WO2023065544A1

WO2023065544A1 - Intention classification method and apparatus, electronic device, and computer-readable storage medium

Info

Publication number: WO2023065544A1
Application number: PCT/CN2022/071077
Authority: WO
Inventors: 舒畅; 陈又新
Original assignee: 平安科技（深圳）有限公司
Priority date: 2021-10-18
Filing date: 2022-01-10
Publication date: 2023-04-27
Also published as: CN113792818A; CN113792818B

Abstract

Embodiments of the present application provide an intention classification method and apparatus, an electronic device, and a computer-readable storage medium, and relate to the technical field of deep learning in artificial intelligence. The method comprises: acquiring request text; extracting entity features from the request text, and obtain first text comprising a target query parameter; inputting the first text into a pre-trained comparison model to perform matrix multiplication with a reference word embedding matrix in the comparison model, and obtain multiple target word embedding vectors; performing classification on the target word embedding vectors by using a pre-trained intention classification model, and obtain an intention classification probability value and a target word embedding vector comprising an intention category label; performing matching on the first text by using a pre-trained intention matching model, and obtain an intention matching value; and obtaining intention classification data according to the intention matching value and the intention classification probability value. The embodiments of the present invention can achieve the accurate classification of user intention and improve the accuracy of intention classification.

Description

Intent Classification Method, Device, Electronic Device, and Computer-Readable Storage Medium

This application claims the priority of the Chinese patent application with the application number 202111212210.3 submitted to the China Patent Office on October 18, 2021, and the title of the invention is "Intention Classification Method, Device, Electronic Equipment, and Computer-Readable Storage Medium", the entire content of which Incorporated in this application by reference.

technical field

The present application relates to the technical field of deep learning in artificial intelligence, and in particular to an intent classification method, device, electronic equipment, and computer-readable storage medium.

Background technique

In natural language understanding, user intent needs to be classified. At present, intent classification is usually based on templates or models, but the inventors realized that model-based intent classification is easily affected by the occurrence frequency and data volume of intent, and often cannot solve the intent classification in real scenarios well, affecting the accuracy of intent classification. accuracy. Therefore, how to improve the accuracy of intent classification has become an urgent technical problem to be solved.

Contents of the invention

The main purpose of the embodiments of the present application is to provide an intent classification method, device, electronic device, and computer-readable storage medium, aiming at realizing accurate classification of user intent and improving the accuracy of intent classification.

In order to achieve the above purpose, the first aspect of the embodiments of the present application proposes an intention classification method, the method includes:

get request text;

performing entity feature extraction on the request text to obtain the first text containing target query parameters;

The first text is input to the pre-trained comparison model and the reference word embedding matrix in the comparison model is multiplied by matrix to obtain a plurality of target word embedding vectors;

Classify the target word embedding vector by using the pre-trained intent classification model to obtain the target word embedding vector and the intent classification probability value including the intent category label;

performing matching processing on the first text by using a pre-trained intent matching model to obtain an intent matching value;

According to the intention matching value and the intention classification probability value, the intention classification data is obtained.

In order to achieve the above purpose, the second aspect of the embodiments of the present application proposes an intention classification device, the device includes:

A text acquisition module, configured to acquire the request text;

A feature extraction module, configured to extract entity features from the request text to obtain the first text containing target query parameters;

Comparison module, for inputting the first text to the pre-trained comparison model and performing matrix multiplication with the reference word embedding matrix in the comparison model, to obtain a plurality of target word embedding vectors;

A classification module, configured to classify the target word embedding vector using a pre-trained intent classification model, to obtain a target word embedding vector and an intent classification probability value including an intent category label;

A matching module, configured to use a pre-trained intent matching model to perform matching processing on the first text to obtain an intent matching value;

A calculation module, configured to obtain intention classification data according to the intention matching value and the intention classification probability value.

To achieve the above object, the third aspect of the embodiments of the present application proposes an electronic device, the electronic device includes a memory, a processor, a program stored in the memory and executable on the processor, and a program for A data bus for connection and communication between the processor and the memory is implemented, and when the program is executed by the processor, an intent classification method is implemented, wherein the intent classification method includes:

get request text;

To achieve the above purpose, the fourth aspect of the embodiments of the present application proposes a computer-readable storage medium for computer-readable storage, the computer-readable storage medium stores one or more programs, and the one or more This program can be executed by one or more processors to implement a method for classifying intent, wherein the method for classifying intent includes the following steps:

get request text;

The intent classification method, device, electronic equipment, and computer-readable storage medium proposed in this application obtain the request text, extract the entity features of the request text, and obtain the first text containing the target query parameters. This method can realize the request The feature extraction of the text reduces the data space of the requested text, making it easier to extract the required first text containing the target query parameters. Then input the first text to the pre-trained comparison model and perform matrix multiplication with the reference word embedding matrix in the comparison model to obtain multiple target word embedding vectors, and use the pre-trained intent classification model to classify the target word embedding vectors, The target word embedding vector and intent classification probability value containing the intent category label are obtained, and the problem of uneven distribution of the target word embedding vector can be better solved by comparing the model. At the same time, the depth of the intent classification probability can be determined by comparing the model and the intent classification model. Learning to improve the accuracy of intent classification probability values. In addition, the present application can also use the pre-trained intention matching model to perform matching processing on the first text to obtain the intention matching value, and the intention matching model can be used to calculate the user's intention matching value based on rule matching to improve the accuracy of intention matching. Finally, according to the intention matching value and the intention classification probability value, the intention classification data is obtained. In this application, through the comparison model, intent classification model, and intent matching model, the user's dialogue intent can be identified by integrating the two aspects of intent classification probability and intent matching, so that the final intent classification data can present more accurate intent classification results. The accurate classification of user intent is achieved, and the accuracy of intent classification is improved.

