Index data extraction method and program product

Abstract

According to the index data extraction method and the program product, the subject-index mapping associated with the business knowledge is constructed, meanwhile, the original text is sliced to obtain the target text block, the target text block and the subject-index mapping are dynamically matched, on the basis of completing matching, the related indexes are accurately extracted, and the extraction efficiency of the index data is improved. The interference of the unstructured original text on extraction is avoided, and the extraction efficiency and accuracy of the index data are improved.

Description

Technical Field

[0001] The present application relates to the field of large model applications, and more specifically, to an index data extraction method and a program product. Background Art

[0002] With the development of various industries, the number of reports generated by enterprises and institutions has increased dramatically. The contents of these announcements cover a wide range of business areas, with diverse forms and expressions. Much of the information is presented in the form of unstructured natural language text, which increases the difficulty of analysis and processing.

[0003] At present, in order to improve the efficiency of analysis and processing, it is often necessary to extract information from these reports, such as extracting specific indicators from the reports. In the task of indicator extraction, the existing technology mainly relies on traditional structured segmentation technology and static extraction methods. These methods lack flexibility and adaptability when dealing with complex semantic structures, and are difficult to cope with the diversity and complexity of report content. Summary of the invention

[0004] The purpose of this application is to provide a method and program product for extracting indicator data, which are used to avoid interference in the extraction due to unstructured original text and improve the extraction efficiency and accuracy of indicator data.

[0005] In order to achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:

[0006] In a first aspect, an embodiment of the present application provides a method for extracting indicator data, including:

[0007] Matching the target text block with all the topic tags to obtain a matching score between the target text block and each topic tag; the target text block is obtained by slicing the original text; the topic tags and the indicator group have a hierarchical correspondence; the indicator group includes various indicators corresponding to the business knowledge;

[0008] When the matching score satisfies the matching condition, a target topic tag corresponding to the target text block is determined; the target topic tag represents a topic tag among the topic tags that corresponds to the original text;

[0009] Determine a corresponding target indicator group according to the target theme; the target indicator group represents the indicator group corresponding to the original text in the indicator group;

[0010] The target indicator group and the target text block are input into a large model, the original text is extracted, and indicator information corresponding to the target indicator group is obtained.

[0011] Optionally, before the step of matching the target text block with all the topic tags to obtain the matching score between the target text block and each topic tag, the step further includes:

[0012] Parsing text data in the original text;

[0013] The text data is segmented to obtain the target text block.

[0014] Optionally, also include:

[0015] Determining hierarchical sets of topic labels and indicators based on the business knowledge;

[0016] Establishing a mapping relationship between the topic label of each layer and the corresponding indicator set respectively;

[0017] Generate a corresponding embedding vector for the topic label to form a topic space corresponding to the topic label.

[0018] Optionally, the step of matching the target text block with all the topic tags to obtain a matching score between the target text block and each topic tag includes:

[0019] Calculating first similarities between the target text block and each of the topic tags in the topic space respectively;

[0020] Calculating a second similarity between the target text block and each of the topic tags in the topic space based on a Gestalt pattern;

[0021] A matching score of each of the topic tags is determined according to each of the first similarities and the corresponding second similarities.

[0022] Optionally, when the matching score satisfies the matching condition, the step of determining the target topic tag corresponding to the target text block includes:

[0023] Comparing each of the matching scores with a matching threshold respectively;

[0024] If the matching score is greater than or equal to the matching threshold, the corresponding topic tag is determined as the target topic tag corresponding to the target text block.

[0025] In a second aspect, an embodiment of the present application provides a device for extracting labeled data, including:

[0026] A matching module is used to match a target text block with all topic tags to obtain a matching score between the target text block and each topic tag; the target text block is obtained by slicing the original text; the topic tags and the indicator groups have a hierarchical correspondence; the indicator group contains various indicators corresponding to business knowledge; when the matching score meets the matching condition, the target topic tag corresponding to the target text block is determined; the target topic tag represents the topic tag corresponding to the original text among the topic tags; the corresponding target indicator group is determined according to the target topic; the target indicator group represents the indicator group corresponding to the original text among the indicator group;

[0027] The extraction module is used to input the target indicator group and the target text block into the large model, extract the original text, and obtain indicator information corresponding to the target indicator group.

