公司 | Corpus labeling correction method, corpus labeling correction device, terminal equipment and medium |
| WO2022142006A1 (en) * | 2020-12-30 | 2022-07-07 | 平安科技（深圳）有限公司 | Semantic recognition-based verbal skill recommendation method and apparatus, device, and storage medium |
| CN113160800A (en) * | 2021-02-04 | 2021-07-23 | 吉林科讯信息科技有限公司 | Voice data labeling method, electronic equipment and storage device |
| CN113160800B (en) * | 2021-02-04 | 2022-12-06 | 吉林科讯信息科技有限公司 | Voice data labeling method, electronic equipment and storage device |
| WO2022213864A1 (en) * | 2021-04-06 | 2022-10-13 | 华为云计算技术有限公司 | Corpus annotation method and apparatus, and related device |
| CN113343709A (en) * | 2021-06-22 | 2021-09-03 | 北京三快在线科技有限公司 | Method for training intention recognition model, method, device and equipment for intention recognition |
| WO2023040153A1 (en) * | 2021-09-18 | 2023-03-23 | 平安科技（深圳）有限公司 | Method, apparatus, and device for updating intent recognition model, and readable medium |
| CN113792540A (en) * | 2021-09-18 | 2021-12-14 | 平安科技（深圳）有限公司 | Intention recognition model updating method and related equipment |
| CN113792540B (en) * | 2021-09-18 | 2024-03-22 | 平安科技（深圳）有限公司 | Method for updating intention recognition model and related equipment |
| CN114372446A (en) * | 2021-12-13 | 2022-04-19 | 北京五八信息技术有限公司 | Vehicle attribute labeling method, device and storage medium |
| CN114637848A (en) * | 2022-03-15 | 2022-06-17 | 美的集团(上海)有限公司 | Semantic classification method and device |
| CN114818703A (en) * | 2022-06-28 | 2022-07-29 | 珠海金智维信息科技有限公司 | Multi-intention recognition method and system based on BERT language model and TextCNN model |
| CN116485450A (en) * | 2023-06-21 | 2023-07-25 | 江苏网进科技股份有限公司 | Automatic demand identification method |
| CN116485450B (en) * | 2023-06-21 | 2023-08-29 | 江苏网进科技股份有限公司 | Automatic demand identification method |
| CN117392983A (en) * | 2023-12-08 | 2024-01-12 | 四川霍尼尔电气技术有限公司 | Intelligent voice interaction control system and control method for drilling machine |
| CN117392983B (en) * | 2023-12-08 | 2024-02-20 | 四川霍尼尔电气技术有限公司 | Intelligent voice interaction control system and control method for drilling machine |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112131890A (en) | Method, device and equipment for constructing intelligent recognition model of conversation intention | |
| CN110609891B (en) | Visual dialog generation method based on context awareness graph neural network | |
| US20230100376A1 (en) | Text sentence processing method and apparatus, computer device, and storage medium | |
| CN111753060B (en) | Information retrieval method, apparatus, device and computer readable storage medium | |
| CN108733792B (en) | Entity relation extraction method | |
| CN107798140B (en) | Dialog system construction method, semantic controlled response method and device | |
| CN111783474B (en) | Comment text viewpoint information processing method and device and storage medium | |
| CN110969020A (en) | CNN and attention mechanism-based Chinese named entity identification method, system and medium | |
| CN109918663A (en) | A kind of semantic matching method, device and storage medium | |
| Arsov et al. | Network embedding: An overview | |
| CN112052684A (en) | Named entity identification method, device, equipment and storage medium for power metering | |
| CN113220886A (en) | Text classification method, text classification model training method and related equipment | |
| JP6738769B2 (en) | Sentence pair classification device, sentence pair classification learning device, method, and program | |
| CN111027292B (en) | Method and system for generating limited sampling text sequence | |
| CN114780831A (en) | Sequence recommendation method and system based on Transformer | |
| CN111985228A (en) | Text keyword extraction method and device, computer equipment and storage medium | |
| CN113255366B (en) | Aspect-level text emotion analysis method based on heterogeneous graph neural network | |
| Dai et al. | Hybrid deep model for human behavior understanding on industrial internet of video things | |
| CN111046178A (en) | Text sequence generation method and system | |
| CN116304020A (en) | Industrial text entity extraction method based on semantic source analysis and span characteristics | |
| CN117132923A (en) | Video classification method, device, electronic equipment and storage medium | |
| CN116842934A (en) | Multi-document fusion deep learning title generation method based on continuous learning | |
| CN116881689A (en) | Knowledge-enhanced user multi-mode online comment quality evaluation method and system | |
| CN109033413B (en) | Neural network-based demand document and service document matching method | |
| CN117216281A (en) | Knowledge graph-based user interest diffusion recommendation method and system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20201225 |