A method for processing insurance reports based on table processing language model

Abstract

The present invention relates to an insurance report processing method based on a table processing language model, and to the technical field of large language model application. The present invention encodes the information in various insurance reports into information encoding including semantic position encoding, table position encoding, embedded position encoding and information embedding; trains a language model through information encoding, so that the language model understands the information contained in the insurance report; uses the request target in the question-answer decision data set as a question, and uses the table intermediate processing request for realizing the request target and the operation instruction for realizing the table intermediate processing request as an answer to train the language model, giving the language model the ability to establish a corresponding relationship between the processing request expressed in natural language and the table processing operation instruction; and based on the question-answer decision data set, trains the language model to decompose the request target into a number of intermediate processing requests; and uses the obtained language model to process the insurance report based on the prompt words containing the request target and the insurance report to be processed.

Description

Technical Field

[0001] The present invention relates to the technical field of large language model application, and in particular to an insurance report processing method based on a table processing language model. Background Art

[0002] In the insurance industry, the extraction and analysis of data reports are an indispensable part of daily work. The industry has accumulated a huge amount of data with diverse data types and complex data analysis requirements. The traditional data extraction method is for scientific and technological personnel to extract data based on demand descriptions. This method is not only time-consuming and labor-intensive, but also prone to errors.

[0003] With the rapid development of big language model technology, big language models have shown excellent performance in text analysis and processing. However, compared with the natural language text processed by big language models, the cell strings in tables are usually shorter in length and more concise in meaning, and in specific fields, such as the insurance industry, there are a large number of professional vocabulary. In addition, the cells in each table establish richer information through non-grammatical connections, such as text, data, location and format, which is different from the text logic of natural language. In summary, the big language model's understanding of tables and its ability to process tables are still limited. Summary of the invention

[0004] In order to solve the above technical problem or at least partially solve the above technical problem, the present invention provides an insurance report processing method, device and storage medium based on a table processing language model.

[0005] In a first aspect, the present invention provides an insurance report processing method based on a table processing language model, comprising:

[0006] According to the titles, formats and contents of various insurance reports, the information in various insurance reports is encoded into information encoding including semantic position encoding, table position encoding, embedded position encoding and information embedding;

[0007] The language model is trained by the information encoding to understand the information contained in the insurance report according to the information embedding, the semantic position encoding of the information embedding and the embedding position encoding;

[0008] Using the initial training insurance report, the request target, the table intermediate processing request to achieve the request target, the table processing operation instruction to achieve the table intermediate processing request, the intermediate training insurance report corresponding to the table intermediate processing request and the result training insurance report corresponding to the request target to build a question-answering decision data set, taking the request target in the question-answering decision data set as the question, taking the table intermediate processing request to achieve the request target and the operation instruction to achieve the table intermediate processing request as the answer to train the language model, giving the language model the ability to establish a corresponding relationship between the processing request expressed in natural language and the table processing operation instruction; and based on the question-answering decision data set, the language model is trained through reinforcement learning to decompose the request target into the decision-making ability of several intermediate processing requests;

[0009] When processing insurance reports: the request target expressed in natural language and the initial state of the insurance report to be processed are used as prompt words for initial request generation, and the control language model decides to obtain a table intermediate processing request and its corresponding table processing operation instructions for achieving the request target; the table processing software executes the obtained table processing operation instructions to obtain the next state of the insurance report to be processed as the current state; the request target expressed in natural language, the current state of the insurance report to be processed, the historical state of the insurance report to be processed and the table processing operation instructions executed historically are used as prompt words for current request generation, and the control language model decides to obtain the next table intermediate processing request and its corresponding table processing operation instructions for achieving the request target, and repeats until a result insurance report that meets the request target expressed in natural language is obtained.