[0028] Optionally, also include:

[0029] A mapping module is used to determine hierarchical topic labels and indicator sets based on the business knowledge; establish a mapping relationship between the topic labels of each layer and the corresponding indicator sets; generate corresponding embedding vectors for the topic labels to form a topic space corresponding to the topic labels.

[0030] In a third aspect, an embodiment of the present application provides an electronic device, including:

[0031] A memory for storing one or more programs;

[0032] processor;

[0033] When the one or more programs are executed by the processor, the method as described in any one of the first aspects above is implemented.

[0034] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method as described in any one of the above-mentioned first aspects.

[0035] In a fifth aspect, an embodiment of the present application provides a program product, which, when executed by a processor, implements a method as described in any one of the above-mentioned first aspects.

[0036] Compared with the prior art, the present invention provides a

[0037] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 A flowchart of a method for extracting indicator data provided by an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of a flow chart of another method for extracting indicator data provided by an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of a flow chart of another method for extracting indicator data provided by an embodiment of the present invention;

[0042] Figure 4 A schematic diagram of a flow chart of another method for extracting indicator data provided by an embodiment of the present invention;

[0043] Figure 5 A schematic diagram of an index data extraction architecture provided by an embodiment of the present invention;

[0044] Figure 6 A schematic diagram of an index data extraction device provided by an embodiment of the present invention;

[0045] Figure 7 A schematic diagram of another device for extracting indicator data provided by an embodiment of the present invention;

[0046] Figure 8 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0048] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0049] It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and cannot be understood as indicating or implying relative importance.

[0050] It should be noted that, in this article, relational terms such as first and second, etc. 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 these entities or operations. Moreover, the terms "include", "comprise" or any other variants 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, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0051] With the development of digital technology, more and more enterprises and organizations need to display data related to their operations and management. For example, in the financial industry, the number of financial reports generated by enterprises and financial institutions has increased sharply. These announcements cover a wide range of business areas and involve complex information of various different businesses. At the same time, the forms and expressions of financial announcements and reports are diversified, and much information is presented in the form of unstructured natural language text, which increases the difficulty of analysis and processing.

[0052] At the same time, with the development of artificial intelligence and big model technology, it is often hoped that key data related to big model interaction and information processing can be obtained from the above announcements and reports through information extraction technology, such as indicator data for specific business areas.

[0053] However, existing technologies mainly rely on traditional structured segmentation techniques and static extraction methods based on rules or large models in indicator extraction tasks, such as:

[0054] Paragraph division methods based on traditional text segmentation: paragraph segmentation is performed based on punctuation marks, sentences, or structural information such as titles and chapters. These methods lack flexibility in dealing with complex semantic structures.

[0055] Static extraction method based on rules and keywords: extract indicators through predefined rules or keyword-based matching, usually extracting indicators according to a fixed pattern. This type of method has poor adaptability and is difficult to cope with the diversity and complexity of report content.

[0056] Static extraction methods based on Retrieval-Augmented Generation (RAG) technology and large models: This type of method usually uses a pre-trained large language model to analyze text, and uses two schemes, one-by-one extraction and batch extraction, to extract indicators according to the pattern defined during training. One-by-one extraction is suitable for accurate extraction of a single indicator, while batch extraction can improve extraction efficiency.

[0057] However, the above-mentioned omission is only applicable to unstructured text data in reports and announcements. Due to the natural language characteristics of such text data, it is often impossible to accurately and efficiently extract the indicator data therein.

[0058] Therefore, an embodiment of the present application provides an indicator data extraction solution, the core idea of ​​which is to construct a mapping of topics and indicators associated with business knowledge, slice the original text to obtain a target text block, dynamically match the target text block with the "topic-indicator mapping", and accurately extract relevant indicators on the basis of completing the matching, avoiding interference in the extraction due to unstructured original text, and improving the extraction efficiency and accuracy of indicator data.

[0059] Optionally, a possible implementation of an indicator data extraction method is provided below. Specifically, Figure 1 A flow chart of a method for extracting index data provided by an embodiment of the present invention is shown in FIG. Figure 1 , the method comprising:

[0060] Step 105: Match the target text block with all the topic tags to obtain a matching score between the target text block and each topic tag.

