Insurance scene intention recognition method and device, electronic equipment and storage medium
By using convolutional neural networks and generative adversarial networks for intent recognition in insurance scenarios, combined with BERT models and FP-Growth algorithms, the problem of low intent recognition accuracy in non-motor insurance fields is solved, achieving higher intent recognition accuracy and process optimization.
Patent Information
- Application Number
- CN202510913019.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-10-17
AI Technical Summary
Existing technologies have low accuracy in intent recognition in non-auto insurance fields, especially in health insurance and property insurance, where semantic expression differs significantly from that in auto insurance, making it difficult for intent recognition to meet the needs.
An intent recognition model trained on a first convolutional neural network and feature vector samples is adopted, combined with the BERT model to obtain word embedding vectors, generative adversarial networks are used to expand the dataset, and domain recognition models are used to reduce domain differences. The FP-Growth algorithm is used to mine intent chains.
It improved the accuracy of intent recognition in non-motor insurance sectors, enhanced the ability to recognize customer intent in insurance scenarios, and promoted process optimization and improved customer experience.
Smart Images
Figure CN120804258A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, and particularly relates to an insurance scenario intent recognition method and device, electronic equipment and a storage medium. BACKGROUND
[0002] Both the financial and medical fields can involve insurance business. In the customer service scenario of the insurance industry, the dialogue data between the agents and the customers contains rich intent information. The recognition and analysis of the intent information have important value for optimizing the service process and improving the customer experience. However, the current insurance dialogue data has a significant domain imbalance problem. For the car insurance field, due to the large business volume and rich historical accumulation, the dialogue data is sufficient and well-annotated, and the intent information is relatively direct and simple to recognize. However, the dialogue data of the non-car insurance field (such as health insurance, life insurance, and property insurance) is relatively sparse, and the implied intent is complex. For example, the health insurance involves "medical expense dispute" and "reimbursement ratio calculation", and the property insurance includes "disaster claim identification" and "clause exemption clause explanation". The semantic expression of the non-car insurance field is significantly different from that of the car insurance field, which makes it difficult for the existing solutions to meet the precise requirements of intent recognition in the non-car insurance field. SUMMARY
[0003] The present application aims to provide an insurance scenario intent recognition method, device, electronic equipment and storage medium to solve the technical problem of low precision of intent recognition in the non-car insurance field in the prior art.
[0004] The technical solution of the present application is as follows. An insurance scenario intent recognition method is provided, comprising:
[0005] obtaining dialogue text data;
[0006] inputting the dialogue text data into an intent recognition model to obtain a probability distribution of intent labels, forming an intent sequence according to the intent label with the largest probability, wherein the intent recognition model is trained according to a first convolutional neural network and a feature vector sample, the feature vector sample is obtained by obtaining a word embedding vector from a dialogue text data sample, creating a car insurance field embedding vector and a non-car insurance field embedding vector, and obtaining the word embedding vector, the car insurance field embedding vector and the non-car insurance field embedding vector by concatenating them;
[0007] generating an intent chain according to the intent sequence.
[0008] Further, the dialogue text data is obtained, comprising:
[0009] The dialog record of a customer and an agent in an insurance field scene, an implicit intent dataset and an insurance field corpus are obtained, the dialog record, the implicit intent dataset and the insurance field corpus are marked with explicit intent and / or implicit intent to obtain insurance scene data, the insurance scene data is cleaned, and the cleaned data is processed by a word segmenter to obtain dialog text data.
[0010] Further, the dialog text data is input into an intent recognition model to obtain a probability distribution of intent labels, including:
[0011] The dialog text data is input into a full connection layer of the intent recognition model and is nonlinearly transformed by a ReLU activation function, and a Softmax function is used to generate a probability distribution of intent labels.
[0012] Further, a word embedding vector is obtained according to the dialog text data sample, including:
[0013] The dialog text data sample is input into a BERT model, the dialog text data sample is encoded by a Transformer encoder to obtain a word embedding vector.