[0010] Furthermore, the semantic position encoding of any target data in any insurance report is constructed as follows:

[0011] Obtain the report name of the insurance report, construct an insurance report name set, and use the position code of each report name in the report name set as the report name position code;

[0012] For the target insurance report where the target data is located, the target insurance report is decomposed into a plurality of row-based sub-tables and column-based sub-tables according to rows and columns, respectively. The row-based sub-table includes hierarchical row headers and table data corresponding to the rightmost level row headers, and the column-based sub-table includes hierarchical column headers and table data corresponding to the lowest level column headers;

[0013] For any target data, its semantic position encoding includes:

[0014] The column semantic position code is obtained based on the report name position code, the sorting position of the header of the column base subtable to which the target data belongs in each column base subtable, the hierarchical relationship of the header of the column base subtable to which the target data belongs, and the sorting position of the sub-level header in the header of the column base subtable to which the target data belongs;

[0015] The row semantic position coding is obtained based on the report name position coding, the sorting position of the header of the row basic sub-table to which the target data belongs in each row basic sub-table, the header hierarchy relationship of the row basic sub-table to which the target data belongs, and the sorting position of the sub-level header in the header of the row basic sub-table to which the target data belongs.

[0016] Furthermore, the column semantic position encoding is expressed as:

[0017] in, is a learnable column semantic position encoding matrix, Cat() represents the concatenation operation, is a vector representing the sorting position of the first-level header to which the target data belongs, a vector representing the sorting position of the second-level header to which the target data belongs, and so on, a vector representing the sorting position of the nth-level header to which the target data belongs. are a number of empty filling vectors with the same dimensions as the vectors representing the sorting positions of the table headers, and nmax is the maximum level of the table header of the column base subtable;

[0018] The row semantic position encoding is represented as:

[0019] in, is a learnable column semantic position encoding matrix, Cat() represents the concatenation operation, is a vector representing the sorting position of the first-level header to which the target data belongs, a vector representing the sorting position of the second-level header to which the target data belongs, and so on, a vector representing the sorting position of the m-th-level header to which the target data belongs. are several empty filling vectors with the same dimensions as the vector representing the sort position of the header, and mmax is the maximum level of the header of the row-based sub-table.

[0020] Furthermore, the table position code includes table column position code and table row position code, and the table column and row position codes are codes based on the column and row coordinates of the table cells;

[0021] The table column position encoding of the target data is expressed as: in, represents the learnable table column position encoding matrix, A vector representing the column coordinates of the table cell where the target data is located;

[0022] The table row position encoding is represented as: in, represents the learnable table row position encoding matrix, A vector representing the row coordinates of the table cell where the target data is located.

[0023] Furthermore, the information encoding is expressed as:

[0024] E t =CAT(P,P emb ,E).

[0025] Among them, the table column position is encoded Column Semantic Position Encoding The column-related position encoding is obtained by concatenation, namely: Combining the table row position encoding with the row semantic position encoding yields the row-related position encoding, namely: Then the column-related position code and the row-related position code are concatenated P = CAT (P l ,P h );

[0026] Embedded Position Encoding P emb Use sine and cosine position encoding;

[0027] Information embedding is obtained by encoding the column header information, row header information and target data in sequence through the Embedding model, expressed as:

[0028] E=Embedding( 1 I l , 2 I l ,…… n I l , 1 I h , 2 I h ,…… m I h , I tar ),in, 1 I l , 2 I l ,…… n I l For column header information, 1 I h , 2 I h ,…… m I h For row header information, I tar is the target data.

[0029] Furthermore, when the language model is trained to understand insurance reports through the information encoding, the first random mask is used to randomly cover the vocabulary of the information in the cell, and the language model is trained to predict the covered part of the content, so that the language modeling understands the contextual connection of the information in each cell; when the language model is trained to understand insurance reports, the second random mask is used to randomly cover any level of headers, and the language model is trained to predict the header information of the covered level, so as to learn the contextual connection between the header levels in the insurance report header; the cross entropy between the predicted meaning of the target data obtained by the language model based on the information encoding of the target information and the true meaning of the target data is calculated; and the above constraints are used to train the language model to understand the information of the insurance report.

[0030] Furthermore, the request generation prompt contains a control language model for a specific form processing software, which gradually obtains a form intermediate processing request to achieve the request target based on the request target expressed in natural language, and obtains the form processing operation instructions of the specific form processing software corresponding to the form intermediate processing request.

