Claim dialogue intention recognition method and device, equipment and storage medium

Abstract

The invention discloses a claim dialogue intention recognition method and device, equipment and a storage medium, and belongs to the technical field of artificial intelligence. According to the method, the dialogue intention recognition model of the vector embedding layer, the multi-head self-attention layer, the feedforward network layer and the classification layer is introduced, so that the accuracy and robustness of intention recognition in the claim dialogue are effectively improved. Compared with the prior art, the method has the advantages that the text semantics are deeply represented by using the vector embedding layer, so that the understanding ability of the model to different expression modes is enhanced; a multi-head self-attention mechanism can capture key information in the context, and the perception ability of the model to the dialogue context is improved; the feed-forward network layer is used for realizing nonlinear combination of the features and is beneficial to mining implied domain features in the corpora; and the classification layer performs label matching and probability distribution calculation in combination with a preset scoring rule, so that efficient and accurate intention recognition is realized. According to the scheme, the user intention recognition effect is improved, and the requirement for high real-time performance in the actual claim settlement service is met.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a claim settlement dialogue intent recognition method and device, a computer device and a storage medium. BACKGROUND

[0002] In the modern insurance industry, claim settlement service is an important part of customer experience. With the development of information technology, more and more insurance companies have begun to use artificial intelligence (AI) and natural language processing (NLP) technology to improve claim settlement efficiency and improve customer service. Claim settlement dialogue usually involves multiple rounds of communication between customers and customer service, and the questions and needs raised by customers often have complexity and diversity. Effective identification of customer intent in the dialogue is crucial for quickly responding to customer needs and providing accurate services.

[0003] In recent years, pre-trained language models such as BERT have made significant progress in natural language understanding tasks. BERT models can better understand the semantic relationship and contextual dependency within a sentence through the learning of bidirectional context information. This makes BERT-based intent recognition methods gradually become a research hotspot.

[0004] Although BERT-based models have excellent text understanding capabilities, there are still challenges in claim settlement dialogue intent recognition. Customers express claim settlement intent in various ways and grammatical structures, making it difficult for models to accurately identify the same intent; claim settlement dialogue is conducted in multiple rounds, and customer intent often depends on the previous text, so it is difficult to obtain the true intent by analyzing a single round of dialogue alone, and the model needs to have the ability to process contextual information; the claim settlement process contains professional terms and complex clauses, and ordinary language models may lack domain knowledge, which can affect the accuracy of claim settlement dialogue intent recognition; in addition, in practical applications, customers expect quick feedback, but the current BERT model claim settlement dialogue intent recognition cannot meet the customers' immediate needs, affecting user experience. SUMMARY

[0005] The purpose of the embodiments of the present application is to propose a claim settlement dialogue intent recognition method, device, computer device and storage medium, aiming to solve the problems of diversified expression, contextual dependency, lack of domain knowledge and real-time demand in the prior art.

[0006] To solve the above technical problems, the embodiments of the present application provide a claim settlement dialogue intent recognition method, which adopts the following technical solutions:

[0007] A claim settlement dialogue intent recognition method, comprising:

[0008] acquiring pre-collected claim settlement question and answer corpus, and inputting the claim settlement question and answer corpus into a preset dialogue intent recognition model, wherein the dialogue intent recognition model comprises a vector embedding layer, a multi-head self-attention layer, a feedforward network layer and a classification layer;

[0009] The vector embedding layer is used to perform vector embedding representation on the claim question and answer corpus to obtain a text feature vector corresponding to the claim question and answer corpus;

[0010] The multi-head self-attention layer is used to perform feature weighting combination on the text feature vector to obtain a weighted combination vector;

[0011] The feedforward network layer is used to perform feature nonlinear transformation on the weighted combination vector to obtain a multi-dimensional feature vector;

[0012] The multi-dimensional feature vector is matched with an intent classification label in the classification layer, and scores of the respective intent classification labels are obtained based on a preset scoring rule;

[0013] The probability distribution of the classification label is calculated according to the scores of the respective intent classification labels, and the intent of the claim dialogue is determined according to the probability distribution.

[0014] To solve the above technical problems, the embodiment of the application further provides a claim dialogue intent recognition device, which adopts the technical scheme as follows:

[0015] A claim dialogue intent recognition device comprises:

[0016] A corpus acquisition module is configured to acquire a pre-collected claim question and answer corpus, and input the claim question and answer corpus into a preset dialogue intent recognition model, wherein the dialogue intent recognition model comprises a vector embedding layer, a multi-head self-attention layer, a feedforward network layer and a classification layer;

[0017] An embedding representation module is configured to perform vector embedding representation on the claim question and answer corpus by using the vector embedding layer to obtain a text feature vector corresponding to the claim question and answer corpus;

[0018] A feature weighting module is configured to perform feature weighting combination on the text feature vector by using the multi-head self-attention layer to obtain a weighted combination vector;

[0019] A feature transformation module is configured to perform feature nonlinear transformation on the weighted combination vector by using the feedforward network layer to obtain a multi-dimensional feature vector;

[0020] A label matching module is configured to match the multi-dimensional feature vector with an intent classification label in the classification layer, and obtain scores of the respective intent classification labels based on a preset scoring rule;

[0021] An intent recognition module is configured to calculate the probability distribution of the classification label according to the scores of the respective intent classification labels, and determine the intent of the claim dialogue according to the probability distribution.

[0022] To solve the above technical problems, the embodiment of the application further provides a computer device, which adopts the technical scheme as follows:

[0023] A computer device comprises a memory and a processor, the memory stores computer readable instructions, and the processor implements the steps of the claim settlement dialogue intent recognition method according to any one of the above when executing the computer readable instructions.

[0024] To solve the above technical problems, the embodiment of the present application also provides a computer readable storage medium, which adopts the technical scheme as follows:

[0025] A computer readable storage medium stores computer readable instructions, and the computer readable instructions implement the steps of the claim settlement dialogue intent recognition method according to any one of the above when executed by a processor.