[0014] Further, the insurance scene intent recognition method further includes: in the training process of the first convolutional neural network, the data set of the feature vector sample is expanded by a trained generative adversarial network; the training process of the generative adversarial network includes: the feature vector sample is input into a generator multi-layer perceptron network structure to generate an implicit intent label, and a real implicit intent label and a generated implicit intent label are input into a discriminator multi-layer perceptron network structure to determine the authenticity of the generated implicit intent label.
[0015] Further, before the feature vector sample is input into the generator multi-layer perceptron network structure, it further includes:
[0016] The feature vector sample is input into a full connection layer of a domain recognition model and is nonlinearly transformed by a ReLU activation function, a Sigmoid function is used to generate a probability distribution of domain labels, and the feature vector sample is adjusted according to the probability distribution of the domain labels; the domain recognition model is trained according to a second convolutional neural network and the feature vector sample.
[0017] Further, an intent chain is generated according to the intent sequence, including:
[0018] The minimum support and the minimum confidence of the mining parameters are set, the intent sequence is mined for frequent patterns, the frequently occurring intent combination order is obtained, and an intent chain is generated.
[0019] Another technical solution of the present application is as follows, providing an insurance scene intent recognition device, comprising a data acquisition module, an intent sequence acquisition module and an intent chain generation module;
[0020] The data acquisition module is configured to acquire dialogue text data.
[0021] The intent sequence acquisition module is configured to input the dialogue text data into an intent recognition model to obtain a probability distribution of intent labels, form an intent sequence according to the intent label with the maximum probability, and train the intent recognition model according to a first convolutional neural network and a feature vector sample.
[0022] The intent chain generation module is configured to generate an intent chain according to the intent sequence.
[0023] Another technical solution of the present application is as follows, providing an electronic device comprising a memory and a processor, wherein the memory stores a computer program executable by the processor, and the processor implements the insurance scene intent recognition method according to any one of the above technical solutions when executing the computer program.
[0024] Another technical solution of the present application is as follows, providing a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the insurance scene intent recognition method according to any one of the above technical solutions.
[0025] The present application has the advantages that dialogue text data is acquired, the dialogue text data is input into an intent recognition model to obtain a probability distribution of intent labels, an intent sequence is formed according to the intent label with the maximum probability, the intent recognition model is trained according to a first convolutional neural network and a feature vector sample, the feature vector sample is obtained by acquiring a word embedding vector according to a dialogue text data sample, creating a car insurance domain embedding vector and a non-car insurance domain embedding vector, and splicing the word embedding vector, the car insurance domain embedding vector and the non-car insurance domain embedding vector, an intent chain is generated according to the intent sequence, and the above technical solution can realize intent recognition of a customer in an insurance scene and improve the accuracy of intent recognition in a non-car insurance domain. BRIEF DESCRIPTION OF DRAWINGS
[0026] Figure 1 A flowchart of an insurance scene intent recognition method provided by the present application embodiment is shown in the figure.
[0027] Figure 2A structural schematic diagram of an insurance scenario intent recognition device provided by an embodiment of the present application is shown in the figure.
[0028] Figure 3 A structural schematic diagram of an electronic device provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION
[0029] The technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.
[0030] In the description of the present application, the words "first", "second", etc. are only used for distinguishing the purpose of description, and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying sequence. In the present specification, the terms "comprise", "contain", "have" and their variants all mean "including but not limited to", unless otherwise specifically emphasized.
[0031] In the description of the embodiments of the present application, unless otherwise specified, " / " represents or, for example, A / B can represent A or B; "and / or" in the text only describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, in the description of the embodiments of the present application, "multiple" means two or more than two, and other quantifiers similar thereto should be understood. The preferred embodiments described herein are only used to illustrate and explain the present application, and do not limit the present application, and the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.
[0032] In this text, the reference to "embodiments" means that the specific features, structures or properties described in conjunction with the embodiments can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor is it an independent or alternative embodiment to other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.
