Insurance entity identification method and device, electronic equipment and storage medium
By combining the BERT model and graph neural networks, the embedding vectors and edge relationship sets of insurance texts are obtained, solving the problem of accuracy in entity recognition in the insurance and finance field. This enables effective understanding of professional terms and complex contexts, improving the accuracy and robustness of information extraction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-22
- Publication Date
- 2026-04-10
AI Technical Summary
Existing general NER models suffer from errors in entity boundary identification, misclassification, and insufficient generalization ability in complex contexts of insurance and financial information, resulting in inaccurate information extraction.
By combining the BERT model and graph neural network, we obtain the embedding vectors and edge relationship sets of insurance texts. By fusing and generating fused vectors and decoding them, we can explicitly capture the grammatical dependencies and semantic associations between words, thereby improving the accuracy of insurance financial entity identification.
It improves the accuracy and robustness of identifying insurance and financial entities, better understands professional terminology and complex contexts, and enhances the accuracy of information extraction.
Smart Images
Figure CN121835682A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of insurance text recognition, in particular to an insurance entity recognition method and device, an electronic device and a storage medium. BACKGROUND
[0002] With the rapid development of artificial intelligence technology in the field of natural language processing, named entity recognition (NER) as one of the core tasks of information extraction has been widely used in knowledge graph construction, intelligent question answering system and other scenarios.
[0003] However, the insurance and finance field as a professional vertical field, its text has characteristics significantly different from general corpus: on the one hand, it involves a large number of actuarial terms (such as "deductible" "waiting period"), exclusive entity types (such as "policy number" "claim coding") and complex clause expression structures; on the other hand, due to the sensitivity of industry data, the scale of available high-quality labeled samples is limited. The existing general NER model in the complex context of insurance and finance information, the existing model often has entity boundary recognition errors, class misjudgment and insufficient generalization ability, which seriously restricts the accuracy and practicality of information extraction. SUMMARY
[0004] The purpose of the embodiments of the present application is to provide an insurance entity recognition method, device, electronic device and storage medium, to solve the technical problem of insufficient information extraction accuracy of the existing general NER model in the complex context of insurance and finance information.
[0005] In a first aspect, the present application provides an insurance entity recognition method, which comprises: obtaining an embedding vector set of the insurance text to be recognized, the embedding vector set comprising an embedding vector corresponding to each character in the insurance text to be recognized; obtaining an edge relationship set of the insurance text to be recognized, the edge relationship set comprising syntactic dependency information of each word in the insurance text to be recognized; fusing the embedding vector set and the edge relationship set to generate a fusion vector of the insurance text to be recognized; and decoding the fusion vector to obtain an entity annotation result in the insurance text to be recognized.
[0006] In an optional implementation, the embedding vector set of the insurance text to be recognized is obtained in the following manner: The insurance text to be recognized is input into a pre-trained BERT model to obtain a first embedding vector corresponding to each character in the insurance text to be recognized output by the BERT model as the embedding vector set.
[0007] In an optional implementation, the embedding vector set of the insurance text to be recognized is obtained in the following manner: inputting the to-be-identified insurance text into a pre-trained BERT model to obtain a first embedding vector corresponding to each character in the to-be-identified insurance text output by the BERT model as a character embedding sequence; inputting the character embedding sequence into a bidirectional long short-term memory network to obtain a second embedding vector corresponding to each character in the to-be-identified insurance text output by the bidirectional long short-term memory network as an embedding vector set.
[0008] In an optional implementation, the to-be-identified insurance text is also segmented to determine all words in the to-be-identified insurance text.
[0009] In an optional implementation, the edge relation set of the to-be-identified insurance text is obtained in the following manner: inputting the to-be-identified insurance text into a pre-trained graph neural network model to obtain an edge relation set output by the graph neural network model.
[0010] In an optional implementation, for each word in the to-be-identified insurance text, an embedding vector sequence corresponding to the word is averaged and pooled to obtain a word vector sequence corresponding to the word. generating a word vector set of the to-be-identified insurance text based on all word vector sequences; fusing the word vector set and the edge relation set to generate a fusion vector of the to-be-identified insurance text.