Description of drawings

FIG. 1 is a flow chart of an intent classification method provided in an embodiment of the present application;

Fig. 2 is the flowchart of step S102 in Fig. 1;

Fig. 3 is the flowchart of step S103 in Fig. 1;

Fig. 4 is another flow chart of the intent classification method provided by the embodiment of the present application;

Fig. 5 is the flowchart of step S104 in Fig. 1;

Fig. 6 is the flowchart of step S105 in Fig. 1;

Fig. 7 is the flowchart of step S106 in Fig. 1;

Fig. 8 is a schematic structural diagram of an intention classification device provided by an embodiment of the present application;

FIG. 9 is a schematic diagram of a hardware structure of an electronic device provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present application, not to limit the present application.

It should be noted that although the functional modules are divided in the schematic diagram of the device, and the logical sequence is shown in the flowchart, in some cases, it can be executed in a different order than the module division in the device or the flowchart in the flowchart. steps shown or described. The terms "first", "second" and the like in the specification and claims and the above drawings are used to distinguish similar objects, and not necessarily used to describe a specific sequence or sequence.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application, and are not intended to limit the present application.

First, analyze some nouns involved in this application:

Artificial Intelligence (AI): It is a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence; artificial intelligence is a branch of computer science. Intelligence attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a manner similar to human intelligence. Research in this field includes robotics, language recognition, image recognition, natural language processing, and expert systems. Artificial intelligence can simulate the information process of human consciousness and thinking. Artificial intelligence is also a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

Natural language processing (NLP): NLP uses computers to process, understand and use human languages (such as Chinese, English, etc.). NLP belongs to a branch of artificial intelligence and is an interdisciplinary subject between computer science and linguistics. Known as computational linguistics. Natural language processing includes syntax analysis, semantic analysis, text understanding, etc. Natural language processing is often used in technical fields such as machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, information retrieval, information extraction and filtering, text classification and clustering, public opinion analysis and opinion mining. It involves language processing Related data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research and linguistics research related to language computing, etc.

Information Extraction (Information Extraction, NER): A text processing technology that extracts specified types of factual information such as entities, relationships, and events from natural language texts, and forms structured data output. Information extraction is a technique to extract specific information from text data. Text data is composed of some specific units, such as sentences, paragraphs, and chapters. Text information is composed of some small specific units, such as words, words, phrases, sentences, paragraphs, or combinations of these specific units. . Extracting noun phrases, personal names, and place names in text data is all text information extraction. Of course, the information extracted by text information extraction technology can be various types of information.

Entity: Refers to something that is distinguishable and exists independently. Such as a certain person, a certain city, a certain plant, a certain commodity, etc. Everything in the world is made up of concrete things, which refer to entities. Entities are the most basic elements in knowledge graphs, and different entities have different relationships.

Concept: A collection of entities of a certain type.

Semantic class (concept): A collection of entities with the same characteristics, such as countries, nations, books, computers, etc. Concepts mainly refer to collections, categories, object types, and types of things, such as people, geography, etc.

Self-supervised learning: Self-supervised learning mainly uses auxiliary tasks (pretext) to mine its own supervision information from large-scale unsupervised data, and trains the network through this structured supervision information, so that it can learn to be valuable for downstream tasks representation. That is to say, the supervision information of self-supervised learning is not manually labeled, but the algorithm automatically constructs supervision information in large-scale unsupervised data for supervised learning or training.

Contrastive learning is a kind of self-supervised learning, which does not need to rely on manually labeled category label information, and directly uses the data itself as supervisory information. Contrastive learning is an approach to the task of describing similar and dissimilar things for deep learning models. Using contrastive learning methods, machine learning models can be trained to distinguish between similar and dissimilar images. Self-supervised learning in the image field is divided into two types: generative self-supervised learning and discriminative self-supervised learning. Contrastive learning applies typical discriminative self-supervised learning. The core point of comparative learning is: by automatically constructing similar instances and dissimilar instances, that is, positive samples and negative samples, learning to compare positive samples and negative samples in the feature space, so that similar instances are closer in the feature space, The distance between dissimilar instances in the feature space is farther, and the difference becomes larger. The model representation obtained through such a learning process can perform downstream tasks and fine-tune on a smaller labeled data set to achieve unsupervised model learning process. The guiding principle of comparative learning is: by automatically constructing similar instances and dissimilar instances, a learning model is obtained through learning, and using this model, similar instances are relatively close in the projection space, while dissimilar instances can be compared in the projection space. Far.

Embedding: embedding is a kind of vector representation, which refers to representing an object with a low-dimensional vector, which can be a word, or a commodity, or a movie, etc.; the nature of this embedding vector is that it can Make objects corresponding to vectors with similar distances have similar meanings. Embedding is essentially a mapping from semantic space to vector space, while maintaining the relationship of the original sample in the semantic space in the vector space as much as possible, such as two semantically close The positions of the words in the vector space are also relatively close. Embedding can encode an object with a low-dimensional vector and retain its meaning. It is often used in machine learning. In the process of building a machine learning model, the object is encoded as a low-dimensional dense vector and then passed to DNN to improve efficiency.

Batch: Batch size (i.e., batch size) is a hyperparameter used to define the number of samples to be processed before updating the internal model parameters, that is, to control the number of training samples before the internal parameters of the model are updated. The training data set can be divided into one or more batches, where when all training samples are used to create a batch, the learning algorithm is called batch gradient descent; when the batch is the size of a sample, the learning algorithm is called stochastic gradient descent; When the batch size is more than one sample and less than the size of the training dataset, the learning algorithm is called mini-batch gradient descent. The batch size is the number of samples processed before updating the model.

Data enhancement: Data enhancement is mainly used to prevent overfitting and optimize the dataset when the dataset is small. Through data enhancement, the amount of training data can be increased, the generalization ability of the model can be improved, and noise data can be increased. Improve the robustness of the model. Data enhancement can be divided into two categories, offline enhancement and online enhancement. Among them, offline enhancement is to directly process the data set, and the number of data will become the enhancement factor x the number of original data sets. Offline enhancement is often used when the data set is small. ;Online enhancement is mainly used to enhance the batch data after obtaining the batch data, such as rotation, translation, flipping and other corresponding changes. Since some data sets cannot accept linear level growth, online enhancement is often used for larger data Set, many machine learning frameworks already support online enhancement methods, and can use GPU to optimize calculations.