[0026] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0027] The present application discloses a claim settlement dialogue intent recognition method, device, equipment and storage medium, belonging to the field of artificial intelligence. The present application introduces a dialogue intent recognition model of vector embedding layer, multi-head self-attention layer, feedforward network layer and classification layer, which effectively improves the accuracy and robustness of intent recognition in claim settlement dialogue. Compared with the prior art, the present application uses the vector embedding layer to deeply represent the text semantics, which enhances the understanding ability of the model to different expression methods; the multi-head self-attention mechanism can capture key information in the context, which improves the perception ability of the model to the dialogue context; the feedforward network layer realizes the nonlinear combination of features, which helps to mine the implicit domain features in the corpus; the classification layer combines the preset scoring rules to perform label matching and probability distribution calculation, thereby realizing efficient and accurate intent recognition. The present application not only improves the recognition effect of user intent in complex context, but also meets the demand for high real-time performance and strong domain adaptability in actual claim settlement business. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the scheme in the present application, the drawings needed in the description of the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0029] Figure 1 An exemplary system architecture diagram to which the present application can be applied is shown;

[0030] Figure 2 A flowchart of one embodiment of the claim settlement dialogue intent recognition method according to the present application is shown;

[0031] Figure 3Fig. 1 shows a structural schematic diagram of an embodiment of a claim settlement dialogue intention recognition apparatus according to the present application;

[0032] Figure 4 Fig. 1 shows a structural schematic diagram of an embodiment of a claim settlement dialogue intention recognition apparatus according to the present application; DETAILED DESCRIPTION

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application; the use herein of terms such as "comprise" and "have" and any variations thereof are intended to cover a non-exclusive inclusion; the use herein of terms such as "first", "second" and the like are used for distinguishing between similar objects and not necessarily for describing a specific sequential or chronological order.

[0034] Reference herein to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of the application. The appearances of the phrase that in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily mutually exclusive of one another. It is expressly understood that the embodiments described herein are merely examples and are not a limitation of the application.

[0035] In order to make the technical personnel in the art better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings.

[0036] As shown in Fig. 1, the system architecture 100 can include a terminal device 101, a network 102 and a server 103, and the terminal device 101 can be a notebook computer 1011, a tablet computer 1012 or a mobile phone 1013. The network 102 is a medium for providing a communication link between the terminal device 101 and the server 103. The network 102 can include various connection types, such as wired, wireless communication links or optical fiber cables, etc. Figure 1 A user can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0037]

[0038] ​The terminal device 101 can be various electronic devices with a display screen and supporting web browsing, in addition to the notebook computer 1011, the tablet computer 1012 or the mobile phone 1013, the terminal device 101 can also be an electronic book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV) player, a laptop computer and a desktop computer, etc.

[0039] The server 103 can be a server providing various services, for example, a background server supporting a page displayed on the terminal device 101.

[0040] It should be noted that the claim settlement dialogue intent recognition method provided by the embodiments of the present application is generally executed by a server / terminal device, and accordingly, the claim settlement dialogue intent recognition apparatus is generally arranged in a server / terminal device.

[0041] It should be understood that, Figure 1 The number of terminal devices, networks and servers in the above system is only illustrative, and the above system can have any number of terminal devices, networks and servers according to the implementation needs.

[0042] With reference to Figure 2 , a flow chart of one embodiment of the claim settlement dialogue intent recognition method according to the present application is shown. The claim settlement dialogue intent recognition method includes the following steps:

[0043] S201, a pre-collected claim settlement question and answer corpus is acquired, and the claim settlement question and answer corpus is input to a preset dialogue intent recognition model, wherein the dialogue intent recognition model includes a vector embedding layer, a multi-head self-attention layer, a feedforward network layer and a classification layer;

[0044] Specifically, the claim settlement question-answer corpus mainly includes various questions that users may ask in the claim settlement process and their corresponding standard answers, covering multiple intent categories such as reporting a case, claim settlement progress inquiry, claim settlement material submission, claim settlement status explanation, and compensation rule explanation. When collecting the corpus, artificial customer service logs, historical dialogue records, and preset question templates can be combined to ensure the diversity and coverage of the samples. The obtained corpus needs to be cleaned, labeled, and structured to ensure data quality and semantic accuracy. Before inputting the model, the corpus usually undergoes preprocessing operations such as word segmentation, stop word removal, and stemming. Subsequently, these processed corpora are input into the constructed intent recognition model, which includes a vector embedding layer for text semantic representation, a multi-head self-attention layer for capturing context relationships, a feedforward network layer for feature nonlinear mapping, and a classification layer for outputting intent labels. This model structure aims to deeply mine semantic information from the corpus and improve the understanding of diverse expressions and complex contexts.

[0045] For example, in practical applications, the claim settlement scenario usually involves a large number of user queries and insurance company responses. For example, users may ask questions such as "How do I apply for claim settlement if my car has been involved in a collision accident?" or "How much of my medical expenses can be reimbursed?" First, the system collects and organizes a large amount of historical claim settlement question-answer data, including customer questions and related insurance company responses, to form a claim settlement question-answer corpus. These question-answer pairs will be input into a preset dialogue intent recognition model. The model includes multiple layers, such as a vector embedding layer for converting text into numerical representation, a multi-head self-attention layer for capturing dependency relationships between different words, and a feedforward network layer and a classification layer to help the model identify the specific intent of the user based on the input corpus. Through this process, the system can accurately understand and respond to complex claim settlement scenarios.

[0046] The dialogue intent recognition model is modified based on the pre-training language model (BERT model) with a Transformer encoder structure, wherein:

[0047] The attention layer of the pre-training language model is replaced with a multi-head self-attention layer;

[0048] And a feedforward network layer is added after the multi-head self-attention layer to obtain the dialogue intent recognition model.

[0049] BERT (Bidirectional Encoder Representations from Transformers) is a pre-training language model based on the Transformer encoder structure. It is pre-trained on large-scale unsupervised text data to learn rich language representations, and then fine-tuned on various downstream natural language processing tasks, achieving significant performance improvement. The core architecture of BERT mainly consists of an input embedding layer, multiple Transformer encoder layers, and a task-related output layer.

[0050] Among them, the multi-head self-attention (MHA) mechanism is one of the core components of the Transformer model, which is an extension and improvement of the self-attention mechanism. The multi-head self-attention mechanism allows the model to simultaneously focus on important information in different positions in the sequence when processing each word. It maps the input sequence to multiple different query (Query), key (Key) and value (Value) spaces respectively, then calculates the attention weight in each space independently, and finally splices the attention results of different spaces. Through the multi-head self-attention mechanism, the model can capture long-distance dependencies and complex semantic information in the sequence.