[0011] In an optional implementation, the graph neural network model is trained in the following manner: obtaining general sample texts and performing syntax dependency relation annotation; inputting the annotated general sample texts into the graph neural network model to make it output corresponding syntax dependency graphs, the syntax dependency graphs at least including syntax dependency information of each word in the text, the syntax dependency information being used to indicate a head word, a dependent word, and a dependency relation type between the head word and the dependent word.
[0012] In a second aspect, the present application provides an insurance entity recognition device, which comprises: a first processing module configured to obtain an embedding vector set of a to-be-identified insurance text, the embedding vector set including an embedding vector corresponding to each character in the to-be-identified insurance text; a second processing module configured to obtain an edge relation set of the to-be-identified insurance text, the edge relation set including syntax dependency information of each word in the to-be-identified insurance text; a fusion module configured to fuse the embedding vector set and the edge relation set to generate a fusion vector of the to-be-identified insurance text; a decoding module configured to decode the fusion vector to obtain an entity annotation result in the to-be-identified insurance text.
[0013] In a third aspect, the present application provides an electronic device, comprising a processor, a memory and a bus, the memory storing machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the processor executes the machine readable instructions to perform the steps of the method for identifying insurance entities according to any one of the preceding embodiments.
[0014] In a fourth aspect, the present application provides a computer readable storage medium, the computer readable storage medium storing a computer program, when the computer program is run by a processor, the steps of the method for identifying insurance entities according to any one of the preceding embodiments are performed.
[0015] The present application provides a method and device for identifying insurance entities, an electronic device and a storage medium. The method comprises obtaining an embedding vector set of a to-be-identified insurance text, the embedding vector set comprising an embedding vector corresponding to each character in the to-be-identified insurance text; obtaining a edge relation set of the to-be-identified insurance text, the edge relation set comprising syntactic dependency information of each word in the to-be-identified insurance text; fusing the embedding vector set and the edge relation set to generate a fusion vector of the to-be-identified insurance text; and decoding the fusion vector to obtain an entity annotation result in the to-be-identified insurance text. By introducing a dependency parsing graph structure, the present application can explicitly capture the syntactic dependency and semantic association between words in the field text, thereby enhancing the understanding of professional vocabulary and complex context and effectively improving the accuracy and robustness of insurance financial entity recognition. BRIEF DESCRIPTION OF DRAWINGS
[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.
[0017] Figure 1 A flowchart of a method for identifying insurance entities provided by the embodiments of the present application; Figure 2 A structural schematic diagram of a device for identifying insurance entities provided by the embodiments of the present application; Figure 3 A structural schematic diagram of an electronic device provided by the embodiments of the present application. DETAILED DESCRIPTION
[0018] The technical solutions of the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.
[0019] Embodiment one Figure 1 A flowchart of an insurance entity identification method provided by an embodiment of the present application. As shown in the figure, the insurance entity identification method provided by the present application can be implemented by a deep learning model, and specifically includes the following steps: Figure 1 S10, obtaining an embedding vector set of the insurance text to be identified, the embedding vector set including an embedding vector corresponding to each character in the insurance text to be identified.
[0020] The insurance text to be identified here can be insurance policies, contracts, and other insurance and financial field texts, including a large number of professional expressions and complex structured expressions.
[0021] In step S10, the embedding vector set of the insurance text to be identified can be obtained in the following manner: The insurance text to be identified is input into a pre-trained BERT model to obtain a first embedding vector corresponding to each character in the insurance text to be identified output by the BERT model as the embedding vector set.
[0022] In a feasible embodiment, a pre-trained BERT model can be used to embed the characters in the insurance and financial information field sentences. The BERT model can learn the word representation in the context, provide rich semantic information, and help feature learning in the subsequent steps. Specifically, the embedding representation of the characters in the insurance and financial information sentences is as follows:
[0023] wherein, is a character sequence in the insurance and financial information sentence, wherein, is the nth character in the insurance text to be identified, is the length of the sentence, is the embedding representation of the character sequence by the BERT model, is the first embedding vector corresponding to the nth character in the insurance text to be identified.
[0024] In another feasible embodiment, the embedding vector set of the insurance text to be identified can also be obtained in the following manner: The insurance text to be identified is input into a pre-trained BERT model to obtain a first embedding vector corresponding to each character in the insurance text to be identified output by the BERT model as the embedding vector set. The character embedding sequence is input into a bidirectional long short-term memory network to obtain a second embedding vector corresponding to each character in the insurance text to be identified output by the bidirectional long short-term memory network as the embedding vector set.