Dropout (discard): dropout is a technique to prevent model overfitting. It means that during the training process of the deep learning network, for the neural network unit, it is temporarily discarded from the network according to a certain probability, so that the model can be more accurate. Robust, because it does not depend too much on some local features (because local features may be discarded).

Mask (mask, mask): mask is a common operation in deep learning; in simple terms, mask is equivalent to putting a mask on the original tensor to shield or select some specific elements, so it is often used to construct tensors volume filter. The linear activation function Relu (simple and rough dichotomy based on the positive and negative range of the output) and the dropout mechanism (division according to the probability) can be understood as generalized mask operations.

Encoder: encoding is to convert the input sequence into a fixed-length vector; decoding (decoder) is to convert the previously generated fixed vector into an output sequence; where the input sequence can be text, voice, image, video; output sequence Can be text, image.

Backpropagation: The general principle of backpropagation is: input the training set data into the input layer of the neural network, pass through the hidden layer of the neural network, and finally reach the output layer of the neural network and output the result; because the output result of the neural network is different from the actual If there is an error in the result, the error between the estimated value and the actual value is calculated, and the error is backpropagated from the output layer to the hidden layer until it is propagated to the input layer; in the process of backpropagation, various parameters are adjusted according to the error The value of ; continue to iterate the above process until convergence.

With the rapid development of artificial intelligence technology, various application products based on dialogue systems are gradually increasing, and the demand for voice interaction is also increasing. A dialogue system is a human-computer interaction system based on natural language. Through the dialogue system, users can use natural language to interact with the computer in multiple rounds to complete specific tasks. At present, dialogue systems are widely used in different fields, such as search, intelligent question answering, sentiment analysis, etc. Among them, natural language understanding is the core module of dialogue systems. The goal of natural language understanding is to convert the text information of natural language into a semantic representation that can be processed by a computer, that is, to use a structured data to represent the meaning expressed in a sentence. That is to say, the goal of natural language understanding is to determine the intention that the user wants to express and the conditions to satisfy the user's intention according to the text information to be parsed.

In natural language understanding, user intent needs to be classified. At present, intent classification is usually based on templates or models, but the inventor realized that template-based intent classification is heavily dependent on the coverage of templates, and is easily affected by data scale and data quality; model-based intent classification is vulnerable to the occurrence frequency of intent And the impact of the amount of data, it is often unable to solve the intent classification in the real scene well, affecting the accuracy of the intent classification results. Therefore, how to provide and realize accurate classification of user intent and improve the accuracy of intent classification has become a technical problem to be solved urgently.

Based on this, the embodiments of the present application provide an intention classification method, device, electronic device, and storage medium, which can realize accurate classification of user intentions and improve the accuracy of intention classification results.

The intent classification method, device, electronic device, and computer-readable storage medium provided in the embodiments of the present application are specifically described through the following embodiments. First, the intent classification method in the embodiments of the present application is described.

The embodiments of the present application may acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. .

Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

The intent classification method provided in the embodiment of the present application relates to the technical field of deep learning in artificial intelligence. The intent classification method provided in the embodiment of the present application can be applied to a terminal, can also be applied to a server, and can also be software running on the terminal or the server. In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.; the server end can be configured as an independent physical server, or can be configured as a server cluster or a distributed system composed of multiple physical servers, or It can be configured as a cloud that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server; the software may be an application that implements the intent classification method, but is not limited to the above forms.

The application can be used in numerous general purpose or special purpose computer system environments or configurations. Examples: personal computers, server computers, handheld or portable devices, tablet-type devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, including A distributed computing environment for any of the above systems or devices, etc. This application may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The application may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including storage devices.

Fig. 1 is an optional flow chart of the intention classification method provided by the embodiment of the present application, the method in Fig. 1 may include but not limited to include steps S101 to S106.

Step S101, obtaining request text;

Step S102, extracting entity features from the request text to obtain the first text containing the target query parameters;

Step S103, inputting the first text to the pre-trained comparison model and performing matrix multiplication with the reference word embedding matrix in the comparison model to obtain a plurality of target word embedding vectors;

Step S104, using the pre-trained intent classification model to classify the target word embedding vector, and obtain the target word embedding vector and the intent classification probability value including the intent category label;

Step S105, using the pre-trained intent matching model to perform matching processing on the first text to obtain an intent matching value;

Step S106, according to the intention matching value and the intention classification probability value, the intention classification data is obtained.

In step S101 of some embodiments, the request text can be obtained by writing a web crawler, setting a data source, and then performing targeted crawling of data. It should be noted that the request text is a natural language text.

Referring to FIG. 2, in some embodiments, the first text includes character text and semantic text, and step S102 may include, but is not limited to, step S201 to step S202:

Step S201, extracting the entity features of the request text according to the feature extraction model based on the prefix tree to obtain the character text;

Step S202, using the pre-trained lexical analysis model to perform recognition processing on the request text to obtain semantic text.

In step S201 of some embodiments, a feature extraction model based on a prefix tree may be constructed according to various types of knowledge databases. For example, the prefix tree-based feature extraction model contains multiple prefix trees constructed according to the music knowledge database, and these prefix trees are constructed from pre-stored data such as song names, singer names, and album names contained in the music knowledge database. The root node of each tree in the feature extraction model represents the first character of each pre-stored data. By comparing the character data of the entity feature in the request text with the first character of each pre-stored data, the entity feature in the current request text can be determined conveniently, and then information extraction is performed on the entity feature to obtain the character text.

In step S202 of some embodiments, the request data thesaurus needs to be constructed in advance, and the request data thesaurus may include various proper nouns, terms, non-proper names and the like related to query management. Through this request data lexicon, the preset lexical analysis model can enumerate specific query management names, for example, user complaints, query categories, and so on. Input the request text into the preset lexical analysis model, and identify the entity features in the request text through the specific query management contained in the preset lexical analysis model and the preset part-of-speech category. The entity features may include the above-mentioned and Query and manage entity vocabulary in multiple dimensions such as proper nouns, terms, non-proper names, modifiers, and time information.