[0061] The target text block is obtained by slicing the original text; the topic labels and indicator groups have a hierarchical correspondence.

[0062] Optionally, the topic labels and indicator groups have a hierarchical correspondence, which can be understood as follows: for a certain business knowledge, its topic can be divided into a multi-level topic relationship of topic-subtopic according to the hierarchy, and then, for each level of topic, there is a corresponding indicator group. Each indicator group contains various indicators corresponding to the business knowledge.

[0063] Step 106: When the matching score meets the matching condition, the target topic tag corresponding to the target text block is determined.

[0064] The target topic tag represents the topic tag corresponding to the original text in the topic tags.

[0065] Step 107: Determine the corresponding target indicator group according to the target theme.

[0066] The target indicator group represents the indicator group in the indicator group that corresponds to the original text.

[0067] Step 108: Input the target indicator group and the target text block into the large model, extract the original text, and obtain indicator information corresponding to the target indicator group.

[0068] An embodiment of the present invention provides a method for extracting indicator data, which constructs a mapping of topics and indicators associated with business knowledge, slices the original text to obtain a target text block, and dynamically matches the target text block with the "mapping of topics and indicators". After completing the matching, relevant indicators are accurately extracted to avoid interference in the extraction due to unstructured original text, thereby improving the extraction efficiency and accuracy of indicator data.

[0069] Before executing the above-mentioned indicator information capture solution, it is necessary to first obtain the target text block from the original text. The following is a possible implementation method. Specifically, Figure 1 On the basis of Figure 2 A flow chart of another method for extracting index data provided by an embodiment of the present invention is shown in FIG. Figure 2 , the method further comprising:

[0070] Step 100: parse text data in the original text.

[0071] Step 101: segment the text data to obtain target text blocks.

[0072] It should be noted that, for the example of this application, what is of concern is the "text data" in the original text. That is, the unstructured text data described above. Therefore, in step 100, the core idea is to parse the original text, such as a financial report file, and remove the noise data (structured data such as icons, headers, and footers). Abnormal characters or symbols in the format can also be cleaned up. Special marks or tags (such as paragraph numbers, reference numbers, etc.) in the original text are uniformly processed to obtain text data.

[0073] In addition, there may be a variety of segmentation mechanisms for the segmentation process in step 101, and several possible implementations are exemplarily described below.

[0074] Segmentation method 1: Structural information segmentation.

[0075] Use the structural information of text data (such as titles, sections, numbers, etc.) to segment. Identify structural markers such as titles, sections, and numbers in the text data and use them as segmentation points. The beginning of each title or section is used as the starting position of a new paragraph to ensure that each paragraph forms an independent unit according to the logical structure of the text data. δ(i) indicates whether to insert a segmentation point at position i:

[0076]

[0077] Segmentation method 2: semantic recognition segmentation.

[0078] Paragraph segmentation is performed based on the language model (LM) and pointwise mutual information (PMI). First, a language model (such as BERT or GPT) is used to evaluate the fluency between paragraphs A and B. The joint probability P(λ, B) of the two paragraphs A and B is calculated.

[0079] After that, the PMI score is calculated to measure the co-occurrence between paragraphs. If the PMI value is low, it may mean that the two paragraphs belong to different semantic units in terms of semantics and are suitable for segmentation:

[0080]

[0081] Weighted calculation of semantic segmentation score:

[0082] Score(A,B)=α×P(A,B)+β×PMI(A,B)

[0083] Where α represents the weight of the language model score (fluency) and β represents the weight of the PMI score (relevance).

[0084] Finally, based on the semantic segmentation score and threshold γ, decide whether to insert a segmentation point:

[0085]

[0086] Splitting method three: sliding window splitting.