[0051] The feedforward neural network is a two-layer fully connected network that performs independent nonlinear transformation on the vector of each position. It uses activation functions such as ReLU to introduce nonlinearity and enhance the model's expression ability. The feedforward neural network further extracts and converts the output of the self-attention mechanism, enabling the model to learn more complex language patterns.

[0052] In the above embodiment, structural modification is made on the basis of the pre-training language model BERT. The original attention layer is replaced by a multi-head self-attention layer, and a feedforward network layer is added after it to build a special model suitable for dialog intent recognition tasks. Compared with the standard attention mechanism, the multi-head self-attention layer can capture semantic relationships of different levels in the text from multiple subspaces in parallel, enhancing the model's ability to understand complex dialog context. The feedforward network layer further performs nonlinear transformation on the attention output, improving the expression ability and discrimination performance of the features. This modification scheme makes the model maintain the strong semantic modeling ability of BERT while better meeting the needs of the intent recognition task, significantly improving the recognition accuracy and adaptability to diverse user expressions.

[0053] S202, vector embedding layer is used to represent the claim settlement question and answer corpus by vector embedding, and the text feature vector corresponding to the claim settlement question and answer corpus is obtained;

[0054] Specifically, the vector embedding layer is a basic component in the model, mainly used to convert the input text data into a low-dimensional, continuous dense vector representation. This step uses a pre-trained word vector model (such as Word2Vec, GloVe or BERT) or through an end-to-end training method to map each word or subword in the claim settlement question and answer corpus into a semantic space, so that semantically similar words have similar vector representations. Vector embedding not only captures the semantic relationship between words, but also provides basic context information, especially when dealing with professional terms such as "claim settlement", "reporting", "inspection", etc. The embedding layer outputs a set of text feature vectors representing the basic semantic information of the input sentence, providing rich input features for further modeling.

[0055] S203, using a multi-head self-attention layer to combine the text feature vectors, obtaining a weighted combination vector;

[0056] Specifically, the multi-head self-attention mechanism can capture the relationship between each word in the text from different subspaces by constructing multiple attention heads, thereby achieving global modeling of the input sentence. Each attention head calculates the attention weight between words, i.e. the degree of attention of each word to other words, and applies the attention weight to the corresponding word vector to obtain the representation of the current word in the global context. The outputs of multiple attention heads are spliced or averaged to form the final weighted combination vector. This mechanism enables the model to not only identify "what the user asked", but also understand "how the user expressed it", for example, understanding that "how long does the claim settlement take to arrive" and "how long can the claim settlement be received" are semantically consistent. In addition, the multi-head self-attention layer can also identify the dependency relationship in the context, improving the model's processing ability in the face of complex sentence patterns or implicit intent scenarios, and further enhancing the recognition accuracy and generalization ability.

[0057] The multi-head self-attention (MHA) mechanism can significantly enhance the model's expression ability and help the model learn more diverse and complex features. In the multi-head mechanism, the input sequence data is divided into multiple heads, each of which performs independent calculation to generate different outputs. These outputs are finally spliced together to form the final result of the model. The output of each head can be represented as:

[0058]

[0059] In the formula, Head i represents the output of the i-th head, Q, K and V represent the query sub-vector, key sub-vector and value sub-vector, are the query transformation matrix, key transformation matrix and value transformation matrix of the i-th head respectively, and Attention is an attention calculation function. In MHA, the commonly used Attention function expression is as follows:

[0060]

[0061] In the formula, d k is the dimension of the key vector, softmax similarity normalization calculates the weight of each key vector, then multiplies the weight with the value vector, and finally performs weighted summation to obtain the output of the multi-head attention layer.

[0062] In the above specific embodiments, after the system inputs the claim settlement QA corpus into the model, the system will perform vectorization processing on the input claim settlement QA corpus. First, the vector embedding layer converts each word into a high-dimensional vector with semantic features through word vector embedding. For example, when the input is "collision accident", the vector embedding layer can convert the word into a vector containing relevant semantics and preserve the relative relationship between words. Then, the multi-head self-attention layer processes the word vectors to calculate the similarity between each word vector, and obtains the weighted combination vector of each word by combining the context information of different words. This process can effectively capture the difference in context between "car accident claim settlement" and "medical claim settlement", even if they contain similar words, the system can identify the semantic difference and accurately distinguish the claim types.

[0063] S204, performing feature nonlinear transformation on the weighted combination vector through the feedforward network layer to obtain a multi-dimensional feature vector;

[0064] Specifically, the feedforward network layer mainly performs nonlinear mapping and conversion on the weighted combination vector obtained in the previous step to further mine the deep semantic relationship in the text. This layer is usually composed of multiple fully connected neural networks and cooperates with activation functions such as ReLU and GELU to introduce nonlinear factors, thereby improving the model's expression ability for complex semantic patterns. This layer can optimize the training process through layer normalization, residual connection and other means to prevent problems such as gradient vanishing and overfitting. Through this process, the original semantic representation will be converted into more abstract and multi-dimensional feature representation, so that the model can identify similar intentions under different semantic expressions and improve the classification accuracy of complex and diverse sentences. These multi-dimensional feature vectors provide more discriminative inputs for the classification layer and are the key bridge connecting semantic extraction and intention classification.

[0065] S205, matching the multi-dimensional feature vector with the intention classification labels in the classification layer, and obtaining the scores of each intention classification label based on a preset scoring rule;

[0066] Specifically, the classification layer inputs the multi-dimensional feature vector extracted by the model into one or more linear transformation functions, each corresponding to a specific intent category (such as "claim progress query", "material supplement", "explanation of claim standards", etc.), for calculating the matching scores of the input vector with each category. These scores can be obtained through dot product, similarity calculation or neural network output, measuring the semantic closeness of the input sentence to each intent category. At the same time, the model can combine preset scoring rules, such as weight adjustment mechanism, category confidence threshold or prior knowledge guidance, to optimize the output scores of different categories, so as to ensure that the final result has higher accuracy and stability. This step is to calculate the matching degree of each intent category by precise calculation, so that the system has the ability to distinguish subtle semantic differences.