In order to extract semantic text more accurately, the entity features in the request text can also be marked by a pre-trained sequence classifier, so that these entity features can carry preset labels to improve classification efficiency.

It should be noted that, in some specific embodiments, the pre-trained sequence classifier can be a maximum entropy Markov model (MEMM model) or a model based on conditional random field algorithm (CRF) or based on a two-way long short-term memory algorithm ( bi-LSTM) model. For example, a sequence classifier can be constructed based on the bi-LSTM algorithm. In the model based on the bi-LSTM algorithm, the input word wi and character embedding are passed through left-to-right long-short-term memory and right-to-left long-short-term memory, so that the output is The connected locations generate a single output layer. The sequence classifier can pass the input entity features directly to the softmax classifier through this output layer, and create a probability distribution on the preset part-of-speech category label through the softmax classifier, so as to mark and classify the entity feature data according to the probability distribution . Finally, feature extraction is performed on the entity feature data containing category labels to obtain the required semantic text.

In addition, in order to achieve data storage, the BERT encoder can also be used to convert the semantic text from the text form to the encoded form through the preset encoding function to realize the storage of the semantic text. The method can realize the semantic recognition processing and feature extraction of the request text, reduce the total amount of data, and make it more convenient to extract the required semantic text.

Referring to FIG. 3, in some embodiments, step S103 may include but not limited to include steps S301 to S303:

Step S301, performing word segmentation and encoding processing on the first text to obtain a plurality of query word segment vectors;

Step S302, inputting a plurality of query word vectors into the pre-trained comparison model, so that the query word vector and the reference word embedding matrix in the comparison model are matrix multiplied to obtain a plurality of basic word embedding vectors;

Step S303, performing mapping processing on the basic word embedding vector to obtain the target word embedding vector.

In some embodiments, step S301 may include, but is not limited to, the following steps:

A pre-trained text segmentation model is used to perform word segmentation processing on the first text to obtain multiple text word segments.

In step S301 of some embodiments, the pre-trained Jieba word breaker can be used to perform word segmentation processing on the original text to obtain text segments; specifically, when using the Jieba word breaker for word segmentation, first by comparing the dictionary in the Jieba word breaker Generate the directed acyclic graph corresponding to the original text, and then find the shortest path on the directed acyclic graph according to the preset selection mode and dictionary, and intercept the original text according to the shortest path, or directly intercept the original text , to get the text segment. Further, for text word segments that are not in the dictionary, HMM (Hidden Markov Model) can be used for new word discovery. Specifically, the positions B, M, E, and S of the characters in the text segment are taken as the hidden state, and the characters are the observed state, where B/M/E/S represent the words that appear at the beginning, middle, end, and composition of words, respectively. word. The dictionary file is used to store the performance probability matrix, initial probability vector and transition probability matrix between characters respectively. Then use the Viterbi algorithm to solve the maximum possible hidden state, so as to obtain the text field.

In step S301 of some other embodiments, it is also necessary to perform part-of-speech tagging on the text word segment, that is, perform part-of-speech tagging on the text word segment according to the preset part-of-speech category to obtain a text word segment containing the part-of-speech category tag, wherein the preset The part-of-speech categories include names, verbs, modifiers, adjectives, and more.

Through the above steps, word segmentation processing of the first text can be realized, making it more convenient to extract required text word segments.

Further, in step S301 of other embodiments, it may include but not limited to include the following steps:

Use the index function in the target thesaurus model to extract the elements of each text phrase to obtain the element value of each text phrase;

The position recognition of the text word segment is carried out according to the element value, and the target position of the text word segment is obtained.

Because the index function can return the element value in the table or array. Therefore, the element value of each text word segment is extracted through the index function in the form of an array to obtain the element value of the text word segment. Wherein, the element value of the text word segment includes the index value of the line number and the column number of the text word segment. Therefore, by searching the line number and column number of the text word segment through the index function, the text word segment at the specified position can be obtained. Use the index function to search the line number and column number of the text word segment, traverse each text field in the original text, and generate a position sequence table of the text word segment, which can reflect the text field, line number, and column number (element value) correspondence. That is to say, the target position of the text word segment is determined according to the element value, so that the location of each text word segment can be more accurately identified.

According to the target position, each text word segment is normalized to obtain a standard word segment;

Perform one-hot encoding on the standard word segment to obtain the text word segment vector.

Specifically, the target position is the index position of the text word segment. According to the index position of each text segment, each text segment is extracted from the first text, and each text segment is linearly scaled to [-1,1], or each text segment is scaled The average value is 0, and the variance is 1, so as to realize the normalization processing for each text segment and obtain the standard term segment.

It should be noted that one-hot encoding is One-Hot encoding, also known as one-bit effective encoding. The method is to use an N-bit state register to encode N states, each state has its own independent register bit, and at any time, only one bit is valid.

The length of the standard word segment can be expressed as a vector form by one-hot encoding, and multiple text word segment vectors can be obtained. For example, assuming that a certain original text consists of 3 text segments, the index positions of these 3 text segments can be obtained through the aforementioned steps. One-hot encoding is to use a vector representation of length V for each text segment, and this V is the number of dictionary words corresponding to the text segment in the target thesaurus model. The vector marks the index position of the text segment in the original text as 1, and the others are 0. Suppose a sentence is composed of 3 text segments, then there are 3 1s in this vector, and the position of this 1 can be compared with Corresponds to the index position of the text segment.

Through the above steps, each text word segment can be more conveniently encoded according to the target position to obtain a query word segment vector, so as to obtain a target word embedding vector through the query word segment vector.

Furthermore, step S302 is executed, and the value of the reference word embedding matrix in the comparison model can be completely fixed by training the comparison model, and other model parameters of the comparison model are also fixed. Therefore, when the query word vector is input into the comparison model, the fixed reference word embedding matrix can be used to perform matrix multiplication with each query word vector to obtain the basic word embedding vector.

Finally, step S303 is executed, using the fixed MLP network in the comparison model to perform mapping processing on the basic word embedding vector to obtain the target word embedding vector. Among them, the MLP network includes a linear layer, a ReLu activation function, and a linear layer.