[0031] Furthermore, based on the question-answering decision dataset, the language model is trained to decompose the request target into several intermediate decision-making capabilities for processing the request, including:

[0032] The request target of the question-answering dataset and the current state of the initial training insurance report are used as prompt words for the initial request generation, and the language model is controlled to make a decision to obtain an expected form intermediate processing request and its corresponding expected form processing operation instruction to achieve the request target;

[0033] The form processing software executes the obtained expected form processing operation instruction to obtain the next state of the initial training insurance report, that is, the expected intermediate training insurance report as the current state;

[0034] The request target expressed in natural language, the current state of the training insurance report, the historical state of the training insurance report, and the expected form processing request and expected form processing operation instruction executed in history are used as prompt words for generating the current request, and the control language model is used to make a decision to obtain the next expected form intermediate processing request and its corresponding expected form processing operation instruction to achieve the request target;

[0035] Iterate until the expected results are obtained to train the insurance report;

[0036] The cross entropy between the expected table intermediate processing requests, expected table processing operation instructions, expected intermediate training insurance reports, expected result training insurance reports obtained by calculating the actual decision of the language model and the real table intermediate processing requests, table processing operation instructions, intermediate training insurance reports and result training insurance reports in the question and answer decision dataset is used as the loss function, and the language model parameters are trained to minimize the loss function.

[0037] In a second aspect, the present invention provides an insurance report processing device based on a table processing language model, comprising: at least one processing unit, the processing unit is connected to a storage unit and a collection unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the insurance report processing method based on the table processing language model is implemented.

[0038] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the insurance report processing method based on the table processing language model as described above.

[0039] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art:

[0040] This application encodes the information in various insurance reports according to the titles, formats and contents of various insurance reports into information codes including: semantic position coding, table position coding, embedded position coding and information embedding; the semantic position coding and table position coding constitute the position information of the information embedded in the global insurance report, and the combination of embedded position coding and information embedding enables the language model to better understand the information of the insurance report.

[0041] The question-answering decision data set is constructed using the initial training insurance report, the request target, the table intermediate processing request to achieve the request target, the table processing operation instruction to achieve the table intermediate processing request, the intermediate training insurance report corresponding to the table intermediate processing request and the result training insurance report corresponding to the request target. The request target in the question-answering decision data set is used as the question, and the table intermediate processing request to achieve the request target and the operation instruction to achieve the table intermediate processing request are used as the answer to train the language model, and the language model is given the ability to establish a corresponding connection between the processing request expressed in natural language and the table processing operation instruction; and based on the question-answering decision data set, the language model is trained to decompose the request target into a number of intermediate processing requests. The language model is enabled to generate a table intermediate processing request and its corresponding table processing operation instruction based on the request target provided by the user under the control of the prompt word, and the table processing software gradually executes the table processing operation instruction to obtain a result insurance report that satisfies the request target expressed in natural language. Through the semantic understanding ability of the language model, the natural language request target that fits the spoken language is connected with a number of table processing operation instructions obtained by the language model decision, and non-professionals who are not proficient in table processing operations are supported to perform table processing, reduce the demand for table operations, and improve processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0044] Figure 1 A flowchart of an insurance report processing method based on a table processing language model provided by an embodiment of the present invention;

[0045] Figure 2 A flowchart of a method for constructing semantic position coding of any target data in any insurance report provided by an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the insurance amount statistics table;

[0047] Figure 4 A flowchart of the decision-making capability of training a language model to decompose a request target into several intermediate processing requests based on a question-answer decision data set provided by an embodiment of the present invention;

[0048] Figure 5 A schematic diagram of an insurance report processing device based on a table processing language model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0051] Example 1

[0052] like Figure 1 As shown, the technology of the present invention implements an insurance report processing method based on a table processing language model, including:

[0053] According to the titles, formats and contents of various insurance reports, the information in various insurance reports is encoded into information encoding including semantic position encoding, table position encoding, embedded position encoding and information embedding encoding.

[0054] Among them, the semantic position coding and the table position coding constitute the position information of the information embedded in the global insurance report, and the embedded position coding is the position coding of the vocabulary contained in the information itself.

[0055] like Figure 2 As shown in the figure, the semantic position encoding of any target data in any insurance report is constructed as follows:

[0056] The report names of all insurance reports are obtained in advance, an insurance report name set is constructed, and the position of each report name in the report name set is encoded, such as one-hot encoding, as the report name position code.

[0057] For the target insurance report where the target data is located, the target insurance report is decomposed into several row-based sub-tables and column-based sub-tables according to rows and columns respectively. The row-based sub-table contains hierarchical row headers and tabular data corresponding to the rightmost level row headers, and the column-based sub-table contains hierarchical column headers and tabular data corresponding to the lowest level column headers.