[0087] Based on the segmentation of structural information, a sliding window technique is used. A fixed-length window is set, and the text data is traversed in order. The text data is segmented into paragraphs within the window size range to ensure that each paragraph is of moderate length. The window can be set to overlap according to actual needs to maintain a certain contextual continuity between adjacent paragraphs and avoid semantic fragmentation. Let the window length be W and the overlap ratio be r, and the starting position x of the kth window be k for:

[0088]

[0089] when:

[0090] δ(x k +(1-r)×W)=0

[0091] The end position y of the kth window k for:

[0092] y k =x k +(1-r)×W

[0093] Otherwise, you need to subtract the step size i and find the previous split point that meets the following conditions:

[0094] δ(x k +(1-r)×Wi)=0

[0095] At this time, the end position y of the kth window k for:

[0096] y k =x k +(1-r)×Wi

[0097] Specifically, during the segmentation process of the sliding window, when the window slides to a suitable segmentation point, it can decide whether to segment based on the semantic information in the window. For paragraphs without obvious structural markers (such as titles, subtitles) and containing complex business information, the sliding window ensures the semantic integrity of each paragraph by gradually traversing the text, and can effectively avoid problems caused by mis-segmentation (such as when cutting the key part of a sentence or paragraph).

[0098] The above three segmentation methods can be combined based on specific business requirements and characteristics of text data to achieve better segmentation effects. The implementation process of the above segmentation methods is described in detail below with reference to specific examples.

[0099] Assume that this application is processing a financial report that contains the company's annual financial data and descriptions of ESG (environmental, social and governance) indicators. The structure and content of the report are unstructured, information-intensive and complex, and the content is mostly described in natural language.

[0100] One implementation method is to combine semantic recognition segmentation with sliding window segmentation.

[0101] It is applicable in the following scenarios: When the text in the report does not have obvious structural information, but the semantics of the content is relatively complex and has multiple sub-topics, semantic recognition segmentation is first used to identify paragraph boundaries, and then the sliding window technology is used to ensure that each paragraph is moderate in length and context.

[0102] The operation steps are:

[0103] First, use semantic recognition segmentation: analyze the fluency and semantic relationships of the text based on the language model and identify the boundaries of natural paragraphs.

[0104] Then, use sliding window segmentation: For each long paragraph, the sliding window will traverse the paragraphs in order to ensure that each paragraph is segmented reasonably while maintaining the continuity of the context.

[0105] Its advantages are:

[0106] Solve the problem of complex semantic segmentation: Semantic recognition segmentation can dynamically identify semantic units and use them as segmentation points, thus avoiding the problem of mis-segmentation caused by fixed rules.

[0107] Improve contextual consistency: Sliding window segmentation can ensure that paragraphs will not be arbitrarily cut in texts without clear titles or chapter markers, maintaining a certain contextual coherence.

[0108] Another implementation method is: structural information segmentation + sliding window segmentation

[0109] It is suitable for scenarios where there are clear titles or chapter divisions in the report, but these structural information is not detailed enough to make it difficult to further accurately segment the content. For example, the "Financial Data" section of the report has a big title, but the subtopics and paragraphs within the section are not clear.

[0110] The operation steps are:

[0111] First, use structural information segmentation: roughly segment the text according to the chapter titles in the report (such as "Financial Overview" or "ESG Assessment"), and extract each large block of text.

[0112] Then, use sliding window segmentation within each large block: In the "Financial Overview" section, the sliding window segmentation method can further segment each paragraph, that is, the target text block, and adjust the window size and overlap ratio according to semantic needs.

[0113] Its advantages are:

[0114] More detailed text processing: Structural information segmentation can provide clear boundaries for each large topic, while sliding window segmentation can further refine these large paragraphs, ensuring that each paragraph is of moderate length and maintains contextual coherence.

[0115] High flexibility: Even if the report lacks detailed structural information, the sliding window segmentation can automatically adjust and adapt to the text content to ensure the integrity of information extraction.

[0116] Another implementation method is: structural information segmentation + semantic recognition segmentation + sliding window segmentation

[0117] It is applicable to scenarios where the report contains both structural information (such as chapter titles) and a large amount of complex content without obvious separation. This combination mode is suitable for processing reports with both structural and unstructured information, and can take into account both outline segmentation and fine-grained semantic and context processing.

[0118] The operation steps are:

[0119] First, use structural information to segment: roughly divide the report and divide the text into multiple parts based on chapter headings.

[0120] Then, semantic recognition segmentation is used: within each chapter, a language model is used to perform further semantic analysis on the text to identify fine-grained paragraph boundaries.

[0121] Finally, use sliding window segmentation: perform sliding window processing on long paragraphs to ensure that the paragraphs are not over-segmented, and adjust the window size and overlap ratio when necessary.