[0025] Here, on the basis of the BERT model, BiLSTM can be further used for sequence feature encoding. By adopting a bidirectional long short-term memory network (BiLSTM), the character sequence of the embedding representation is encoded. BiLSTM can effectively capture the context information between words in the sequence and improve the model's understanding of the sentence structure. BiLSTM can be one layer or multiple layers. The input of the first layer BiLSTM is the output E of the BERT model. The input of other layers is the output of the last BiLSTM.
[0026] The calculation formula of LSTM can be:
[0027]
[0028]
[0029]
[0030]
[0031]
[0032] wherein is the input of the input sequence at time step . is the hidden state of the previous time step. is the weight matrix of the input gate. is the weight matrix of the forget gate. is the bias of the input gate. is the bias of the forget gate. is the activation function. tanh is the hyperbolic tangent activation function.
[0033] The forward propagation and backward propagation of BiLSTM are respectively:
[0034]
[0035] wherein, is the first embedding vector input in the BiLSTM model at time step . and are the forward and backward hidden states of the BiLSTM model at time step . The final output of a single second embedding vector result can be represented as: .
[0036] S11, obtain an edge relation set of the insurance text to be recognized, the edge relation set including syntactic dependency information of each word in the insurance text to be recognized.
[0037] In step S11, the insurance text to be recognized can be segmented to determine all the words in the insurance text to be recognized. Then, the edge relation set of the insurance text to be recognized is obtained in the following manner: The insurance text to be recognized is input into a pre-trained graph neural network model to obtain an edge relation set output by the graph neural network model.
[0038] Here, a general domain graph neural network model can be used to analyze the syntactic dependency relationship of the insurance text to be recognized. The syntactic dependency relationship refers to the subordinate and modifying relationship between words in a sentence in terms of grammatical structure, such as subject-predicate relationship, verb-object relationship, and determiner-noun relationship, etc., which reflects the hierarchical structure and semantic association within the sentence.
[0039] Here, the graph neural network model can be trained in the following manner: Obtain general sample texts and perform syntactic dependency relationship annotation. The annotated general sample texts are input into the graph neural network model to make it output corresponding syntactic dependency graphs, the syntactic dependency graphs including at least syntactic dependency information of each word in the text, the syntactic dependency information being used to indicate a head word, a dependent word, and a dependency relationship type between the head word and the dependent word.
[0040] For example, the syntactic dependency annotation of the sentence “Li Lei studies computer science at Peking University.” can be expressed as:
[0041] Here, “study” is the root of the sentence. “Li Lei” is the subject (nsubj) of “study”. “Computer science” is the object of the sentence as a whole, in which “computer” modifies “science”. “At Peking University” is an adverbial phrase, in which “Beijing” modifies “University”. The preposition “at” is attached to the noun “University”. This dependency structure reveals the grammatical hierarchy within the sentence, which helps the graph neural network to capture the semantic and syntactic association between words.
[0042] Then, the syntactic dependency structure can be used by the graph neural network to learn the grammatical hierarchy and semantic relationship of the text.
[0043] S12, fuse the embedding vector set and the edge relation set to generate a fusion vector of the insurance text to be recognized.
[0044] In step S12, for each word in the insurance text to be recognized, the embedding vector sequence corresponding to the word can be averaged-pooled to obtain a word vector sequence corresponding to the word. Based on all the word vector sequences, a word vector set of the insurance text to be recognized is generated. The word vector set and the edge relation set are fused to generate a fusion vector of the insurance text to be recognized.
[0045] Here, a general domain graph neural network model can be used to further encode the sentence based on the syntax dependency graph. The general domain graph neural network model is combined with the embedding representation of the insurance financial information sentence to obtain a deeper encoding of the sentence based on the syntax dependency graph. This step can provide more global and syntax-aware features, which can help to more accurately understand the information in the sentence. The financial sentence is first segmented using a segmentation tool, and then the inputs of the word-level graph neural network are aggregated using the average pooling method.
[0046]
[0047] wherein, represents the embedding vector corresponding to the word after segmentation and average pooling of the output result of the BiLSTM, m is the number of words in the insurance text to be recognized, and D is the edge relation obtained by the graph neural network model based on the syntax dependency graph.
[0048] S13, decoding the fusion vector to obtain the entity annotation result in the insurance text to be recognized.