[0033] Figure 1 A flowchart of an insurance scenario intent recognition method of an embodiment of the present application is shown in the figure. It should be noted that the insurance scenario intent recognition method of the present application is not limited to the flow order shown in the figure if there is substantially the same result. As shown in the figure, the insurance scenario intent recognition method mainly includes the following steps: Figure 1 Figure 1
[0034] S101, obtain dialogue text data;
[0035] In an optional embodiment, obtaining dialogue text data comprises:
[0036] Obtaining dialogue records of customers and agents in the insurance field, an implicit intent dataset and an insurance field corpus, performing explicit intent labeling and / or implicit intent labeling on the dialogue records, the implicit intent dataset and the insurance field corpus to obtain insurance scenario data, cleaning the insurance scenario data, and performing word segmentation on the cleaned data using a word segmenter to obtain dialogue text data.
[0037] In some embodiments, dialogue records of customers and agents in various scenarios of car insurance and non-car insurance are collected, and the dialogue records can include dialogue text, intent labels, timestamps, etc. Artificially annotated implicit intent datasets and publicly available insurance field corpora can also be collected, and the insurance field corpora include insurance clause texts and common question banks. Explicit intent labeling and / or implicit intent labeling are performed on the dialogue records, the implicit intent dataset and the insurance field corpus to obtain insurance scenario data. The explicit intent labeling can be annotated by domain experts to indicate the intent, such as claim consultation, policy inquiry, objection processing, etc. The implicit intent labeling can be annotated by pre-set rules and review methods to indicate the implicit intent, such as the customer saying that the service is too bad, which can be annotated as complaint tendency.
[0038] In some embodiments, cleaning the insurance scenario data can include removing irrelevant characters (such as special symbols, repeated spaces), and unifying insurance terms (such as combining compulsory insurance for motor vehicle traffic accident liability and compulsory insurance for motor vehicle traffic accident liability into the same expression). The word segmentation of the cleaned insurance scenario data using a word segmenter can include using a WordPiece segmenter of a BERT pre-training model to segment the cleaned insurance scenario data (text), which includes segmenting dialogue text exceeding the maximum length by periods or paragraphs to obtain segmented text, i.e. dialogue text data.
[0039] S102, inputting the dialogue text data into an intent recognition model to obtain a probability distribution of intent labels, forming an intent sequence according to the intent label with the largest probability, the intent recognition model being trained according to a first convolutional neural network and a feature vector sample, the feature vector sample being obtained by acquiring a word embedding vector from a dialogue text data sample, creating a car insurance field embedding vector and a non-car insurance field embedding vector, and obtaining the word embedding vector, the car insurance field embedding vector and the non-car insurance field embedding vector after splicing;
[0040] In some embodiments, the dialogue text data sample forming process is similar to the dialogue text data, and dialogue records of customers and agents in various scenarios of car insurance and non-car fields are collected, and an artificial annotated implicit intent dataset and a public insurance field corpus are collected to form an insurance scenario data sample; the insurance scenario data sample is annotated with explicit intent and / or implicit intent to obtain the dialogue text data sample.
[0041] In an optional implementation, the dialogue text data is input into an intent recognition model to obtain a probability distribution of intent labels, including:
[0042] The dialogue text data is input into a full connection layer of the intent recognition model, and a ReLU activation function is used for nonlinear transformation, and a Softmax function is used to generate a probability distribution of intent labels.
[0043] In some embodiments, the dialogue text data is input into a full connection layer of the intent recognition model, and a ReLU activation function is used for nonlinear transformation, and a Softmax function is used to generate a probability distribution of intent labels. The full connection layer is an intent classifier layer used to predict the explicit intent and implicit intent of the dialogue. When training the intent recognition model, a word embedding vector is obtained according to the dialogue text data sample, a car insurance field embedding vector and a non-car insurance field embedding vector are created, the word embedding vector, the car insurance field embedding vector and the non-car insurance field embedding vector are spliced to obtain a feature vector sample, and the intent recognition model is trained according to the first convolutional neural network and the feature vector sample.