[0122] Its advantages are:

[0123] Strong comprehensiveness: It combines the advantages of structured segmentation, semantic analysis and sliding windows, and can flexibly handle texts with complex structures and contents.

[0124] Strong comprehensiveness: It combines the advantages of structured segmentation, semantic analysis and sliding windows, and can flexibly handle texts with complex structures and contents.

[0125] Optionally, in order to improve the efficiency and accuracy of capturing indicator information, topic tags can be pre-built. The following is a possible implementation method. Specifically, Figure 1 On the basis of Figure 3 A flow chart of another method for extracting index data provided by an embodiment of the present invention is shown in FIG. Figure 3 , the method further comprising:

[0126] Step 102: Determine hierarchical topic labels and indicator sets based on business knowledge.

[0127] Optionally, the business knowledge may include information such as business requirements and domain knowledge. The indicator set may be divided into several groups by topic or subtopic, and each topic or subtopic corresponds to a group of related indicator sets:

[0128] T k = {I k1 , I k2 , ..., I kn}

[0129] Among them, Tk is the corresponding indicator set of the kth topic, I ki Theme T k The i-th index in .

[0130] Optionally, the following is an example of a theme-sub-theme-indicator, taking ESG business as an example:

[0131] Theme: Environmental Responsibility

[0132] Sub-theme: Carbon emissions

[0133] The corresponding indicator set:

[0134] 1. Total carbon emissions

[0135] 2. Carbon emission reduction rate

[0136] Sub-theme: Water Resources Management

[0137] The corresponding indicator set:

[0138] 1. Total water consumption

[0139] 2. Water saving measures

[0140] Sub-topic: Energy consumption

[0141] The corresponding indicator set:

[0142] 1. Total energy consumption

[0143] 2. Proportion of renewable energy use

[0144] Step 103: Establish a mapping relationship between each layer of topic labels and the corresponding indicator set.

[0145] Optionally, construct a set of hierarchical topic labels. The construction of labels follows the principle of gradual refinement from major topics to sub-topics, using example sentences and keywords as topic labels.

[0146] Step 104: Generate a corresponding embedding vector for the topic tag to form a topic space corresponding to the topic tag.

[0147] Optionally, generate an embedding vector v of the topic label based on a deep learning model (such as a Transformer model) T ,In this way, a standardized topic space is established, laying the foundation for the subsequent ,topic matching and indicator extraction.

[0148] The following is a possible calculation method for the matching score mentioned above. Figure 3 On the basis of Figure 4 A flow chart of another method for extracting index data provided by an embodiment of the present invention is shown in FIG. Figure 4 , step 105, comprising:

[0149] Step 105 - 1 , respectively calculating the first similarity between the target text block and each topic label in the topic space.

[0150] Optionally, the expression of the first similarity may be:

[0151]

[0152] sim(P, T) is the first similarity, v P is the embedding vector of the target text block, v T is the embedding vector of the topic label. P With v T All of them can be generated by the deep learning model described above.

[0153] Step 105 - 2 : Calculate the second similarity between the target text block and each topic tag in the topic space based on the Gestalt pattern.

[0154] Specifically, the Gestalt pattern is a sequence comparison algorithm that can be used to determine the matching relationship between text paragraphs and tags. Its implementation ideas are as follows:

[0155] Phase 1 - Data Preparation: Preprocess a set of target text blocks, including but not limited to: tokenization (segmenting text paragraphs into sequences of words or vocabulary), stop word removal (removing common, meaningless stop words), lemmatization (restoring words to their basic form (e.g., restoring "running" to "run")), and stemming (e.g., converting "fishing" and "fished" to "fish"). Obtain a set of preprocessed text segments.

[0156] Phase 2 - Pattern matching: Using the initialization matcher, the preprocessed text segment set is matched with each topic label in the topic space one by one to obtain the similarity ratio between the text segment and the topic label, which is the second similarity mentioned above.

[0157] Step 105 - 3 : Determine a matching score for each topic tag according to each first similarity and the corresponding second similarity.

[0158] Optionally, the expression of the matching score can be:

[0159] sim_gestalt(P,T)=λ×sim(P,T)+(1-λ)×Gestalt_Metric(P,T)

[0160] Among them, sim_gestalt(P, T) is the matching score, sim(P, T) is the first similarity, Gestalt_Metric(P, T) is the second similarity, and λ is the weight of the first similarity in the matching score.