[0067] It should be noted that the intent label identification of the present application is identified in the style of j son, and each intent label is represented in the form of key-value pair, where the key is the name of the intent label and the value is the corresponding score or other related information. Through this structured way, the system can clearly show the identification result of each intent, and sort and classify the user's intent according to different score values, so as to improve the response accuracy and efficiency of the dialogue system.

[0068] S206, according to the score of each intent classification label, the probability distribution of the classification label is calculated, and the intent of the claim dialogue is determined according to the probability distribution.

[0069] Specifically, this step inputs the intent score obtained in the previous stage into the Softmax function or other normalization tools, and converts it into a standard probability distribution form, so as to intuitively reflect the possibility of each intent label as the final identification result. In actual application, this probability distribution is not only used to select the label with the highest score as the main intent of the current claim dialogue, but also can be used for multi-intent recognition, confidence evaluation, result interpretation and other extended functions. For example, in the case of small probability gap, the system can prompt "multiple possible intents" to improve user experience; or combined with manual review mechanism for secondary confirmation to enhance the reliability of decision-making. Through this probability-based judgment method, the system can balance accuracy and controllability, especially when facing complex, ambiguous or cross-intent expressions, it can still maintain stable recognition performance and good business practicability.

[0070] After obtaining the weighted combination vector, the feedforward network layer performs a nonlinear transformation on it to further extract potential complex features. For example, for the query of "traffic accident compensation", the feedforward network layer can perform dimensionality reduction and dimensionality reduction on the input vector to extract high-dimensional features related to the compensation process, accident type, insurance policy, etc. Then, the system matches these feature vectors with pre-set intent classification labels (such as "traffic accident compensation", "medical reimbursement", etc.) and calculates their similarity to generate a score for each intent classification label. For example, if the system identifies the user's "traffic accident compensation" question, it will give the "traffic accident compensation" label a higher score, while the "medical compensation" label will have a lower score. Through this series of operations, the model can calculate the final probability distribution according to the scores of each intent classification label to determine the user's intent and quickly and accurately respond to compensation needs.

[0071] In the above embodiments, the present application introduces a dialogue intent recognition model with vector embedding layer, multi-head self-attention layer, feedforward network layer and classification layer, effectively improving the accuracy and robustness of intent recognition in the compensation dialogue. Compared with the prior art, the present application uses the vector embedding layer to deeply represent the text semantics, enhancing the model's understanding ability of different expression forms; the multi-head self-attention mechanism can capture key information in the context, improving the model's perception ability of the dialogue context; the feedforward network layer realizes the nonlinear combination of features, which helps to mine the implicit domain features in the corpus; the classification layer combines the pre-set scoring rules for label matching and probability distribution calculation, thereby realizing efficient and accurate intent recognition. The present application not only improves the recognition effect of user intent in complex context, but also meets the demand for high real-time performance and strong domain adaptability in actual compensation business.

[0072] Further, the step of obtaining the text feature vector corresponding to the compensation question and answer corpus by vector embedding the compensation question and answer corpus through the vector embedding layer comprises:

[0073] performing text segmentation processing on the compensation question and answer corpus to obtain initial text segmentation;

[0074] performing semantic recognition on the initial text segmentation to obtain compensation question and answer segmentation;

[0075] determining the position of the compensation question and answer segmentation in the compensation question and answer corpus to obtain segmentation position information;

[0076] loading a bag-of-words model pre-set in the vector embedding layer and using the bag-of-words model to perform word vector embedding representation on the compensation question and answer segmentation to obtain a text segmentation vector;

[0077] encoding the segmentation position information to obtain a segmentation position vector;

[0078] The text feature vector is obtained by combining the text segmentation vector and the segmentation position vector.

[0079] In this embodiment, in order to realize more accurate text semantic modeling, first, the standard text segmentation processing is performed on the claim settlement question and answer corpus, the basic lexical units in the sentence are extracted, and the initial text segmentation is formed. Then, the initial segmentation is further filtered, classified and standardized by the semantic recognition module, so that the extracted claim settlement question and answer segmentation has higher semantic consistency and business relevance, such as normalizing expressions such as “claim settlement amount” and “payout limit”. Subsequently, the corresponding segmentation position information is generated in combination with the position relationship of each segmentation in the corpus, so as to maintain the influence of the word order structure in the text on the semantics. In the embedding link, the bag-of-words model (such as pre-trained word vectors or customized word vectors) embedded in the vector embedding layer is loaded to map the segmentation into a high-dimensional dense vector, thereby obtaining the text segmentation vector representing the semantic content. At the same time, the corresponding position information is also converted into a vector form through a position encoding mechanism. Finally, the semantic vector and the position vector are fused to generate a text feature vector that not only retains the word meaning but also reflects the context structure, thereby providing high-quality input features for attention modeling and classification.

[0080] It should be noted that the vector embedding of the claim settlement question and answer corpus text input by the user includes word vector embedding and position embedding. The position information is obtained through the embedding layer and also participates in the fine-tuning training process, so that the model can not only learn the semantic relationship between words during training, but also dynamically adjust the perception ability of the word order structure. Unlike traditional static position encoding, the position embedding vector in this scheme is a trainable parameter that can be optimized and updated according to the language distribution and context logic in the actual corpus. Through joint training of the word vector and the position vector, the model is more accurate when processing dialog content involving time sequence, causal relationship or sequential dependence, especially for problems involving sequential semantics in the claim settlement scenario, such as “what information to submit first” and “when to receive the payment”. This mechanism enhances the model's understanding of text structure and semantic combination, and improves the context perception effect and expression diversity of intent recognition.

[0081] Through the above steps, the model can effectively retain the semantic and position information of the words, improve the understanding of the context dependence and complex language structure, and thus enhance the accuracy and robustness of intent recognition.

[0082] Further, the step of using a multi-head self-attention layer to perform feature weighting combination on the text feature vector to obtain a weighted combination vector includes:

[0083] obtaining query sub-vectors, key sub-vectors and value sub-vectors corresponding to the text feature vector based on the multi-head self-attention layer;

[0084] calculating multi-head attention weights of the text feature vector based on the query sub-vector and the key sub-vector;

[0085] performing dot product operation or linear transformation on the multi-head attention weights and the value sub-vector to obtain a weighted combination vector.