[0044] In an optional implementation, the word embedding vector is obtained according to the dialogue text data sample, including:
[0045] The dialogue text data sample is input into a BERT model, and the dialogue text data sample is encoded by a Transformer encoder to obtain a word embedding vector.
[0046] In some embodiments, after obtaining the word embedding vector, a car insurance field embedding vector and a non-car insurance field embedding vector are created for each of the car insurance field and the non-car insurance field, the initial values of the car insurance field embedding vector and the non-car insurance field embedding vector are generated using a normal distribution or a uniform distribution, and these embedding vectors are learnable parameters that are automatically adjusted in the subsequent model training process.
[0047] The word embedding vector is obtained according to the dialogue text data sample, including:
[0048] The dialogue text data sample is input into a BERT model, and the dialogue text data sample is encoded by a Transformer encoder to obtain a word embedding vector.
[0049] In an optional implementation, the insurance scene intent recognition method further includes: in the training process of the first convolutional neural network, expanding the data set of the feature vector sample by a trained generative adversarial network; the training process of the generative adversarial network includes: inputting the feature vector sample into a generator multi-layer perceptron network structure to generate an implicit intent label, and inputting the real implicit intent label and the generated implicit intent label into a discriminator multi-layer perceptron network structure to determine the authenticity of the generated implicit intent label.
[0050] In some embodiments, the data set of the dialogue text data sample is expanded by a trained generative adversarial network, specifically, an implicit intent label, i.e., an implicit intent data set, is generated by the trained generative adversarial network, so as to expand the data set of the dialogue text data sample. The training process of the generative adversarial network can further include defining a loss function of the generator and the discriminator, and training using an Adam optimizer. The loss function of the generator is to minimize the probability that the generated sample is judged as generated by the discriminator, i.e., to maximize the probability that the discriminator considers the generated sample as real. The loss function of the discriminator is to maximize the probability that the real sample is real and the generated sample is generated.
[0051] In an optional implementation, before the feature vector sample is input into the generator multi-layer perceptron network structure, it further includes:
[0052] The feature vector sample is input into a full connection layer of a domain recognition model, and is subjected to nonlinear transformation by a ReLU activation function. A Sigmoid function is used to generate a probability distribution of a domain label, and the feature vector sample is adjusted according to the probability distribution of the domain label. The domain recognition model is trained according to a second convolutional neural network and the feature vector sample.
[0053] In some embodiments, the feature vector sample is input into a full connection layer of a domain recognition model, and is subjected to nonlinear transformation by a ReLU activation function. A Sigmoid function is used to generate a probability distribution of a domain label. The full connection layer is a domain classifier layer, which is used to distinguish the features of car insurance and non-car insurance. During model training, common features across domains are extracted through adversarial training. Adversarial training can refer to an adversarial process between a feature extractor that extracts word embedding vectors from original text and a domain classifier. The feature extractor extracts semantic features from the original text and tries to remove domain-specific features so that the domain classifier cannot distinguish whether the features come from which domain. However, the domain classifier wants to distinguish whether the features extracted by the feature extractor come from the car insurance domain or the non-car insurance domain. Through this adversarial process, the feature extractor is forced to extract common features across domains, reducing the impact of domain differences on model performance.
[0054] S103, generate an intent chain according to the intent sequence.
[0055] In an optional embodiment, generating the intent chain according to the intent sequence comprises:
[0056] Setting a minimum support degree and a minimum confidence degree of mining parameters, performing frequent pattern mining on the intent sequence, obtaining a frequently-occurring intent combination sequence, and generating the intent chain.
