Enterprise annual report intelligent question and answer method and system based on large generative model

Abstract

The invention discloses an enterprise annual report intelligent question and answer method and system based on a generative large model, and the method comprises the following steps: obtaining and preprocessing an enterprise annual report text, and constructing a theme tag, an emotion tag and an annual report knowledge graph based on the preprocessed enterprise annual report text; carrying out keyword extraction and extension on the preprocessed enterprise annual report text, storing the enterprise annual report text into a keyword library, carrying out classification and layering processing on keywords in the keyword library, and constructing a keyword knowledge graph; analyzing a question text input by a user, matching the question text with a keyword library, and querying the keyword knowledge graph and the annual report knowledge graph according to a matching result to locate related enterprise annual report text fragments; and generating a final answer corresponding to the question text input by the user by calling a generative large model based on the positioned enterprise annual report text fragment and the theme tag and the emotion tag thereof. According to the method, efficient, accurate and comprehensive enterprise annual report intelligent question answering can be realized, and annual report data analysis in multiple fields is supported.

Description

Technical Field

[0001] The present invention belongs to the technical field of enterprise information management, and specifically relates to an intelligent question-answering method and system for enterprise annual reports based on a generative large model. Background Art

[0002] With the advent of the digital economy and the intelligent era, listed companies are disclosing more and more digital and intelligent information in their annual reports. This information is of great significance for investors, regulators and researchers to understand the company's development status, evaluate investment value and formulate regulatory policies.

[0003] Traditional manual analysis methods are difficult to efficiently and accurately extract and analyze this information from massive annual report texts. In the existing technology, natural language processing (NLP) technology is increasingly used to extract and analyze information from listed companies' annual reports. For example, the required corporate information is extracted from the annual report through keyword matching, rule extraction, etc., and then analyzed. However, these methods still have the following limitations:

[0004] (1) Limitations of keyword matching: The keyword matching method is highly dependent on a predefined keyword list, which makes it difficult to cover all relevant information and is easily affected by synonyms and antonyms, resulting in an incomplete and inaccurate information extraction process.

[0005] (2) Challenges of efficiency and adaptability of rule extraction: Rule extraction methods require manual formulation of complex extraction rules, which is a time-consuming and labor-intensive process and difficult to adapt to the diversity and complexity of annual report texts, resulting in low efficiency and accuracy of information extraction;

[0006] (3) Problems with semantic fuzziness and ambiguity: Most existing methods remain at the surface level of text analysis. For uncertain words such as “may” and “expected”, it is difficult to deeply analyze the true intention behind them, resulting in ambiguity in the analysis results.

[0007] (4) Insufficient contextual understanding ability: Certain information in the annual report requires context to be correctly understood. For example, the meaning of the same term or financial indicator may have different interpretations in different chapters. Existing NLP models find it difficult to accurately capture these nuances, which affects the in-depth understanding of the annual report content.

[0008] In recent years, generative large models (such as the GPT series) have made significant progress in the field of natural language processing. These models have powerful text generation, understanding and reasoning capabilities. For example, the Chinese patent application with publication number CN117312511A provides an intelligent question-answering learning method based on a generative large model, and the Chinese patent application with publication number CN118170894A provides a knowledge graph question-answering method, device and storage medium. Although these intelligent question-answering methods have been developed using large models, they ignore the graph relationships between local libraries or the deep and complex relationships in the knowledge graph of corporate annual reports. Generating answers only based on question text or using fuzzy calculations will result in answers that are inaccurate, too single and lack interpretability.

[0009] Therefore, the existing technology has the problems of incomplete, inaccurate and inefficient information extraction. How to use generative big models to effectively extract, analyze and evaluate information in the annual reports of listed companies remains a difficult problem that needs to be solved urgently. Summary of the invention

[0010] In view of the above, the purpose of the present invention is to provide a method and system for intelligent question answering of annual reports based on a generative big model, which utilizes the powerful language understanding and generation capabilities of the generative big model to conduct in-depth analysis of the annual report text, extract and expand keywords, and construct multi-dimensional and multi-level keyword knowledge graphs and annual report knowledge graphs. Through question keywords, the keyword knowledge graph and the annual report knowledge graph are quickly associated to accurately locate text fragments, realize intelligent question answering, and achieve comprehensive and accurate evaluation and prediction of the enterprise. At the same time, through a variety of natural language processing technologies and statistical methods, the accuracy of information extraction is ensured, the query response speed and quality are improved, and annual report data analysis in multiple fields is supported.

[0011] In order to achieve the above-mentioned invention object, the technical solution provided by the present invention is as follows:

[0012] An embodiment of the present invention provides an intelligent question-answering method for annual reports of enterprises based on a generative large model, comprising the following steps:

[0013] Obtain and preprocess the company's annual report text, and build topic tags, sentiment tags, and annual report knowledge graphs based on the preprocessed company's annual report text;

[0014] Extract and expand keywords from the pre-processed annual report text and store them in the keyword library. Classify and layer the keywords in the keyword library and build a keyword knowledge graph.

[0015] After parsing the question text entered by the user, it is matched with the keyword library, and the keyword knowledge graph and annual report knowledge graph are queried based on the matching results to locate relevant corporate annual report text fragments;

[0016] Based on the located corporate annual report text fragments and their topic tags and sentiment tags, the final answer corresponding to the question text entered by the user is generated by calling the generative big model.