[0086] In this embodiment, the text feature vector is processed using a multi-head self-attention layer. First, the input text feature vector is mapped to multiple different subspaces, each corresponding to an attention head. In each head, query, key and value sub-vectors are extracted from the text feature vector, which are linearly transformed by different weight matrices. Then, the attention weights of each position are determined based on the similarity between the query sub-vector and the key sub-vector (usually calculated by dot product and scaled and Softmax normalized), so as to measure the degree of attention of the word to other words. Subsequently, the attention weights are applied to the corresponding value sub-vector, and the output vector of each attention head is obtained by weighted summation or linear transformation. Finally, the outputs of all attention heads are spliced or integrated to form a complete weighted combination vector. The multi-head mechanism can capture different dimensional semantic features in the text, such as context dependence, sentiment orientation or the relevance between domain terms, so that the model has stronger expression ability and flexibility in understanding complex claim sentences.

[0087] Through the above steps, the model can fully mine the deep relationships between words in the text, improve the modeling ability of context semantics and the dependence between keywords, and thus enhance the accuracy and generalization ability of intent recognition.

[0088] Further, the step of calculating multi-head attention weights of the text feature vector based on the query sub-vector and the key sub-vector specifically includes:

[0089] calculating the feature vector similarity of the query sub-vector and the key sub-vector to obtain a first feature similarity;

[0090] performing weighting operation according to the feature vector similarity to obtain the multi-head attention weights.

[0091] In this embodiment, in order to effectively obtain the dependency relationship between different words in the text, the similarity of the feature vector is calculated based on the query sub-vector and the key sub-vector to obtain the attention weight. Specifically, first, the dot product operation is performed between the position of each word in the query vector of the sentence and the key vector of the remaining words to obtain the correlation score between the word and other words, i.e., the first feature similarity. In order to enhance the stability and training effect of the model, the similarity is usually divided by a scaling factor (such as the square root of the key vector dimension), and then normalized by the Softmax function to convert it into a set of weighting coefficients, which are the multi-head attention weights. These weights reflect the degree of attention of the current word to other words when modeling the entire sentence. This process is not performed only once, but in parallel in multiple attention heads to ensure that information is extracted from different semantic levels. Finally, the attention weight of each head guides the weighted summation of the value sub-vector, which retains the potential long-distance dependency information between words and enhances the comprehensive understanding of the model to the sentence structure and semantics.

[0092] It should be noted that for a specific query vector, the multi-head self-attention layer will perform weighted summation on the value sub-vector, and the weight is calculated based on the similarity between the query sub-vector and the key sub-vector. Then, the obtained multi-head attention weight is multiplied by the value sub-vector to generate the final output. In the calculation of the similarity, the commonly used methods include dot product and bilinear calculation.

[0093] Through the above steps, the model can dynamically capture the semantic association between words, thereby improving the recognition ability of the implied intent in complex sentence and long text.

[0094] Further, the step of performing feature nonlinear transformation on the weighted combination vector through the feedforward network layer to obtain a multi-dimensional feature vector includes:

[0095] performing dimensionality increasing operation on the weighted combination vector through the feedforward network layer to obtain a high-dimensional feature vector;

[0096] performing dimensionality decreasing operation on the weighted combination vector through the feedforward network layer to obtain a low-dimensional feature vector;

[0097] mapping the high-dimensional feature vector and the low-dimensional feature vector into the same feature space;

[0098] performing multi-dimensional feature extraction in the feature space to obtain a multi-dimensional feature vector.

[0099] In this embodiment, the feedforward network layer performs a nonlinear transformation on the weighted combination vector output by the multi-head self-attention layer, mainly including dimension increasing, activation, and dimension reduction operations, to enhance the expression ability of the features. Specifically, first, the weighted combination vector is input into a fully connected layer for dimension increasing operation, so as to map it to a higher-dimensional feature space, and to mine deep feature relationships hidden under the original semantic expression; then, a nonlinear factor is introduced through an activation function (such as ReLU or GELU), so that the model can learn complex semantic changes and patterns; subsequently, another fully connected layer is used to reduce the dimension of the high-dimensional feature vector, so as to map it back to a space with the same or a set dimension as the original input, thereby retaining key information and compressing redundant content. After dimension reduction, the high-dimensional and low-dimensional features are fused and uniformly mapped to a shared feature space, on the basis of which multi-dimensional feature extraction is performed, thereby generating a multi-dimensional feature vector containing rich semantic representations. This process significantly improves the modeling ability of the model for nonlinear semantic changes in the text, and provides multi-dimensional features for the classification layer to accurately identify the intent.

[0100] Through the above steps, the model can effectively extract high-order semantic features, enhance the expression and discrimination ability of complex semantic structures, and improve the accuracy and robustness of claim settlement intent recognition.

[0101] Further, the step of matching the multi-dimensional feature vector with the intent classification labels in the classification layer and obtaining the scores of the intent classification labels based on a preset scoring rule includes:

[0102] mapping the multi-dimensional feature vector into the dimensions of the intent classification labels, respectively;

[0103] obtaining a label embedding vector corresponding to each intent classification label;

[0104] calculating the feature vector similarity between the multi-dimensional feature vector and each label embedding vector to obtain a second feature similarity;

[0105] weighting the intent classification labels based on the second feature similarity to obtain classification label weights;

[0106] obtaining initial scores corresponding to each intent classification label, and calculating the final score of each intent classification label in combination with the classification label weights.

[0107] In the embodiment, the multi-dimensional feature vector generated by the feedforward network layer is first mapped into the intent classification label space, so that it can be docked with the feature representation of each label. Specifically, the model assigns each intent classification label an independent label embedding vector according to the number of labels and their dimensions, which are obtained by training and can reflect the semantic features of different intents. Then, the similarity between the multi-dimensional feature vector and each label embedding vector is calculated, usually using cosine similarity or dot product calculation, to obtain the second feature similarity of each label. Based on these similarity values, the model assigns a weight to each intent classification label, which reflects the matching degree of the label with the input text. Then, according to the initial score of each label and the calculated weight, the score of each intent classification label is adjusted to obtain the final intent classification score. This process enables the model to accurately classify the claim conversation input by the user by the similarity between labels and feature weights, helping to identify the most matched claim intent.

[0108] Through the above steps, the model can efficiently dynamically adjust the classification label weight according to the semantic similarity, improve the accuracy and relevance of intent classification, and is particularly suitable for complex claim conversation scenarios.