Referring to FIG. 4, in some embodiments, before step S103, the method further includes training a comparison model, which may specifically include but not limited to steps S401 to S405:

Step S401, acquiring sample data;

Step S402, performing data enhancement processing on the sample data to obtain positive example pairs;

Step S403, input positive example pairs into the comparative learning model;

Step S404, calculate the first similarity of the positive example pair and the second similarity of the negative example pair by comparing the loss function of the learning model;

Step S405, optimizing the loss function of the contrastive learning model according to the first similarity and the second similarity, so as to update the contrastive learning model.

Specifically, firstly, the sample data is mapped to the embedding space, and the sample data is expressed as a vector, so that initial embedded data (that is, initial embedding data) can be obtained, and the initial embedded data includes positive sample data and negative sample data.

In step S402 of some embodiments, data enhancement processing is performed on the initial embedded data through the dropout mask mechanism; the embodiment of the present application replaces the traditional data enhancement method through the dropout mask mechanism, that is, the same sample data is input into the dropout encoder twice to obtain The two vectors of the two vectors are used as positive example pairs for comparative learning, and the effect is good enough, because for example, a different dropout mask is randomly generated for each dropout inside BERT, so only the same sample data (that is, the initial Embedding data) is input to the simCSE model twice, and the two vectors obtained are the results of applying two different dropout masks. It is understandable that the dropout mask is a random network model, which is the mask of the model parameter W, which prevents overfitting.

In a batch, the data (that is, the first vector and the second vector) obtained through data enhancement processing are positive example pairs, and other data that have not undergone data enhancement are negative example pairs. In the embodiment of the present application, some of the initial embedded data in a batch can be processed through data enhancement to obtain positive example pairs, and the other part of the initial embedded data can be used as negative example pairs.

In some embodiments, positive pairs are generated by randomly sampling the dropout mask.

In some specific application scenarios, in the stage of comparative learning, the typical comparative learning method in the batch is used to perform data enhancement processing within the batch, that is, to perform data enhancement processing on the complete initial embedding data obtained above, so that the positive examples There are differences between the two samples. In the embodiment of the present application, dropout is directly regarded as data enhancement, that is, the positive example pair is generated by randomly sampling the dropout mask, that is, the same first sample data and second sample data are respectively input to the dropout encoder for data enhancement processing, so that it can be obtained Two different representation vectors x (first vector) and x′ (second vector), the first vector and the second vector are taken as a positive example pair <x, x′>.

In step S404 of some embodiments, the first similarity and the second similarity are both cosine similarity, and the loss function of the comparative learning model is optimized according to the first similarity and the second similarity, which may include but not limited to include :

Maximize the first similarity to the first value and minimize the second similarity to the first value to optimize the loss function; where the first similarity is the numerator of the loss function, the first similarity and the second The similarity is the denominator of the loss function, the first value is 1, and the second value is 0. In this loss function, the numerator is the first similarity of the corresponding positive example pair, the denominator is the first similarity and the second similarity of all negative example pairs, and then the molecular formula value composed of the numerator and denominator is wrapped in -log() , so that the loss function can be minimized by maximizing the numerator and minimizing the denominator. In the embodiment of this application, minimizing the loss function infoNCE loss is to maximize the numerator and minimize the denominator, that is, to maximize the first similarity of the positive pair and minimize the second similarity of the negative pair, and the loss The function is minimized to realize the optimization of the loss function. More specifically, the loss function is shown in formula (1):

Among them, f(x) ^T is the transpose of f(x), f(x) is the original sample, f(x ⁺ ) is a positive sample, f(x _j ) is a single negative sample, and then all negative samples Added up, the denominator includes a positive sample and N-1 negative samples;

The loss function represents the loss of sample N; in this loss function, the numerator is the similarity of the positive pair, the denominator is the similarity of the positive pair and all negative pairs, and then wrap the value in -log() In this way, the loss function can be minimized by maximizing the numerator and minimizing the denominator.

It should be noted that the similarity (first similarity) of the positive example pair and the similarity (second similarity) of the negative example pair meet the conditions:

Score(f(x), f(x ⁺ ))>>Score(f(x), f(x ^- )) formula (2)

It can be seen from the above formula that this method needs to satisfy: the similarity of the positive example pair is greater than or equal to the similarity of the negative example pair, where x+ refers to the data similar to x, that is, the positive sample pair data; here x- refers to the data with x Dissimilar data, that is, negative sample pair data, f(x ⁺ ) is a positive sample, and f(x ^- ) is a negative sample.

Further, the preset measurement function is:

Score(f(x), f(x ⁺ )) = f(x) ^T f(x ⁺ ) formula (3)

Score(f(x), f(x ^- )) = f(x) ^T f(x ^- ) formula (4)

Among them, Score is a measurement function used to evaluate the similarity between two features. The default metric function is one that uses the dot product as the fraction function.

In step S405 of some embodiments, optimizing the loss function of the comparative learning model according to the first similarity and the second similarity may include but not limited to:

Backpropagation is performed according to the loss function, and the loss parameters of the loss function are updated to optimize the loss function.

In the embodiment of the present application, backpropagation is performed according to the loss function, so as to update the contrastive learning model by optimizing the loss function, and update the internal parameters (ie, loss parameters) of the contrastive learning model. It can be understood that conventional backpropagation principles may be applied to the backpropagation principle, which is not limited in this embodiment of the present application.

Referring to FIG. 5, in some embodiments, step S104 may also include but not limited to include steps S501 to S502:

Step S501, using the pre-trained intention classification model and the preset intention category to classify the word embedding vector, and obtain the word embedding vector containing the intention category label and the intention probability value corresponding to each intention category;

Step S502, according to the intention probability value, the intention classification probability value is obtained.

Specifically, in step S501, the intention classification model includes a softmax multi-class classifier, wherein the softmax multi-class classifier includes an input layer, a first feature layer and a second feature layer. The word embedding vector is input into the intent classification model, and the word embedding vector is encoded and pooled sequentially through the input layer, the first feature layer and the second feature layer to obtain the feature vector. The softmax multi-category classifier can be used in the pre- Create a probability distribution on the set intention category label, so as to mark and classify the feature vector according to the probability distribution, and obtain the word embedding vector containing the intention category label and the corresponding intention probability value of each intention category.