[0049] Here, a full connection layer can be used for decoding to obtain a BIO annotation result: the insurance financial information sentence with the syntax dependency graph and the BERT embedding representation is annotated through the decoding of the full connection layer to obtain a BIO annotation result. This result can indicate the role of each word in the entity naming, such as start (B), internal (I), or other (O).
[0050] The insurance entity recognition method provided in the present application can explicitly capture the syntax dependency and semantic association between words in the field text by introducing a dependency parsing graph structure, thereby enhancing the understanding of professional vocabulary and complex context, and effectively improving the accuracy and robustness of insurance financial entity recognition.
[0051] Embodiment two In an embodiment of the present application, a processing system based on insurance entity recognition is provided, which can be an intelligent entity recognition tool specially designed for the insurance financial field, aiming to efficiently and accurately extract key information from unstructured insurance texts. The core of the system is based on natural language processing technology, especially deep optimization for professional vocabulary and complex context in the insurance field.
[0052] Users only need to log in to the system through a simple interface and select the model formed by the insurance and financial information entity recognition method based on dependency parsing graphs in this invention. The system can then complete the analysis and processing of common entities (such as the insured's name, policy number, insurance amount, term, etc.) in policy, report and other texts in real time, and output structured and visualized entity recognition results. This greatly improves information processing efficiency and data utilization value, and provides strong technical support for the automation process, intelligent risk control and data analysis of insurance business.
[0053] Example 3 Figure 2 This is a schematic diagram of the structure of an insurance entity identification device provided in an embodiment of this application. Figure 2 As shown, based on the same inventive concept, this application also provides an insurance entity identification device 20, the device comprising: The first processing module 210 is used to obtain an embedding vector set of the insurance text to be identified, the embedding vector set including the embedding vector corresponding to each character in the insurance text to be identified; The second processing module 220 is used to obtain the edge relation set of the insurance text to be identified, which includes the syntactic dependency information of each word in the insurance text to be identified. The fusion module 230 is used to fuse the set of embedded vectors and the set of edge relationships to generate a fused vector of the insurance text to be identified; Decoding module 240 is used to decode the fused vector to obtain entity annotation results in the insurance text to be identified.
[0054] In a preferred embodiment, the first processing module 210 obtains the set of embedding vectors of the insurance text to be identified in the following manner: The insurance text to be identified is input into a pre-trained BERT model to obtain the first embedding vector corresponding to each character in the insurance text to be identified output by the BERT model, which is used as the embedding vector set.
[0055] In a preferred embodiment, the first processing module 210 obtains the set of embedding vectors of the insurance text to be identified in the following manner: The insurance text to be identified is input into the pre-trained BERT model to obtain the first embedding vector corresponding to each character in the insurance text to be identified output by the BERT model, which is used as the character embedding sequence. The character embedding sequence is input into a bidirectional long short-term memory network to obtain the second embedding vector corresponding to each character in the insurance text to be recognized output by the bidirectional long short-term memory network, which is used as the embedding vector set.
[0056] In a preferred embodiment, the process further includes word segmentation of the insurance text to be identified in order to determine all the words in the insurance text.
[0057] In a preferred embodiment, the second processing module 220 obtains the edge relationship set of the insurance text to be recognized in the following manner: inputting the insurance text to be recognized into the pre-trained graph neural network model to obtain an edge relationship set output by the graph neural network model.
[0058] In a preferred embodiment, the fusion module 230 is specifically configured to, for each word in the insurance text to be recognized, perform average pooling on the embedding vector sequence corresponding to the word to obtain a word vector sequence corresponding to the word. generate a word vector set of the insurance text to be recognized based on all the word vector sequences; fuse the word vector set and the edge relationship set to generate a fusion vector of the insurance text to be recognized.
[0059] In a preferred embodiment, the graph neural network model is trained in the following manner: obtain general sample texts and perform syntax dependency relationship annotation; input the annotated general sample texts into the graph neural network model to make it output corresponding syntax dependency graphs, the syntax dependency graphs at least including syntax dependency information of each word in the text, the syntax dependency information being used to indicate a head word, a dependent word, and a dependency relationship type between the head word and the dependent word.