[0058] like Figure 3 In the insurance amount statistics table shown, the column basic sub-tables include: insurance amount statistics column basic sub-table, premium column basic sub-table, effective policy column basic sub-table, compensation expenditure column basic sub-table, surrender amount column basic sub-table, annuity payment column basic sub-table, maturity payment column basic sub-table and death and injury medical payment column basic sub-table; the row basic sub-tables include: insurance amount statistics row basic sub-table and insurance type row basic sub-table.

[0059] Taking the premium column basic sub-table as an example, the first-level header is "Premium", and the first-level header "Premium" contains two second-level headers: "Current Premium" and "Cumulative Premium". Taking the insurance type row basic sub-table as an example, the first-level header is "Insurance Type", and the first-level header "Insurance Type" contains four second-level headers: "Life Insurance", "Accidental Injury Insurance", "Health Insurance" and "Auto Insurance".

[0060] For any target data, its semantic position encoding includes: column semantic position encoding and row semantic position encoding.

[0061] The column semantic position encoding is obtained based on the report name position encoding, the header sorting position of the column base subtable to which the target data belongs in each column base subtable, the header hierarchical relationship of the column base subtable to which the target data belongs, and the sorting position of the sub-level header in the header of the column base subtable to which the target data belongs:

[0062] Report name position code + first-level header sort position code of the column-based sub-table header to which the target data belongs + second-level header sort position code to which the target data belongs...

[0063] The column semantic position encoding is represented as:

[0064] in, is a learnable column semantic position encoding matrix, Cat() represents the concatenation operation, is a vector representing the sorting position of the first-level header to which the target data belongs, a vector representing the sorting position of the second-level header to which the target data belongs, and so on, a vector representing the sorting position of the nth-level header to which the target data belongs. are a number of empty filling vectors that are consistent with the dimensions of the vector representing the sort position of the header, an example of an empty filling vector is all 0;

[0065] The number of header levels of the column-based subtable corresponding to different target data may be different, and it is necessary to ensure that the total dimension of the vector obtained by the Cat() operation is consistent with the learnable column semantic position encoding matrix Matching, since the number of levels of the column-based subtable header has a maximum upper limit, let the learnable column semantic position encoding matrix match the maximum dimension, that is, the dimension of the concatenation result when the number of levels is the maximum upper limit nmax. When the level of the header of the column-based subtable to which the target data belongs is lower than the maximum level nmax of the header of the column-based subtable, the filling vector is controlled to fill in to meet the dimension matching requirements.

[0066] The row semantic position encoding is obtained based on the report name position encoding, the sorting position of the header of the row basic sub-table to which the target data belongs in each row basic sub-table, the hierarchical relationship of the header of the row basic sub-table to which the target data belongs, and the sorting position of the sub-level header in the header of the row basic sub-table to which the target data belongs:

[0067] Report name position code + first-level header sort position code of the row-based sub-table header to which the target data belongs + second-level header sort position code to which the target data belongs...

[0068] The row semantic position encoding is represented as:

[0069] in, is a learnable column semantic position encoding matrix, Cat() represents the concatenation operation, is a vector representing the sorting position of the first-level header to which the target data belongs, a vector representing the sorting position of the second-level header to which the target data belongs, and so on, a vector representing the sorting position of the m-th-level header to which the target data belongs. are a number of empty filling vectors that are consistent with the dimensions of the vector representing the sort position of the header, an example of an empty filling vector is all 0;

[0070] The number of levels of the header of the row-based sub-table corresponding to different target data may be different, and it is necessary to ensure that the total dimension of the vector obtained by the Cat() operation is consistent with the row semantic position encoding matrix Matching, since the number of levels of the header of the row-based subtable has a maximum upper limit, let the learnable column semantic position encoding matrix match the maximum dimension, that is, the dimension of the concatenation result when the number of levels is the maximum upper limit mmax. When the level of the header of the row-based subtable to which the target data belongs is lower than the maximum level mmax of the header of the row-based subtable, the filling vector is controlled to fill in to meet the dimension matching requirements.

[0071] The table position code includes table column position code and table row position code. The table column and row position codes are codes based on the column and row coordinates of the table cells.