[0057] In some embodiments, using a trained intent recognition model, performing intent recognition on each conversation according to paragraph segmentation (conversation text data), taking the intent label with the maximum probability, combining into a sequence in chronological order to obtain the intent sequence, using the FP-Growth algorithm to perform frequent pattern mining, setting the mining parameters of the minimum support degree (the minimum frequency of the pattern in the data set) and the minimum confidence degree (the probability of containing intent B in the transaction containing intent A), and then inputting the intent sequence into the mining algorithm to obtain the frequently-occurring (highest frequency) intent combination sequence to generate the intent chain, for example, the intent chain can be inquiry product->policy query->quotation->decide to purchase.
[0058] In some embodiments, according to the mined intent chain and the corresponding transaction, the next step of the customer's intent can be guided to the intent chain with a higher transaction probability, and the transaction conversion rate can be improved; the next possible intent of the customer can be predicted according to the mined intent chain, and relevant information can be prepared in advance; potential unreasonable service processes can be found by analyzing the distribution and changes of the intent chain, and the processes can be improved to reduce unnecessary steps and improve customer experience.
[0059] The insurance scenario intent recognition method provided by the embodiment of the application obtains conversation text data; inputs the conversation text data into an intent recognition model to obtain a probability distribution of intent labels, forms an intent sequence according to the intent label with the maximum probability, the intent recognition model is obtained according to a first convolutional neural network and a feature vector sample, the feature vector sample is obtained by obtaining a word embedding vector according to a conversation text data sample, creating a car insurance field embedding vector and a non-car insurance field embedding vector, and obtaining the word embedding vector, the car insurance field embedding vector and the non-car insurance field embedding vector after splicing; generates an intent chain according to the intent sequence; the intent recognition of the customer in the insurance scenario can be realized, and the accuracy of the intent recognition in the non-car insurance field can be improved.
[0060] The insurance scene intent recognition method provided by the embodiment of the present application extracts common features irrelevant to the field and retains specific information through learnable field embedding vectors and adversarial training, generates implicit intent labels using a generative adversarial network (GAN), enhances the diversity of training data, alleviates the labeling sparsity problem, then mines high-frequency intent chains appearing in large data sets through the FP-Growth algorithm, helps locate service bottlenecks, promotes process optimization, and ultimately improves customer experience. The entire process realizes closed-loop optimization from data to business scenarios by reusing car insurance knowledge, reducing field differences through adversarial training, and mining intent sequences. The insurance scene intent recognition method provided by the embodiment of the present application uses transfer learning technology to reuse the semantic representation ability of the car insurance scene, reduces field differences through adversarial training, and improves the robustness and generalization ability of the model for intent recognition in non-car insurance fields.
[0061] The insurance scene intent recognition method provided by the embodiment of the present application can be constructed based on artificial intelligence, and the related data can be acquired and processed based on artificial intelligence technology to realize unattended artificial intelligence insurance scene intent recognition. 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.
[0062] 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. The software technology of artificial intelligence mainly includes computer vision technology, robot technology, biometric technology, speech processing technology, natural language processing technology, and machine learning / deep learning, etc.
[0063] Figure 2 is a structural schematic diagram of the insurance scene intent recognition device of the embodiment of the present application, as Figure 2 shown, the insurance scene intent recognition device 20 includes a data acquisition module 21, an intent sequence acquisition module 22, and an intent chain generation module 23;
[0064] The data acquisition module 21 is configured to acquire dialogue text data.
[0065] The intention sequence acquisition module 22 is configured to input the dialogue text data into an intention recognition model to obtain a probability distribution of intention labels, and form an intention sequence according to an intention label with the maximum probability, wherein the intention recognition model is trained according to a first convolutional neural network and a feature vector sample, the feature vector sample is obtained by acquiring a word embedding vector according to a dialogue text data sample, creating a car insurance field embedding vector and a non-car insurance field embedding vector, and concatenating the word embedding vector, the car insurance field embedding vector and the non-car insurance field embedding vector.
[0066] The intention chain generation module 23 is configured to generate an intention chain according to the intention sequence.