Furthermore, step S502 is executed, according to the intention probability value, the multiple intention probability values are arranged in descending order, the highest intention probability value is selected as the intention classification probability value, and the intention category corresponding to the intention probability value is used as the reference intention category.

In the above steps, the deep learning of intent classification probability can be performed by comparing the model and the intent classification model, thereby improving the accuracy of intent classification.

Referring to FIG. 6, step S105 in some embodiments may include but not limited to steps S601 to S602:

Step S601, inputting the first text into the preset intent matching model, so that the first text is matched with the preset sentence template to generate matching data;

Step S602, performing score statistics on the matching data according to a preset reference matching score to obtain an intended matching value.

Specifically, step S601 is executed, the preset intent matching model includes a plurality of preset sentence templates, the first text is input into the preset intent matching model, and the first text (specifically, the character containing the target query parameter text) and the sentence pattern template for character matching, if a certain sentence pattern template includes the character text, then it is considered that the sentence pattern template matches the character text. At the same time, by comparing the text content of the character text and the sentence pattern template, the matching data of each sentence pattern template can also be obtained. The matching data includes whether the target query parameters match, whether the sentence pattern characters match, whether there is a character intersection, the character text and the sentence pattern Whether the template is exactly the same and so on.

Furthermore, step S602 is executed to calculate the score of different matching data according to the preset reference matching score, and the intention matching value corresponding to each sentence template can be obtained. For example, the preset reference matching score includes: target query parameter matching plus 2 points, 1 point for sentence character matching, 0.5 points for character intersection, 100 points for whether the character text is completely consistent with the sentence template, etc. According to the preset reference matching score, the matching data of each sentence template is traversed, the score calculation of each sentence template is realized, and the intention matching value of each sentence template is obtained. By comparing the intent matching values of each sentence template, the one with the highest intent matching value is selected as the final sentence template and intent matching value.

Referring to FIG. 7, in some embodiments, step S106 may include but not limited to include steps S701 to S702:

Step S701, performing weighted calculation on the intention matching value and the intention classification probability value according to the preset weight ratio to obtain the comprehensive intention value;

In step S702, according to the integrated intention value, the intention classification data is obtained.

Specifically, the preset weight ratio may be the intent matching value: the intent classification probability value is 3:2, then the intent matching value and the intent classification probability value are weighted and calculated according to this ratio to obtain the comprehensive intent value. According to the size of the comprehensive intention value, query the comparison table of the comprehensive intention value and the intention category, so as to determine the corresponding intention category. According to the intent category, the intent classification data is obtained, and the intent classification data is the intent data under the intent category.

The embodiment of the present application obtains the request text, extracts the entity features of the request text, and obtains the first text containing the target query parameters. This method can realize the feature extraction of the request text and reduce the data space of the request text, making it more convenient Extract to the desired first text containing the target query parameters. Then input the first text to the pre-trained comparison model and perform matrix multiplication with the reference word embedding matrix in the comparison model to obtain multiple target word embedding vectors, and use the pre-trained intent classification model to classify the target word embedding vectors, The target word embedding vector and intent classification probability value containing the intent category label are obtained, and the problem of uneven distribution of the target word embedding vector can be better solved by comparing the model. At the same time, the depth of the intent classification probability can be determined by comparing the model and the intent classification model. Learning to improve the accuracy of intent classification probability values. In addition, the present application can also use the pre-trained intention matching model to perform matching processing on the first text to obtain the intention matching value, and the intention matching model can be used to calculate the user's intention matching value based on rule matching to improve the accuracy of intention matching. Finally, according to the intention matching value and the intention classification probability value, the intention classification data is obtained. In this application, through the comparison model, intent classification model, and intent matching model, the user's dialogue intent can be identified by integrating the two aspects of intent classification probability and intent matching, so that the final intent classification data can present more accurate intent classification results. The accurate classification of user intent is achieved, and the accuracy of intent classification is improved.

Please refer to FIG. 8, the embodiment of the present application also provides an intention classification device, which can realize the above intention classification method, and the device includes:

A text acquisition module 801, configured to acquire the request text;

A feature extraction module 802, configured to extract entity features from the request text to obtain the first text containing the target query parameters;

Comparison module 803, for inputting the first text to the pre-trained comparison model and the reference word embedding matrix in the comparison model to perform matrix multiplication to obtain a plurality of target word embedding vectors;

The classification module 804 is used to classify the target word embedding vector by using the pre-trained intent classification model to obtain the target word embedding vector and the intent classification probability value including the intent category label;

A matching module 805, configured to match the first text with a pre-trained intent matching model to obtain an intent matching value;

The calculation module 806 is configured to obtain intention classification data according to the intention matching value and the intention classification probability value.

The specific implementation manner of the intention classification device is basically the same as the specific embodiment of the above intention classification method, and will not be repeated here.

The embodiment of the present application also provides an electronic device, the electronic device includes: a memory, a processor, a program stored in the memory and operable on the processor, and a data bus for realizing connection and communication between the processor and the memory , when the program is executed by the processor, the above intention classification method is realized. The electronic device may be any intelligent terminal including a tablet computer, a vehicle-mounted computer, and the like.

Please refer to FIG. 9. FIG. 9 illustrates a hardware structure of an electronic device in another embodiment. The electronic device includes:

The processor 901 may be implemented by a general-purpose CPU (Central Processing Unit, central processing unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute related programs, so as to realize The intention classification method provided by the embodiment of this application;

The memory 902 may be implemented in the form of a read-only memory (ReadOnlyMemory, ROM), a static storage device, a dynamic storage device, or a random access memory (RandomAccessMemory, RAM). The memory 902 can store operating systems and other application programs. When implementing the technical solutions provided by the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 902 and called by the processor 901 to execute the implementation of this application. Example intent classification method;

The input/output interface 903 is used to realize information input and output;

The communication interface 904 is used to realize the communication and interaction between the device and other devices, and the communication can be realized through a wired method (such as USB, network cable, etc.), or can be realized through a wireless method (such as a mobile network, WIFI, Bluetooth, etc.);

bus 905, for transferring information between various components of the device (such as processor 901, memory 902, input/output interface 903 and communication interface 904);

The processor 901 , the memory 902 , the input/output interface 903 and the communication interface 904 are connected to each other within the device through the bus 905 .