[0072] Specifically, the table column position encoding of the target data is expressed as: in, represents the learnable table column position encoding matrix, A vector representing the column coordinates of the table cell where the target data is located.

[0073] The table row position encoding is represented as: in, represents the learnable table row position encoding matrix, A vector representing the row coordinates of the table cell where the target data is located.

[0074] Encode table column positions Column Semantic Position Encoding The column-related position encoding is obtained by concatenation, namely: Combining the table row position encoding with the row semantic position encoding yields the row-related position encoding, namely: Then the column-related position code and the row-related position code are concatenated P = CAT (P l ,P h ). In this way, the semantic position coding and the table position coding constitute the position information embedded in the global insurance report.

[0075] Embedded Position Encoding P emb Sin-cosine position encoding is used.

[0076] Information embedding is obtained by encoding the column header information, row header information and target data in sequence through the Embedding model, expressed as:

[0077] E=Embedding( 1 I l , 2 I l ,…… n I l , 1 I h , 2 I h ,…… m I h , I tar ),in, 1 I l , 2 I l ,…… n I l For column header information, 1 I h , 2 I h ,…… m I h For row header information, I tar is the target data.

[0078] The overall information encoding is expressed as:

[0079] E t =CAT(P,P emb ,E).

[0080] The information encoding is used to train a language model to understand the information contained in the insurance report based on information embedding, semantic position encoding of information embedding and embedding position encoding.

[0081] During the specific implementation process, when the language model is trained to understand the insurance report through the information encoding, the first random mask is used to randomly cover the vocabulary of the information in the cell, and the language model is trained to predict the covered part of the content, so that the language modeling can understand the contextual connection of the information in each cell; when the language model is trained to understand the insurance report, the second random mask is used to randomly cover the header of any level of the header, and the language model is trained to predict the header information of the covered level, so as to learn the contextual connection between the header levels in the insurance report header; the cross entropy between the predicted meaning of the target data obtained by the language model based on the information encoding of the target information and the true meaning of the target data is calculated; the language model is trained using the above constraints so that the language model can better understand the information of the insurance report.

[0082] A question-and-answer decision dataset is constructed using the initial training insurance report, request target, table intermediate processing request to achieve the request target, table processing operation instruction to achieve the table intermediate processing request, intermediate training insurance report corresponding to the table intermediate processing request and result training insurance report corresponding to the request target; the request target in the question-and-answer decision dataset is used as the question, and the table intermediate processing request to achieve the request target and the operation instruction to achieve the table intermediate processing request are used as the answer to train the language model, giving the language model the ability to establish a corresponding relationship between the processing request expressed in natural language and the table processing operation instruction; and based on the question-and-answer decision dataset, the language model is trained to decompose the request target into the decision-making ability of several intermediate processing requests.

[0083] like Figure 4 As shown in the figure, based on the question-answering decision dataset, the decision capabilities of training the language model to decompose the request target into several intermediate processing requests include:

[0084] The request target of the question-answering dataset and the current state of the initial training insurance report are used as the initial request generation prompt words, and the control language model decides to obtain an expected form intermediate processing request to achieve the request target and its corresponding expected form processing operation instruction. Among them, the request generation prompt words contain the control language model for the specific form processing software, based on the request target expressed in natural language, gradually obtains the form intermediate processing request to achieve the request target, and obtains the form processing operation instruction of the specific form processing software corresponding to the form intermediate processing request.

[0085] The form processing software executes the obtained expected form processing operation instruction to obtain the next state of the initial training insurance report, that is, the expected intermediate training insurance report as the current state.

[0086] The request target expressed in natural language, the current state of the training insurance report, the historical state of the training insurance report, and the expected form processing request and expected form processing operation instruction executed in history are used as prompt words for generating the current request, and the control language model is used to make a decision to obtain the next expected form intermediate processing request and its corresponding expected form processing operation instruction to achieve the request target;

[0087] Iterate until the expected results are obtained to train the insurance report;

[0088] The cross entropy between the expected table intermediate processing requests, expected table processing operation instructions, expected intermediate training insurance reports, expected result training insurance reports obtained by calculating the actual decision of the language model and the real table intermediate processing requests, table processing operation instructions, intermediate training insurance reports and result training insurance reports in the question and answer decision dataset is used as the loss function, and the language model parameters are trained to minimize the loss function.