[0109] In the above embodiment, the present application discloses a claim conversation intent recognition method, which belongs to the field of artificial intelligence. The present application introduces a dialogue intent recognition model with vector embedding layer, multi-head self-attention layer, feedforward network layer and classification layer, which effectively improves the accuracy and robustness of intent recognition in claim conversation. Compared with the prior art, the present application uses the vector embedding layer to deeply represent the text semantics, enhancing the model's understanding ability of different expression methods; the multi-head self-attention mechanism can capture key information in the context, improving the model's perception ability of the conversation context; the feedforward network layer realizes the nonlinear combination of features, which helps to mine the implicit domain features in the corpus; the classification layer combines the preset scoring rules for label matching and probability distribution calculation, thereby realizing efficient and accurate intent recognition. The present application scheme not only improves the recognition effect of user intent in complex context, but also meets the demand for high real-time performance and strong field adaptability in actual claim business.

[0110] In the embodiment, the electronic device (e.g. Figure 1 The server) on which the claim conversation intent recognition method runs can receive instructions or obtain data through wired or wireless connection. It should be noted that the above wireless connection can include but is not limited to 3G / 4G connection, Wi-Fi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wide band) connection, and other now known or future developed wireless connection.

[0111] It should be emphasized that, in order to further ensure the privacy and security of the above-mentioned claim settlement question and answer corpus information, the above-mentioned claim settlement question and answer corpus information can also be stored in a node of a block chain.

[0112] The blockchain referred to in the present application is a new application mode of distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm and other computer technologies. Blockchain, in essence, is a decentralized database, which is a series of data blocks associated using cryptographic methods, each data block containing information about a batch of network transactions, used to verify the validity of the information (anti-fake) and generate the next block. The blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer, etc.

[0113] The embodiments of the present application can acquire and process related data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the use of digital computers or digital computer-controlled machines to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. Theory, methods, technologies and application systems.

[0114] The basic technology of artificial intelligence generally includes technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, robot technology, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning, etc. Several major directions.

[0115] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by computer readable instructions instructing related hardware, and the computer readable instructions can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of each method. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM) and other non-volatile storage media, or a random access memory (RAM) and the like.

[0116] It should be understood that although each step in the flowchart of the accompanying drawings is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other sequences. Moreover, at least part of the steps in the flowchart of the accompanying drawings can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence is not necessarily sequential, but can be alternately executed with at least part of other steps or sub-steps or stages of other steps.

[0117] Further referring to Figure 3 , as an implementation of the method shown in the above Figure 2 , the present application provides an embodiment of a claim settlement dialogue intent recognition device, which corresponds to the method embodiment shown in Figure 2 , and the device can be specifically applied to various electronic devices.

[0118] As shown in Figure 3 , the claim settlement dialogue intent recognition device 300 described in the embodiment includes:

[0119] The corpus acquisition module 301 is configured to acquire pre-collected claim settlement question and answer corpus, and input the claim settlement question and answer corpus into a preset dialogue intent recognition model, wherein the dialogue intent recognition model includes a vector embedding layer, a multi-head self-attention layer, a feedforward network layer, and a classification layer.

[0120] The embedding representation module 302 is configured to perform vector embedding representation on the claim settlement question and answer corpus through the vector embedding layer to obtain a text feature vector corresponding to the claim settlement question and answer corpus.

[0121] The feature weighting module 303 is configured to perform feature weighting combination on the text feature vector using the multi-head self-attention layer to obtain a weighted combination vector.

[0122] The feature transformation module 304 is configured to perform feature nonlinear transformation on the weighted combination vector through the feedforward network layer to obtain a multi-dimensional feature vector.

[0123] The label matching module 305 is configured to match the multi-dimensional feature vector with an intent classification label in the classification layer, and obtain a score of each intent classification label based on a preset scoring rule.

[0124] The intent recognition module 306 is configured to calculate a probability distribution of the classification label according to the score of each intent classification label, and determine the intent of the claim settlement dialogue according to the probability distribution.

[0125] Further, the dialogue intent recognition model is improved based on a pre-training language model with a Transformer encoder structure, and the claim dialogue intent recognition device 300 further comprises:

[0126] an attention layer reconstruction module, configured to replace an attention layer of the pre-training language model with a multi-head self-attention layer;

[0127] a feedforward network setting module, configured to add a feedforward network layer after the multi-head self-attention layer, to obtain the dialogue intent recognition model.

[0128] Further, the embedding representation module 302 specifically comprises:

[0129] a word segmentation processing unit, configured to perform text word segmentation processing on the claim question and answer corpus, to obtain initial text word segmentation;

[0130] a semantic recognition unit, configured to perform semantic recognition on the initial text word segmentation, to obtain claim question and answer word segmentation;

[0131] a word segmentation position unit, configured to determine a position of the claim question and answer word segmentation in the claim question and answer corpus, to obtain word segmentation position information;

[0132] an embedding representation unit, configured to load a bag-of-words model preset in a vector embedding layer, and perform word vector embedding representation on the claim question and answer word segmentation using the bag-of-words model, to obtain a text word segmentation vector;

[0133] a position encoding unit, configured to encode the word segmentation position information, to obtain a word segmentation position vector;

[0134] a vector combination unit, configured to combine the text word segmentation vector and the word segmentation position vector, to obtain a text feature vector.

[0135] Further, the feature weighting module 303 specifically comprises:

[0136] a sub-vector reading unit, configured to obtain a query sub-vector, a key sub-vector and a value sub-vector corresponding to the text feature vector based on the multi-head self-attention layer;

[0137] a multi-head attention weight unit, configured to calculate multi-head attention weights of the text feature vector based on the query sub-vector and the key sub-vector;

[0138] a weighted combination unit, configured to perform dot product operation or linear transformation on the multi-head attention weights and the value sub-vector, to obtain a weighted combination vector.

[0139] Further, the multi-head attention weight unit specifically comprises:

[0140] a first feature similarity sub-unit, configured to calculate a feature vector similarity of the query sub-vector and the key sub-vector, to obtain a first feature similarity;

[0141] The first feature weighting subunit is configured to perform a weighting operation according to the feature vector similarity to obtain the multi-head attention weight.