Among them, the intention classification method provided by the embodiment of the present application includes:

get request text;

Input the first text to the pre-trained comparison model and perform matrix multiplication with the reference word embedding matrix in the comparison model to obtain a plurality of target word embedding vectors;

Use the pre-trained intent classification model to classify the target word embedding vector, and obtain the target word embedding vector and intent classification probability value containing the intent category label;

According to the intent matching value and the intent classification probability value, the intent classification data is obtained. An embodiment of the present application also provides a computer-readable storage medium for computer-readable storage. The computer-readable storage medium may be non-volatile or volatile. The computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement an intention classification method. Wherein, the intention classification method includes the following steps: obtaining the request text; performing entity feature extraction on the request text to obtain the first text containing the target query parameters; inputting the first text to the pre-trained comparison model and the reference word embedding in the comparison model Multiply the matrix to get multiple target word embedding vectors; use the pre-trained intent classification model to classify the target word embedding vectors, and obtain the target word embedding vector and intent classification probability value containing the intent category label; use the pre-trained intent classification model The intent matching model performs matching processing on the first text to obtain an intent matching value; and obtains intent classification data according to the intent matching value and the intent classification probability value.

As a non-transitory computer-readable storage medium, memory can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory optionally includes memory located remotely from the processor, and these remote memories may be connected to the processor via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The embodiments described in the embodiments of the present application are to illustrate the technical solutions of the embodiments of the present application more clearly, and do not constitute a limitation to the technical solutions provided by the embodiments of the present application. Those skilled in the art know that with the evolution of technology and new For the emergence of application scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

Those skilled in the art can understand that the technical solutions shown in Figures 1-7 do not constitute a limitation to the embodiments of the present application, and may include more or fewer steps than those shown in the illustrations, or combine certain steps, or be different A step of.

The device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Those of ordinary skill in the art can understand that all or some of the steps in the methods disclosed above, the functional modules/units in the system, and the device can be implemented as software, firmware, hardware, and an appropriate combination thereof.

If the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or part of the contribution to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including multiple instructions to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method in each embodiment of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, referred to as ROM), random access memory (Random Access Memory, referred to as RAM), magnetic disk or optical disc, etc., which can store programs. medium.

The preferred embodiments of the embodiments of the present application have been described above with reference to the accompanying drawings, and are not intended to limit the scope of rights of the embodiments of the present application. Any modifications, equivalent replacements and improvements made by those skilled in the art without departing from the scope and essence of the embodiments of the present application shall fall within the scope of rights of the embodiments of the present application.

Claims

A method for classifying intentions, wherein the method includes:

get request text;

performing entity feature extraction on the request text to obtain the first text containing target query parameters;

The first text is input to the pre-trained comparison model and the reference word embedding matrix in the comparison model is multiplied by matrix to obtain a plurality of target word embedding vectors;

Classify the target word embedding vector by using the pre-trained intent classification model to obtain the target word embedding vector and the intent classification probability value including the intent category label;

performing matching processing on the first text by using a pre-trained intent matching model to obtain an intent matching value;

According to the intention matching value and the intention classification probability value, the intention classification data is obtained.
The intent classification method according to claim 1, wherein the first text includes character text and semantic text, and the step of performing entity feature extraction on the request text to obtain the first text containing target query parameters includes :

performing entity feature extraction on the request text according to a prefix tree-based feature extraction model to obtain character text;

A pre-trained lexical analysis model is used to identify and process the request text to obtain semantic text.
The intent classification method according to claim 1, wherein the input of the first text to the pre-trained comparison model is matrix-multiplied with the reference word embedding matrix in the comparison model to obtain a plurality of target word embeddings Vector steps, including:

performing word segmentation and encoding processing on the first text to obtain a plurality of query word segment vectors;

A plurality of the query word vectors are input into the pre-trained comparison model, so that the query word vector and the reference word embedding matrix in the comparison model are multiplied by matrix to obtain a plurality of basic word embedding vectors;

The basic word embedding vector is mapped to obtain the target word embedding vector.
The method for classifying intent according to claim 1, wherein the input of the first text to the pre-trained comparison model is multiplied with the reference word embedding matrix in the comparison model to obtain a plurality of target words Before the step of embedding vectors, the method also includes training a comparison model, specifically including:

Get sample data;

performing data enhancement processing on the sample data to obtain positive example pairs;

inputting the positive example pair into the contrastive learning model;

Calculate the first similarity of the positive example pair and the second similarity of the negative example pair through the loss function of the comparative learning model;

Optimizing the loss function of the contrastive learning model according to the first similarity and the second similarity, so as to update the contrastive learning model.
The intention classification method according to claim 1, wherein, the step of using the pre-trained intention classification model to classify the target word embedding vector to obtain the target word embedding vector containing the intention category label and the intention classification probability value ,include:

Classify the word embedding vector using a pre-trained intent classification model and preset intent categories to obtain a word embedding vector containing an intent category label and an intent probability value corresponding to each intent category;

According to the intention probability value, an intention classification probability value is obtained.
The intent classification method according to claim 1, wherein the step of using a pre-trained intent matching model to perform matching processing on the first text to obtain an intent matching value includes:

Inputting the first text into a preset intent matching model, so that the first text is character-matched with a preset sentence template to generate matching data;

Score statistics are performed on the matching data according to a preset reference matching score to obtain an intention matching value.
The intention classification method according to any one of claims 1 to 6, wherein the step of obtaining intention classification data according to the intention matching value and the intention classification probability value includes:

performing weighted calculations on the intention matching value and the intention classification probability value according to a preset weight ratio to obtain a comprehensive intention value;

According to the comprehensive intention value, the intention classification data is obtained.
An intention classification device, wherein the device includes:

A text acquisition module, configured to acquire the request text;

A feature extraction module, configured to extract entity features from the request text to obtain the first text containing target query parameters;

Comparison module, for inputting the first text to the pre-trained comparison model and performing matrix multiplication with the reference word embedding matrix in the comparison model, to obtain a plurality of target word embedding vectors;

A classification module, configured to classify the target word embedding vector using a pre-trained intent classification model, to obtain a target word embedding vector and an intent classification probability value including an intent category label;

A matching module, configured to use a pre-trained intent matching model to perform matching processing on the first text to obtain an intent matching value;

A calculation module, configured to obtain intention classification data according to the intention matching value and the intention classification probability value.
An electronic device, wherein the electronic device includes a memory, a processor, a program stored on the memory and operable on the processor, and a program for realizing the connection between the processor and the memory A data bus for communication, when the program is executed by the processor, an intent classification method is implemented, wherein the intent classification method includes:

get request text;

performing entity feature extraction on the request text to obtain the first text containing target query parameters;

The first text is input to the pre-trained comparison model and the reference word embedding matrix in the comparison model is multiplied by matrix to obtain a plurality of target word embedding vectors;

Classify the target word embedding vector by using the pre-trained intent classification model to obtain the target word embedding vector and the intent classification probability value including the intent category label;

performing matching processing on the first text by using a pre-trained intent matching model to obtain an intent matching value;

According to the intention matching value and the intention classification probability value, the intention classification data is obtained.
The electronic device according to claim 9, wherein the first text includes character text and semantic text, and the step of performing entity feature extraction on the request text to obtain the first text containing target query parameters includes:

performing entity feature extraction on the request text according to a prefix tree-based feature extraction model to obtain character text;

A pre-trained lexical analysis model is used to identify and process the request text to obtain semantic text.
The electronic device according to claim 9, wherein the input of the first text to the pre-trained comparison model is matrix-multiplied with the reference word embedding matrix in the comparison model to obtain a plurality of target word embedding vectors steps, including:

performing word segmentation and encoding processing on the first text to obtain a plurality of query word segment vectors;

A plurality of the query word vectors are input into the pre-trained comparison model, so that the query word vector and the reference word embedding matrix in the comparison model are multiplied by matrix to obtain a plurality of basic word embedding vectors;

The basic word embedding vector is mapped to obtain the target word embedding vector.
The electronic device according to claim 9, wherein the step of using a pre-trained intention classification model to classify the target word embedding vector to obtain a target word embedding vector containing an intention category label and an intention classification probability value, include:

Classify the word embedding vector using a pre-trained intent classification model and preset intent categories to obtain a word embedding vector containing an intent category label and an intent probability value corresponding to each intent category;

According to the intention probability value, an intention classification probability value is obtained.
The electronic device according to claim 9, wherein the step of using a pre-trained intent matching model to perform matching processing on the first text to obtain an intent matching value includes:

Inputting the first text into a preset intent matching model, so that the first text is character-matched with a preset sentence template to generate matching data;

Score statistics are performed on the matching data according to a preset reference matching score to obtain an intention matching value.
The electronic device according to any one of claims 9 to 13, wherein the step of obtaining intent classification data according to the intent matching value and the intent classification probability value includes:

performing weighted calculations on the intention matching value and the intention classification probability value according to a preset weight ratio to obtain a comprehensive intention value;

According to the comprehensive intention value, the intention classification data is obtained.
A computer-readable storage medium for computer-readable storage, wherein the computer-readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to A method for classifying intent is realized, wherein the method for classifying intent includes the following steps:

get request text;

performing entity feature extraction on the request text to obtain the first text containing target query parameters;

The first text is input to the pre-trained comparison model and the reference word embedding matrix in the comparison model is multiplied by matrix to obtain a plurality of target word embedding vectors;

Classify the target word embedding vector by using the pre-trained intent classification model to obtain the target word embedding vector and the intent classification probability value including the intent category label;

performing matching processing on the first text by using a pre-trained intent matching model to obtain an intent matching value;

According to the intention matching value and the intention classification probability value, the intention classification data is obtained.
The computer-readable storage medium according to claim 15, wherein the first text includes character text and semantic text, and the step of performing entity feature extraction on the request text to obtain the first text containing target query parameters ,include:

performing entity feature extraction on the request text according to a prefix tree-based feature extraction model to obtain character text;

A pre-trained lexical analysis model is used to identify and process the request text to obtain semantic text.
The computer-readable storage medium according to claim 15, wherein the input of the first text to the pre-trained comparison model is matrix-multiplied with the reference word embedding matrix in the comparison model to obtain a plurality of target The steps of word embedding vector include:

performing word segmentation and encoding processing on the first text to obtain a plurality of query word segment vectors;

A plurality of the query word vectors are input into the pre-trained comparison model, so that the query word vector and the reference word embedding matrix in the comparison model are multiplied by matrix to obtain a plurality of basic word embedding vectors;

The basic word embedding vector is mapped to obtain the target word embedding vector.
The computer-readable storage medium according to claim 15, wherein the target word embedding vector is classified using the pre-trained intent classification model to obtain the target word embedding vector and the intent classification probability value including the intent category label steps, including:

Classify the word embedding vector using a pre-trained intent classification model and preset intent categories to obtain a word embedding vector containing an intent category label and an intent probability value corresponding to each intent category;

According to the intention probability value, an intention classification probability value is obtained.
The computer-readable storage medium according to claim 15, wherein the step of using a pre-trained intent matching model to perform matching processing on the first text to obtain an intent matching value includes:

Inputting the first text into a preset intent matching model, so that the first text is character-matched with a preset sentence template to generate matching data;

Score statistics are performed on the matching data according to a preset reference matching score to obtain an intention matching value.
The computer-readable storage medium according to any one of claims 15 to 19, wherein the step of obtaining intent classification data according to the intent matching value and the intent classification probability value includes:

performing weighted calculations on the intention matching value and the intention classification probability value according to a preset weight ratio to obtain a comprehensive intention value;

According to the comprehensive intention value, the intention classification data is obtained.