[0089] When using the language model to process insurance reports:

[0090] The request target expressed in natural language and the initial state of the insurance report to be processed are used as prompt words for initial request generation, and the control language model decides to obtain a table intermediate processing request and its corresponding table processing operation instruction to achieve the request target; the table processing software executes the obtained table processing operation instruction to obtain the next state of the insurance report to be processed as the current state; the request target expressed in natural language, the current state of the insurance report to be processed, the historical state of the insurance report to be processed and the table processing operation instruction executed historically are used as prompt words for current request generation, and the control language model decides to obtain the next table intermediate processing request and its corresponding table processing operation instruction to achieve the request target, and repeats until a result insurance report that meets the request target expressed in natural language is obtained.

[0091] Example 2

[0092] See also Figure 5 As shown, an embodiment of the present invention provides an insurance report processing device based on a table processing language model, comprising: at least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit is a computer-readable storage medium, and can be used to store software programs, computer executable programs, and modules, such as the software programs, computer executable programs, and modules corresponding to an insurance report processing method based on a table processing language model in an embodiment of the present invention. The processing unit implements the above-mentioned insurance report processing method based on a table processing language model by running the software programs, computer executable programs, and modules stored in the storage unit, comprising:

[0093] According to the titles, formats and contents of various insurance reports, the information in various insurance reports is encoded into information encoding including semantic position encoding, table position encoding, embedded position encoding and information embedding;

[0094] The language model is trained by the information encoding to understand the information contained in the insurance report according to the information embedding, the semantic position encoding of the information embedding and the embedding position encoding;

[0095] A question-answering decision data set is constructed using the initial training insurance report, the request target, the table intermediate processing request to achieve the request target, the table processing operation instruction to achieve the table intermediate processing request, the intermediate training insurance report corresponding to the table intermediate processing request and the result training insurance report corresponding to the request target; the request target in the question-answering decision data set is used as the question, and the table intermediate processing request to achieve the request target and the operation instruction to achieve the table intermediate processing request are used as the answer to train the language model, giving the language model the ability to establish a corresponding relationship between the processing request expressed in natural language and the table processing operation instruction; and based on the question-answering decision data set, the language model is trained to decompose the request target into the decision-making ability of several intermediate processing requests;

[0096] The request target expressed in natural language and the initial state of the insurance report to be processed are used as prompt words for initial request generation, and the control language model decides to obtain a table intermediate processing request and its corresponding table processing operation instruction to achieve the request target; the table processing software executes the obtained table processing operation instruction to obtain the next state of the insurance report to be processed as the current state; the request target expressed in natural language, the current state of the insurance report to be processed, the historical state of the insurance report to be processed and the table processing operation instruction executed historically are used as prompt words for current request generation, and the control language model decides to obtain the next table intermediate processing request and its corresponding table processing operation instruction to achieve the request target, and repeats until a result insurance report that meets the request target expressed in natural language is obtained.

[0097] Of course, the computer program stored in the storage unit of the device for implementing the insurance report processing method based on the table processing language model provided in the embodiment of the present invention is not limited to the method operations described above, and can also execute the relevant operations in the insurance report processing method based on the table processing language model provided in any embodiment of the present invention.

[0098] Example 3

[0099] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed, the insurance report processing method based on the table processing language model is implemented, including:

[0100] According to the titles, formats and contents of various insurance reports, the information in various insurance reports is encoded into information encoding including semantic position encoding, table position encoding, embedded position encoding and information embedding;

[0101] The language model is trained by the information encoding to understand the information contained in the insurance report according to the information embedding, the semantic position encoding of the information embedding and the embedding position encoding;

[0102] A question-answering decision data set is constructed using the initial training insurance report, the request target, the table intermediate processing request to achieve the request target, the table processing operation instruction to achieve the table intermediate processing request, the intermediate training insurance report corresponding to the table intermediate processing request and the result training insurance report corresponding to the request target; the request target in the question-answering decision data set is used as the question, and the table intermediate processing request to achieve the request target and the operation instruction to achieve the table intermediate processing request are used as the answer to train the language model, giving the language model the ability to establish a corresponding relationship between the processing request expressed in natural language and the table processing operation instruction; and based on the question-answering decision data set, the language model is trained to decompose the request target into the decision-making ability of several intermediate processing requests;