[0067] In an optional implementation, the data acquisition module 21 acquires dialogue text data, including:
[0068] acquiring dialogue records of a customer and an agent, an implicit intention data set and an insurance field corpus in an insurance field scenario, performing explicit intention labeling and / or implicit intention labeling on the dialogue records, the implicit intention data set and the insurance field corpus to obtain insurance scenario data, performing cleaning on the insurance scenario data, and performing word segmentation processing on the cleaned data by using a word segmenter to obtain dialogue text data.
[0069] In an optional implementation, the intention sequence acquisition module 22 inputs the dialogue text data into an intention recognition model to obtain a probability distribution of intention labels, including:
[0070] inputting the dialogue text data into a fully connected layer of the intention recognition model, performing nonlinear transformation by using a ReLU activation function, and generating a probability distribution of intention labels by using a Softmax function.
[0071] In an optional implementation, the word embedding vector is obtained according to the dialogue text data sample, including:
[0072] inputting the dialogue text data sample into a BERT model, encoding the dialogue text data sample by using a Transformer encoder to obtain a word embedding vector.
[0073] In an optional embodiment, the insurance scenario intent recognition apparatus 20 further comprises a dataset expansion module, configured to expand the dataset of the feature vector samples by a trained generative adversarial network during the training of the first convolutional neural network; the training of the generative adversarial network comprises inputting the feature vector samples into a generator multi-layer perceptron network structure to generate implicit intent labels, and inputting the real implicit intent labels and the generated implicit intent labels into a discriminator multi-layer perceptron network structure to determine the authenticity of the generated implicit intent labels.
[0074] In an optional embodiment, before the feature vector samples are input into the generator multi-layer perceptron network structure, the method further comprises:
[0075] inputting the feature vector samples into a full connection layer of a domain recognition model, performing nonlinear transformation by a ReLU activation function, generating a probability distribution of domain labels by a Sigmoid function, and adjusting the feature vector samples according to the probability distribution of the domain labels; the domain recognition model is trained according to a second convolutional neural network and the feature vector samples.
[0076] In an optional embodiment, the intent chain generation module 23 generates an intent chain according to the intent sequence, comprising:
[0077] setting a minimum support degree and a minimum confidence degree of mining parameters, performing frequent pattern mining on the intent sequence, obtaining frequently-occurring intent combination sequences, and generating an intent chain.
[0078] Figure 3 is a structural schematic diagram of an electronic device of an embodiment of the present application. As shown in Figure 3 The electronic device 30 comprises a processor 31 and a memory 32 in communication connection with the processor 31.
[0079] The memory 32 stores program instructions for implementing the insurance scenario intent recognition method of any of the above embodiments.
[0080] The processor 31 is configured to execute the program instructions stored in the memory 32 to perform insurance scenario intent recognition.
[0081] The processor 31 can also be referred to as a CPU (Central Processing Unit). The processor 31 can be an integrated circuit chip having a processing capability of signals. The processor 31 can also be a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.
[0082] The storage medium of the embodiment of the present application stores program instructions capable of implementing all the methods described above. The storage medium can be non-volatile or volatile. The program instructions can be stored in the storage medium in the form of a software product, including a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a ROM (Read-Only Memory), a RAM (Random Access Memory), a magnetic disk or an optical disk, and various media capable of storing program codes, or a terminal device such as a computer, a server, a mobile phone, a tablet, etc.
[0083] In several embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented by other means. For example, the apparatus embodiments described above are only schematic, and the division of modules is only a logical function division. In actual implementation, another division manner can be adopted, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed modules can be indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.
[0084] In addition, each functional module in each embodiment of the present application can be integrated in one processing unit, or each module can exist physically separately, or two or more modules can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit. The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation obtained by using the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.
[0085] The above merely describes the embodiments of the present application, and it should be pointed out that, for those skilled in the art, improvements can be made without departing from the inventive concept, and these all belong to the protection scope of the present application.