[0060] Embodiment four Please refer to Figure 3 , Figure 3 A structural schematic diagram of an electronic device provided by an embodiment of the present application is shown in FIG. 3. Figure 3 As shown in FIG. 3, the electronic device 300 includes a processor 310, a memory 320, and a bus 330.
[0061] The memory 320 stores machine-readable instructions executable by the processor 310. When the electronic device 300 is running, the processor 310 and the memory 320 communicate through the bus 330. The machine-readable instructions, when executed by the processor 310, can perform the steps of the insurance entity identification method in any of the above method embodiments. For specific implementation manners, please refer to the method embodiments, which will not be described here again.
[0062] The embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program is run by a processor, the steps of the insurance entity identification method in any of the above method embodiments can be performed. For specific implementation manners, please refer to the method embodiments, which will not be described here again.
[0063] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.
[0064] In the embodiments provided in the present application, it should be understood that the disclosed device and method can be implemented by other manners. The device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and another division manner can be used in actual implementation, and for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some communication interfaces, devices or units, and can be electrical, mechanical or other forms.
[0065] In addition, the units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.
[0066] Further, the function modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0067] It should be noted that if the function is realized in the form of a software function module and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, and various program code storage media.
[0068] In this paper, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations.
[0069] The above merely provides an example of the present application, and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A method for identifying an insurance entity, characterized in that, The method includes: Obtain the set of embedding vectors for the insurance text to be identified, wherein the set of embedding vectors includes the embedding vector corresponding to each character in the insurance text to be identified; Obtain the edge relation set of the insurance text to be identified, wherein the edge relation set includes the syntactic dependency information of each word in the insurance text to be identified; The embedded vector set and the edge relation set are fused to generate a fused vector of the insurance text to be identified; The fused vector is decoded to obtain entity annotation results in the insurance text to be identified.
2. The method according to claim 1, characterized in that, The set of embedding vectors for the insurance text to be identified is obtained through the following method: The insurance text to be identified is input into a pre-trained BERT model to obtain the first embedding vector corresponding to each character in the insurance text to be identified output by the BERT model, which is used as the embedding vector set.
3. The method according to claim 1, characterized in that, The set of embedding vectors for the insurance text to be identified is obtained through the following method: The insurance text to be identified is input into a pre-trained BERT model to obtain the first embedding vector corresponding to each character in the insurance text to be identified output by the BERT model, which is used as the character embedding sequence. The character embedding sequence is input into a bidirectional long short-term memory network to obtain the second embedding vector corresponding to each character in the insurance text to be identified output by the bidirectional long short-term memory network, which is used as the embedding vector set.
4. The method according to claim 2 or 3, characterized in that, It also includes segmenting the insurance text to be identified in order to determine all the words in the insurance text.
5. The method according to claim 4, characterized in that, The set of edge relationships of the insurance text to be identified is obtained in the following way: The insurance text to be identified is input into a pre-trained graph neural network model to obtain the set of edge relationships output by the graph neural network model.
6. The method according to claim 5, characterized in that, For each word in the insurance text to be identified, average pooling is performed on the embedding vector sequence corresponding to that word to obtain the word vector sequence corresponding to that word. Based on all word vector sequences, generate a set of word vectors for the insurance text to be recognized; The set of word vectors and the set of edge relationships are fused to generate a fused vector of the insurance text to be identified.
7. The method according to claim 1, characterized in that, The graph neural network model is trained using the following method: Obtain general sample text and annotate its syntactic dependencies; The labeled general sample text is input into the graph neural network model, which outputs the corresponding syntactic dependency graph. The syntactic dependency graph includes at least the syntactic dependency information of each word in the text. The syntactic dependency information is used to indicate the head word, dependent words, and the type of dependency relationship between the head word and dependent words.
8. A device for identifying an insurance entity, characterized in that, The device includes: The first processing module is used to obtain an embedding vector set of the insurance text to be identified, wherein the embedding vector set includes the embedding vector corresponding to each character in the insurance text to be identified; The second processing module is used to obtain the edge relation set of the insurance text to be identified, wherein the edge relation set includes the syntactic dependency information of each word in the insurance text to be identified; The fusion module is used to fuse the embedded vector set and the edge relation set to generate a fused vector of the insurance text to be identified; The decoding module is used to decode the fused vector to obtain entity annotation results in the insurance text to be identified.
9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the identification method for an insurance entity as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the identification method for an insurance entity as described in any one of claims 1 to 7.