[0103] The request target expressed in natural language and the initial state of the insurance report to be processed are used as prompt words for initial request generation, and the control language model decides to obtain a table intermediate processing request and its corresponding table processing operation instruction to achieve the request target; the table processing software executes the obtained table processing operation instruction to obtain the next state of the insurance report to be processed as the current state; the request target expressed in natural language, the current state of the insurance report to be processed, the historical state of the insurance report to be processed and the table processing operation instruction executed historically are used as prompt words for current request generation, and the control language model decides to obtain the next table intermediate processing request and its corresponding table processing operation instruction to achieve the request target, and repeats until a result insurance report that meets the request target expressed in natural language is obtained.

[0104] A computer-readable storage medium provided in an embodiment of the present invention stores a computer program which is not limited to the method operations described above, but can also execute related operations in an insurance report processing method based on a table processing language model provided in any embodiment of the present invention.

[0105] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, structures or units, which can be electrical, mechanical or other forms.

[0106] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0107] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0108] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for processing insurance reports based on a table processing language model, characterized in that: include: According to the titles, formats and contents of various insurance reports, the information in various insurance reports is encoded into information encoding including semantic position encoding, table position encoding, embedded position encoding and information embedding; Among them, the construction method of the semantic position coding of any target data in any insurance report is as follows: obtain the report name of the insurance report, construct an insurance report name set, and use the position coding of each report name in the report name set as the report name position coding; for the target insurance report where the target data is located, decompose the target insurance report into a plurality of row-based sub-tables and column-based sub-tables according to rows and columns, respectively, the row-based sub-table contains a hierarchical row header and table data corresponding to the rightmost level row header, and the column-based sub-table contains a hierarchical column header and table data corresponding to the lowest level column header; for any target data, Its semantic position coding includes: obtaining a column semantic position coding based on the report name position coding, the sorting position of the header of the column base subtable to which the target data belongs in each column base subtable, the hierarchical relationship of the header of the column base subtable to which the target data belongs, and the sorting position of the sub-level header in the header of the column base subtable to which the target data belongs; obtaining a row semantic position coding based on the report name position coding, the sorting position of the header of the row base subtable to which the target data belongs in each row base subtable, the hierarchical relationship of the header of the row base subtable to which the target data belongs, and the sorting position of the sub-level header in the header of the row base subtable to which the target data belongs; The table position code includes table column position code and table row position code. The table column and row position codes are codes based on the column and row coordinates of the table cells. The table column position code of the target data is expressed as: ,in, represents the learnable table column position encoding matrix, is the characterization vector of the column coordinates of the table cell where the target data is located; the table row position encoding is expressed as: ,in, represents the learnable table row position encoding matrix, is the representation vector of the row coordinates of the table cell where the target data is located; Using the information encoding to train a language model, the language model can understand the information contained in the insurance report based on the information embedding, the semantic position encoding of the information embedding, and the embedding position encoding; A question-answering decision data set is constructed using the initial training insurance report, the request target, the table intermediate processing request that realizes the request target, the table processing operation instruction that realizes the table intermediate processing request, the intermediate training insurance report corresponding to the table intermediate processing request and the result training insurance report corresponding to the request target; the request target in the question-answering decision data set is used as the question, and the table intermediate processing request that realizes the request target and the operation instruction that realizes the table intermediate processing request are used as the answer to train the language model, giving the language model the ability to establish a corresponding relationship between the processing request expressed in natural language and the table processing operation instruction; and based on the question-answering decision data set, the language model is trained to decompose the request target into the decision-making ability of several intermediate processing requests; When processing insurance reports: the request target expressed in natural language and the initial state of the insurance report to be processed are used as prompt words for initial request generation, and the control language model decides to obtain a table intermediate processing request and its corresponding table processing operation instructions for achieving the request target; the table processing software executes the obtained table processing operation instructions to obtain the next state of the insurance report to be processed as the current state; the request target expressed in natural language, the current state of the insurance report to be processed, the historical state of the insurance report to be processed and the table processing operation instructions executed historically are used as prompt words for current request generation, and the control language model decides to obtain the next table intermediate processing request and its corresponding table processing operation instructions for achieving the request target, and repeats until a result insurance report that meets the request target expressed in natural language is obtained.