Claims
1. A method for identifying intentions in insurance scenarios, characterized in that: include: Get conversation text data; Inputting the conversation text data into an intent recognition model to obtain a probability distribution of intent labels, and forming an intent sequence based on the intent label with the highest probability, wherein the intent recognition model is trained based on a first convolutional neural network and feature vector samples, wherein the feature vector samples are obtained by obtaining word embedding vectors based on the conversation text data samples, creating an embedding vector for the auto insurance domain and an embedding vector for the non-auto insurance domain, and concatenating the word embedding vector, the auto insurance domain embedding vector, and the non-auto insurance domain embedding vector; Generate an intent chain according to the intent sequence.
2. The insurance scenario intention recognition method according to claim 1 is characterized in that: Get conversation text data, including: Obtain conversation records between customers and agents in insurance scenarios, implicit intent datasets, and insurance field corpus, perform explicit intent annotation and / or implicit intent annotation on the conversation records, the implicit intent dataset, and the insurance field corpus to obtain insurance scenario data, clean the insurance scenario data, and use a word segmenter to perform word segmentation on the cleaned data to obtain conversation text data.
3. The insurance scenario intention recognition method according to claim 1 is characterized in that: Input the conversation text data into the intent recognition model to obtain the probability distribution of intent labels, including: The conversation text data is input into the fully connected layer of the intent recognition model, and nonlinear transformation is performed through the ReLU activation function, and the Softmax function is used to generate the probability distribution of the intent label.
4. The insurance scenario intention recognition method according to claim 1, characterized in that: Obtaining a word embedding vector according to the conversation text data sample includes: The conversation text data sample is input into the BERT model, and the conversation text data sample is encoded through the Transformer encoder to obtain a word embedding vector.
5. The insurance scenario intention recognition method according to claim 1 is characterized in that: The insurance scenario intention recognition method also includes: during the training process of the first convolutional neural network, expanding the data set of the feature vector samples through the trained generative adversarial network; the training process of the generative adversarial network includes inputting the feature vector samples into the generator multi-layer perceptron network structure to generate implicit intent labels, and inputting the true implicit intent labels and the generated implicit intent labels into the discriminator multi-layer perceptron network structure to determine the authenticity of the generated implicit intent labels.
6. The insurance scenario intention recognition method according to claim 5 is characterized in that: Before the feature vector sample is input into the generator multi-layer perceptron network structure, the method further includes: The feature vector sample is input into the fully connected layer of the domain recognition model and nonlinearly transformed using the ReLU activation function. The probability distribution of the domain label is generated using the Sigmoid function, and the feature vector sample is adjusted according to the probability distribution of the domain label. The domain recognition model is trained based on the second convolutional neural network and the feature vector sample.
7. The insurance scenario intention recognition method according to claim 1, characterized in that: Generate an intent chain according to the intent sequence, including: The minimum support and minimum confidence of the mining parameters are set, and frequent pattern mining is performed on the intent sequence to obtain the frequently occurring sequence of intent combinations to generate an intent chain.
8. An insurance scenario intention recognition device, characterized in that: Includes data acquisition module, intent sequence acquisition module and intent chain generation module; The data acquisition module is used to acquire conversation text data; The intent sequence acquisition module is configured to input the conversation text data into an intent recognition model to obtain a probability distribution of intent labels, and form an intent sequence based on the intent label with the highest probability. The intent recognition model is trained using a first convolutional neural network and feature vector samples. The feature vector samples are obtained by obtaining word embedding vectors based on the conversation text data samples, creating an embedding vector for the auto insurance domain and an embedding vector for the non-auto insurance domain, and concatenating the word embedding vector, the auto insurance domain embedding vector, and the non-auto insurance domain embedding vector. The intention chain generation module is used to generate an intention chain according to the intention sequence.
9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable by the processor, wherein: When the processor executes the computer program, the insurance scenario intention recognition method as described in any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the insurance scenario intention recognition method as described in any one of claims 1 to 7 is implemented.