[0142] Further, the feature transformation module 304 specifically includes:

[0143] The dimension increasing operation unit is configured to perform a dimension increasing operation on the weighted combination vector through the feedforward network layer to obtain a high-dimensional feature vector.

[0144] The dimension decreasing operation unit is configured to perform a dimension decreasing operation on the weighted combination vector through the feedforward network layer to obtain a low-dimensional feature vector.

[0145] The feature mapping unit is configured to map the high-dimensional feature vector and the low-dimensional feature vector into the same feature space.

[0146] The multi-dimensional feature extraction unit is configured to perform multi-dimensional feature extraction in the feature space to obtain a multi-dimensional feature vector.

[0147] Further, the label matching module 305 specifically includes:

[0148] The multi-dimensional mapping unit is configured to map the multi-dimensional feature vector into the dimensions of the intent classification labels respectively.

[0149] The label embedding unit is configured to obtain a label embedding vector corresponding to each intent classification label.

[0150] The second feature similarity subunit is configured to calculate feature vector similarities between the multi-dimensional feature vector and the label embedding vectors to obtain second feature similarities.

[0151] The second feature weighting subunit is configured to perform weighting on the intent classification labels based on the second feature similarities to obtain classification label weights.

[0152] The score weighting unit is configured to obtain initial scores corresponding to the intent classification labels and calculate final scores of each intent classification label in combination with the classification label weights.

[0153] In the above embodiment, the application discloses a claim settlement dialogue intention recognition device, which belongs to the technical field of artificial intelligence. The application introduces a dialogue intention recognition model of a vector embedding layer, a multi-head self-attention layer, a feedforward network layer and a classification layer, effectively improving the accuracy and robustness of intention recognition in claim settlement dialogue. Compared with the prior art, the application uses the vector embedding layer to deeply represent the text semantics, enhances the understanding ability of the model to different expression methods, the multi-head self-attention mechanism can capture key information in the context, improves the perception ability of the model to the dialogue context, the feedforward network layer realizes the nonlinear combination of features, which is helpful to mine the implicit domain features in the corpus, and the classification layer combines the preset scoring rules to realize label matching and probability distribution calculation, thereby realizing efficient and accurate intention recognition. The application scheme not only improves the recognition effect of user intention in complex context, but also meets the demand for high real-time performance and strong field adaptability in actual claim settlement business.

[0154] To solve the above technical problems, the application embodiment further provides a computer device. For details, please refer to Figure 4 , Figure 4 The basic structure block diagram of the computer device of the embodiment is shown in the figure.

[0155] The computer device 4 includes a memory 41, a processor 42 and a network interface 43 which are connected to each other through a system bus. It should be pointed out that only the computer device 4 with the memory 41, the processor 42 and the network interface 43 is shown in the figure, but it should be understood that it is not required to implement all the shown components, and more or fewer components can be alternatively implemented. Among them, those skilled in the art can understand that the computer device here is a device capable of automatically performing numerical calculation and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessor, application specific integrated circuit (ASIC), field programmable gate array (FPGA), digital signal processor (DSP), embedded device, etc.

[0156] The computer device can be a desktop computer, a notebook computer, a palm computer and a cloud server, etc. The computer device can interact with the user through a keyboard, a mouse, a remote controller, a touchpad or a voice control device, etc.

[0157] The memory 41 includes at least one type of readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or a memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the computer device 4. Of course, the memory 41 can also include both the internal storage unit and the external storage device of the computer device 4. In this embodiment, the memory 41 is generally used to store an operating system and various application software installed on the computer device 4, such as computer readable instructions of the claim settlement dialogue intent recognition method, etc. In addition, the memory 41 can also be used to temporarily store various data that have been output or will be output.

[0158] The processor 42 can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run computer readable instructions or process data stored in the memory 41, such as computer readable instructions of the claim settlement dialogue intent recognition method.

[0159] The network interface 43 can include a wireless network interface or a wired network interface, and is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0160] The present application also provides an embodiment, i.e., to provide a computer device, which includes a memory and a processor, the memory stores computer readable instructions, and the processor executes the computer readable instructions to implement the steps of the user insurance demand evaluation method as described above, i.e., to implement:

[0161] A claim settlement dialogue intent recognition method includes:

[0162] The pre-collected claim settlement question and answer corpus is obtained, and the claim settlement question and answer corpus is input into a preset dialogue intent recognition model, wherein the dialogue intent recognition model comprises a vector embedding layer, a multi-head self-attention layer, a feedforward network layer, and a classification layer.

[0163] The claim settlement question and answer corpus is represented by vector embedding through the vector embedding layer to obtain a text feature vector corresponding to the claim settlement question and answer corpus.

[0164] The text feature vector is combined by feature weighting using the multi-head self-attention layer to obtain a weighted combination vector.

[0165] The weighted combination vector is transformed by feature nonlinearity through the feedforward network layer to obtain a multi-dimensional feature vector.

[0166] The multi-dimensional feature vector is matched with an intent classification label in the classification layer, and scores of each intent classification label are obtained based on a preset scoring rule.

[0167] The probability distribution of the classification label is calculated according to the scores of each intent classification label, and the intent of the claim settlement dialogue is determined according to the probability distribution.

[0168] The application also provides another embodiment, that is, a computer readable storage medium storing computer readable instructions executable by at least one processor to cause the at least one processor to perform the steps of the claim settlement dialogue intent recognition method as described above, that is, to implement:

[0169] A claim settlement dialogue intent recognition method comprises:

[0170] The pre-collected claim settlement question and answer corpus is obtained, and the claim settlement question and answer corpus is input into a preset dialogue intent recognition model, wherein the dialogue intent recognition model comprises a vector embedding layer, a multi-head self-attention layer, a feedforward network layer, and a classification layer.

[0171] The claim settlement question and answer corpus is represented by vector embedding through the vector embedding layer to obtain a text feature vector corresponding to the claim settlement question and answer corpus.

[0172] The text feature vector is combined by feature weighting using the multi-head self-attention layer to obtain a weighted combination vector.

[0173] The weighted combination vector is transformed by feature nonlinearity through the feedforward network layer to obtain a multi-dimensional feature vector.

[0174] The multi-dimensional feature vector is matched with an intent classification label in the classification layer, and scores of each intent classification label are obtained based on a preset scoring rule.