[0161] Obviously, the above example takes into account the large difference in length between the target text block and the topic tag, and introduces Gestalt pattern matching to reduce the impact of text length differences and redundant information by identifying patterns and structures in the target text block, thereby improving the robustness of the matching process.

[0162] Furthermore, assuming that the matching threshold is α, each matching score is compared with the matching threshold α; if the matching score is greater than or equal to α, the corresponding topic label is determined as the target topic label corresponding to the target text block. That is, the paragraph P of the target text block matches the topic label T. The expression of this target indicator group is as follows:

[0163] f(P)=T k If sim_gestalt(P, T)>α

[0164] Then, referring to step 108 above, the target indicator group T k The paragraphs P of the target text block are input into the large model for batch extraction. The large model extracts the indicator information corresponding to the target indicator group from the report by deeply understanding the input content, combining context information and predefined indicator sets.

[0165] For each step of the above example, a possible architecture for implementing the index data extraction method is provided below. Specifically, Figure 5 A schematic diagram of an index data extraction architecture provided by an embodiment of the present invention, see Figure 5 The architecture includes: a parsing module 200, a segmentation module 201, a mapping module 202, a matching module 203 and a large model 204.

[0166] The parsing module 200 is used to parse the original text to obtain text data.

[0167] The segmentation module 201 is used for the above segmentation process to obtain the target text block.

[0168] Specifically, the segmentation process may include any one or a combination of the following methods: structural information segmentation, semantic recognition segmentation, or sliding window segmentation.

[0169] The mapping module 202 is used to determine hierarchical topic labels and indicator sets according to business knowledge, and then respectively establish a mapping relationship between each layer of topic labels and corresponding indicator sets, and finally generate corresponding embedding vectors for the topic labels to form a topic space corresponding to the topic labels.

[0170] The matching module 203 is used to match the target text block with the target topic tag.

[0171] The large model 204 is used to extract the original text through the target indicator group and the target text block to obtain the indicator information corresponding to the target indicator group.

[0172] The following provides a set of existing static extraction methods and the extraction method of this solution to compare the original text experimental data for different scenarios, as shown in Table 1.

[0173] Scenario Static extraction accuracy The accuracy of this solution Annual work report 63.23 80.56 Corporate ESG Report 51.41 73.86

[0174] As shown in Table 1, the accuracy of this solution is improved by 27%-44% compared with the existing static extraction.

[0175] Optionally, the present application also provides an implementation method of an index data extraction device, which can execute each step of the above example to achieve the corresponding technical effect. Specifically, Figure 6 A schematic diagram of an index data extraction device provided by an embodiment of the present invention, see Figure 6 , the device 30 comprises: a matching module 300 and an extraction module 301;

[0176] The matching module 300 is used to match the target text block with all the topic tags to obtain the matching score between the target text block and each topic tag; when the matching score meets the matching condition, determine the target topic tag corresponding to the target text block; and determine the corresponding target indicator group according to the target topic;

[0177] The extraction module 301 is used to input the target indicator group and the target text block into the large model, extract the original text, and obtain the indicator information corresponding to the target indicator group.

[0178] exist Figure 6 On the basis of Figure 7 A schematic diagram of another device for extracting index data provided by an embodiment of the present invention, see Figure 7 , the device 30, further includes: a parsing module 302, a segmentation module 303, and a mapping module 304;

[0179] A parsing module 302, used for parsing text data in the original text;

[0180] The segmentation module 303 is used to segment the text data to obtain target text blocks.

[0181] Optionally, a mapping module 304 is used to determine hierarchical topic labels and indicator sets based on business knowledge; establish a mapping relationship between each layer of topic labels and corresponding indicator sets; and generate corresponding embedding vectors for topic labels to form a topic space corresponding to the topic labels.

[0182] Optionally, the mapping module 304 is specifically used to calculate the first similarity between the target text block and each topic tag in the topic space respectively; calculate the second similarity between the target text block and each topic tag in the topic space based on the Gestalt pattern; and determine the matching score of each topic tag according to each first similarity and the corresponding second similarity.

[0183] Optionally, the matching module 300 is specifically configured to compare each matching score with a matching threshold respectively; if the matching score is greater than or equal to the matching threshold, the corresponding topic tag is determined as the target topic tag corresponding to the target text block.