2. The insurance report processing method based on the table processing language model according to claim 1 is characterized in that: The column semantic position encoding is expressed as: ,in, is the learnable column semantic position encoding matrix, Represents a splicing operation, is a vector representing the sorting position of the first-level header to which the target data belongs, a vector representing the sorting position of the second-level header to which the target data belongs, and so on, a vector representing the sorting position of the nth-level header to which the target data belongs. are several empty filling vectors with the same dimensions as the vector representing the sort position of the header, The maximum level of the header of the column base subtable; The row semantic position encoding is represented as: ,in, is the learnable column semantic position encoding matrix, Represents a splicing operation, is a vector representing the sorting position of the first-level header to which the target data belongs, a vector representing the sorting position of the second-level header to which the target data belongs, and so on, a vector representing the sorting position of the m-th-level header to which the target data belongs. are several empty filling vectors with the same dimensions as the vector representing the sort position of the header, The maximum level of the header of the row-based subtable.

3. The insurance report processing method based on the table processing language model according to claim 1 is characterized in that: The information encoding is expressed as: ； Among them, the table column position is encoded Column Semantic Position Encoding The column-related position encoding is obtained by concatenation, namely: ; Combine the table row position encoding with the row semantic position encoding to get the row-related position encoding, that is: ; Then concatenate the column-related position code and the row-related position code ; Embedded Positional Encoding Use sine and cosine position encoding; Information embedding is obtained by encoding the column header information, row header information and target data in sequence through the Embedding model, expressed as: ,in, For column header information, For row header information, is the target data.

4. The insurance report processing method based on the table processing language model according to claim 1 is characterized in that: When the information encoding is used to train the language model to understand the insurance report, the first random mask is used to randomly cover the vocabulary of the information in the cell, and the language model is trained to predict the covered part of the content, so that the language modeling understands the contextual connection of the information in each cell; When training the language model to understand the insurance report, the second random mask is used to randomly cover any level of the header, and the language model is trained to predict the header information of the covered level, thereby learning the contextual connection between the header levels in the insurance report header; The cross entropy between the predicted meaning of the target data obtained by the language model based on the information encoding of the target information and the true meaning of the target data is calculated; and the above constraints are used to train the language model to understand the information of the insurance report.

5. The insurance report processing method based on the table processing language model according to claim 1 is characterized in that: The request generation prompt contains a control language model for a specific form processing software, which gradually obtains a form intermediate processing request to achieve the request target based on the request target expressed in natural language, and obtains the form processing operation instructions of the specific form processing software corresponding to the form intermediate processing request.

6. The insurance report processing method based on the table processing language model according to claim 1 is characterized in that: Based on the question-answering decision dataset, the language model is trained to decompose the request target into several intermediate decision-making capabilities for processing requests, including: The request target of the question-answering dataset and the current state of the initial training insurance report are used as prompt words for the initial request generation, and the language model is controlled to make a decision to obtain an expected form intermediate processing request and its corresponding expected form processing operation instruction to achieve the request target; The form processing software executes the obtained expected form processing operation instruction to obtain the next state of the initial training insurance report, that is, the expected intermediate training insurance report as the current state; The request target expressed in natural language, the current state of the training insurance report, the historical state of the training insurance report, and the expected form processing request and expected form processing operation instruction executed in history are used as prompt words for generating the current request, and the control language model is used to make a decision to obtain the next expected form intermediate processing request and its corresponding expected form processing operation instruction to achieve the request target; Iterate until the expected results are obtained to train the insurance report; The cross entropy between the expected table intermediate processing requests, expected table processing operation instructions, expected intermediate training insurance reports, expected result training insurance reports obtained by calculating the actual decision of the language model and the real table intermediate processing requests, table processing operation instructions, intermediate training insurance reports and result training insurance reports in the question and answer decision dataset is used as the loss function, and the language model parameters are trained to minimize the loss function.

7. An insurance report processing device based on a table processing language model, characterized in that: include: At least one processing unit, the processing unit is connected to a storage unit via a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the insurance report processing method based on the table processing language model as described in any one of claims 1-6 is implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the insurance report processing method based on the table processing language model as described in any one of claims 1 to 6 is implemented.

Images