[0175] The probability distribution of the classification label is calculated according to the scores of the respective intent classification labels, and the intent of the claim conversation is determined according to the probability distribution.

[0176] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for causing a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present application.

[0177] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment, in which tasks are performed by remote processing devices connected by a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0178] It should be noted that the non-company software tools or components appearing in various embodiments of the present application are only illustrative and do not represent actual use.

[0179] Obviously, the above-described embodiments are only a part of the embodiments of the present application, and not all the embodiments. The preferred embodiments of the present application are given in the drawings, but do not limit the patent scope of the present application. The present application can be realized in many different forms, and on the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent replacements to some technical features. Any equivalent structure made by using the contents of the specification and drawings, directly or indirectly applied to other related technical fields, is also within the scope of the patent protection of the present application.

Claims

1. A method for recognizing the intention of a claims dialogue, characterized in that: include: Obtaining pre-collected claims question and answer corpus and inputting the claims question and answer corpus into a preset conversation intent recognition model, wherein the conversation intent recognition model includes a vector embedding layer, a multi-head self-attention layer, a feedforward network layer, and a classification layer; Performing vector embedding representation on the claim question and answer corpus through the vector embedding layer to obtain a text feature vector corresponding to the claim question and answer corpus; Using the multi-head self-attention layer to perform feature weighted combination on the text feature vector to obtain a weighted combination vector; Performing a feature nonlinear transformation on the weighted combination vector through the feedforward network layer to obtain a multi-dimensional feature vector; Matching the multidimensional feature vector with the intent classification labels in the classification layer, and obtaining a score for each intent classification label based on a preset scoring rule; The probability distribution of the classification labels is calculated based on the scores of the respective intention classification labels, and the intent of the claims dialogue is determined based on the probability distribution.

2. The method for recognizing claims dialogue intention according to claim 1, characterized in that: The conversation intent recognition model is modified based on the pre-trained language model of the Transformer encoder structure, where: Replacing the attention layer of the pre-trained language model with the multi-head self-attention layer; And the feedforward network layer is added after the multi-head self-attention layer to obtain the conversation intention recognition model.

3. The method for recognizing claims dialogue intention according to claim 1, wherein: The step of performing vector embedding representation on the claim question and answer corpus through the vector embedding layer to obtain a text feature vector corresponding to the claim question and answer corpus specifically includes: Performing text segmentation processing on the claim question and answer corpus to obtain initial text segmentation; Performing semantic recognition on the initial text segmentation to obtain claim question and answer segmentation; Determining the position of the claim question and answer segmentation word in the claim question and answer corpus to obtain segmentation word position information; Loading the bag-of-words model preset in the vector embedding layer, and using the bag-of-words model to perform word vector embedding representation on the claim question and answer segmentation to obtain a text segmentation vector; Encoding the word segmentation position information to obtain a word segmentation position vector; The text word segmentation vector and the word segmentation position vector are combined to obtain the text feature vector.

4. The method for recognizing claims dialogue intention according to claim 2, wherein: The step of using the multi-head self-attention layer to perform feature weighted combination on the text feature vector to obtain a weighted combination vector specifically includes: Obtaining a query sub-vector, a key sub-vector, and a value sub-vector corresponding to the text feature vector based on the multi-head self-attention layer; Calculating a multi-head attention weight of the text feature vector based on the query sub-vector and the key sub-vector; Perform a dot product operation or a linear transformation on the multi-head attention weight and the value sub-vector to obtain the weighted combination vector.

5. The method for recognizing claims dialogue intention according to claim 4, characterized in that: The step of calculating the multi-head attention weight of the text feature vector based on the query sub-vector and the key sub-vector specifically includes: Calculating a feature vector similarity between the query subvector and the key subvector to obtain a first feature similarity; A weighting operation is performed according to the similarity of the feature vectors to obtain the multi-head attention weight.

6. The method for recognizing claims dialogue intention according to claim 2, characterized in that: The step of performing a nonlinear feature transformation on the weighted combination vector through the feedforward network layer to obtain a multi-dimensional feature vector specifically includes: Performing a dimensionality increase operation on the weighted combination vector through the feedforward network layer to obtain a high-dimensional feature vector; Performing a dimensionality reduction operation on the weighted combination vector through the feedforward network layer to obtain a low-dimensional feature vector; Mapping the high-dimensional feature vector and the low-dimensional feature vector into the same feature space; Multi-dimensional feature extraction is performed in the feature space to obtain the multi-dimensional feature vector.

7. The method for recognizing claims dialogue intention according to claim 1, wherein: The step of matching the multi-dimensional feature vector with the intent classification label in the classification layer and obtaining the score of each intent classification label based on a preset scoring rule specifically includes: Mapping the multi-dimensional feature vectors to the dimensions of the intent classification label respectively; Get the label embedding vector corresponding to each intent classification label; Calculating the feature vector similarity between the multi-dimensional feature vector and each of the label embedding vectors to obtain a second feature similarity; weighting the intent classification label based on the second feature similarity to obtain a classification label weight; The initial scores corresponding to the respective intent classification labels are obtained, and the final scores of the respective intent classification labels are calculated in combination with the classification label weights.

8. A claim dialogue intention recognition device, characterized in that: include: A corpus acquisition module is used to obtain pre-collected claims question and answer corpus and input the claims question and answer corpus into a preset dialogue intent recognition model, wherein the dialogue intent recognition model includes a vector embedding layer, a multi-head self-attention layer, a feedforward network layer, and a classification layer; An embedding representation module, configured to perform vector embedding representation on the claim question and answer corpus through the vector embedding layer to obtain a text feature vector corresponding to the claim question and answer corpus; A feature weighting module, configured to perform feature weighted combination on the text feature vector using the multi-head self-attention layer to obtain a weighted combination vector; A feature transformation module, configured to perform a feature nonlinear transformation on the weighted combination vector through the feedforward network layer to obtain a multi-dimensional feature vector; A label matching module is used to match the multi-dimensional feature vector with the intent classification label in the classification layer, and obtain the score of each intent classification label based on a preset scoring rule; The intent recognition module is used to calculate the probability distribution of the classification labels based on the scores of each of the intent classification labels, and determine the intent of the claims dialogue based on the probability distribution.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the method for identifying the intention of a claims dialogue as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the method for identifying the intention of a claims dialogue as described in any one of claims 1 to 7.

Images