[0184] The embodiment of the invention further provides an electronic device, which can execute all the steps of the above examples of the embodiment of the invention to achieve the corresponding technical effects. Specifically, Figure 8 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, see Figure 8 , the electronic device 50, comprises: a memory 501, a processor 500;

[0185] Memory 501, used to store one or more programs;

[0186] Processor 500;

[0187] When one or more programs are executed by the processor, when the electronic device 50 is used to execute the steps shown in the above-mentioned method examples, it can achieve each step and the corresponding technical effects.

[0188] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can also be implemented in other ways. The device embodiments described above are merely schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to the multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of the code, and a part of the module, program segment or code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0189] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0190] If the function is implemented 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 solution of the present application can be essentially or partly embodied in the form of a program product that contributes to the prior art, and the program product is stored in a computer-readable storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0191] The above are only preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. 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.

[0192] It will be apparent to those skilled in the art that the present application is not limited to the details of the exemplary embodiments described above, and that the present application can be implemented in other specific forms without departing from the spirit or essential features of the present application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present application. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for extracting index data, characterized in that: include: Matching the target text block with all the topic tags to obtain a matching score between the target text block and each topic tag; The target text block is obtained by slicing the original text; the topic label and the indicator group have a hierarchical correspondence; the indicator group includes various indicators corresponding to business knowledge; When the matching score satisfies the matching condition, a target topic tag corresponding to the target text block is determined; the target topic tag represents a topic tag among the topic tags that corresponds to the original text; Determine a corresponding target indicator group according to the target theme; the target indicator group represents the indicator group corresponding to the original text in the indicator group; The target indicator group and the target text block are input into a large model, the original text is extracted, and the indicator information corresponding to the target indicator group is obtained.

2. The method according to claim 1, characterized in that Before the step of matching the target text block with all the topic tags and obtaining the matching score between the target text block and each topic tag, the step further includes: Parsing text data in the original text; The text data is segmented to obtain the target text block.

3. The method according to claim 1, characterized in that Also includes: Determining a hierarchical set of topic labels and indicators based on the business knowledge; Establishing a mapping relationship between the topic label of each layer and the corresponding indicator set respectively; Generate a corresponding embedding vector for the topic label to form a topic space corresponding to the topic label.

4. The method according to claim 3, characterized in that The step of matching the target text block with all the topic tags and obtaining the matching score between the target text block and each topic tag comprises: Calculating first similarities between the target text block and each of the topic tags in the topic space respectively; Calculating a second similarity between the target text block and each of the topic tags in the topic space based on a Gestalt pattern; A matching score of each of the topic tags is determined according to each of the first similarities and the corresponding second similarities.

5. The method according to claim 1, characterized in that When the matching score satisfies the matching condition, the step of determining the target topic tag corresponding to the target text block includes: Comparing each of the matching scores with a matching threshold respectively; If the matching score is greater than or equal to the matching threshold, the corresponding topic tag is determined as the target topic tag corresponding to the target text block.

6. A device for extracting index data, characterized in that: include: A matching module, used to match the target text block with all the topic tags, and obtain a matching score between the target text block and each topic tag; The target text block is obtained by slicing the original text; the topic label and the indicator group have a hierarchical correspondence; the indicator group includes various indicators corresponding to business knowledge; when the matching score meets the matching condition, the target topic label corresponding to the target text block is determined; the target topic label represents the topic label corresponding to the original text among the topic labels; the corresponding target indicator group is determined according to the target topic; the target indicator group represents the indicator group corresponding to the original text among the indicator group; The extraction module is used to input the target indicator group and the target text block into the large model, extract the original text, and obtain indicator information corresponding to the target indicator group.

7. The device according to claim 6, characterized in that Also includes: A mapping module, used to determine hierarchical topic labels and indicator sets according to the business knowledge; and to establish a mapping relationship between the topic labels of each layer and the corresponding indicator sets; Generate a corresponding embedding vector for the topic label to form a topic space corresponding to the topic label.

8. An electronic device, characterized in that: include: A memory for storing one or more programs; processor; When the one or more programs are executed by the processor, the method according to any one of claims 1 to 5 is implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

10. A program product, characterized in that When the program product is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images