[0017] Preferably, constructing a topic tag based on the preprocessed annual report text includes:

[0018] The latent Dirichlet distribution is used to perform topic modeling on the preprocessed annual report text, and corresponding topic labels are generated for each annual report text fragment. The topic labels include financial status, operating performance, R&D innovation, human resources, strategy and risk management, environmental and social responsibility, law and compliance, shareholder information and competitive analysis.

[0019] Preferably, constructing sentiment labels based on the preprocessed annual report text includes:

[0020] A pre-trained large language model is used to perform sentiment analysis on the pre-processed annual report text, and a sentiment label including sentiment polarity and sentiment intensity is generated for each annual report text fragment, where sentiment intensity includes positive, neutral and negative, and sentiment intensity includes low intensity, medium intensity and high intensity. The sentiment analysis results are then adjusted and corrected based on the contextual information of the text fragment.

[0021] Preferably, the annual report knowledge graph is constructed based on the preprocessed annual report text, including:

[0022] A BERT-based named entity recognition model is used to identify entities in the preprocessed annual report text. A relation extraction method based on dependency syntactic analysis is used to extract the relationships between entities from the preprocessed annual report text. An annual report knowledge graph is constructed with entities as nodes and relationships between entities as edges.

[0023] Preferably, extracting and expanding keywords from the pre-processed annual report text and storing them in a keyword library includes:

[0024] The word frequency-inverse document frequency method is used to calculate the high-frequency words in the preprocessed annual report text as candidate keywords. The named entity recognition method is used to identify the entities in the preprocessed annual report text as candidate keywords. The pre-trained large language model is combined with the clustering algorithm to extract representative words in the preprocessed annual report text as candidate keywords. The generative large model is used to semantically expand all candidate keywords, and all candidate keywords and semantically expanded keywords are stored in the keyword library.

[0025] Preferably, the classification and stratification of the keywords in the keyword library and construction of a keyword knowledge graph includes:

[0026] Input the keywords in the keyword library into the pre-trained large language model and convert them into vector representations. Calculate the semantic similarity between the vector representations. Cluster the keywords based on the semantic similarity using a clustering algorithm. Generate a hierarchical structure of keywords based on the semantic similarity and clustering results using a hierarchical clustering algorithm. Classify the keywords in the keyword library hierarchically.

[0027] According to the hierarchically classified keywords, the relationships and connections between each keyword in different dimensions including functions, technologies and application scenarios are further established to establish the association relationship between keywords, so as to form a multi-level and multi-dimensional keyword library, and the keyword knowledge graph is constructed with keywords as nodes and the association relationship between multi-level and multi-dimensional keywords as edges;

[0028] Establish a direct association between the entities in the annual report knowledge graph and the same keywords in the keyword knowledge graph, and establish attribute association between multiple keywords in the keyword knowledge graph corresponding to multiple attributes of the entities in the annual report knowledge graph.

[0029] Preferably, the parsing of the question text input by the user is matched with the keyword library, and querying the keyword knowledge graph and the annual report knowledge graph according to the matching result to locate the relevant corporate annual report text fragments, including:

[0030] The question text entered by the user is parsed and the initial keywords are extracted. The initial keywords are matched with the keyword library to obtain the same initial keywords and expanded related keywords. The graph traversal algorithm is used to query the keyword nodes corresponding to the initial keywords and related keywords in the keyword knowledge graph. Further, based on the direct association and attribute association between the keyword knowledge graph and the annual report knowledge graph, the entity nodes corresponding to the keyword nodes are queried in the annual report knowledge graph, and the relevant corporate annual report text fragments are located according to the entity nodes.

[0031] Preferably, before calling the generative large model to generate the final answer, it also includes:

[0032] For the located relevant annual report text fragments, the frequency of the initial keywords and the expanded related keywords in the question text input by the user are counted through a word frequency counting algorithm.

[0033] Preferably, the generative large model generates a final answer including the word frequency of the initial keywords and related keywords based on the annual report text fragment and context information, topic tags, sentiment tags, and word frequency.

[0034] To achieve the above-mentioned purpose of the invention, the embodiment of the present invention also provides an intelligent question-answering system for annual reports of enterprises based on a generative large model, including: an annual report knowledge graph construction module, a keyword knowledge graph construction module, a question analysis and query positioning module and a model answer generation module;

[0035] The annual report knowledge graph construction module is used to obtain and preprocess the enterprise annual report text, and to construct topic tags, sentiment tags and annual report knowledge graphs based on the preprocessed enterprise annual report text;

[0036] The keyword knowledge graph construction module is used to extract and expand keywords from the pre-processed annual report text and store them in a keyword library, classify and hierarchically process the keywords in the keyword library and construct a keyword knowledge graph;

[0037] The question analysis and query positioning module is used to analyze the question text input by the user and match it with the keyword library, and query the keyword knowledge graph and the annual report knowledge graph according to the matching results to locate the relevant corporate annual report text fragments;

[0038] The model answer generation module is used to generate a final answer corresponding to the question text input by the user by calling the generative large model based on the located corporate annual report text fragments and their theme tags and sentiment tags.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The present invention expands and classifies keywords in a hierarchical manner during the process of establishing a keyword knowledge graph, which helps to cover all relevant keywords and provide a clear hierarchical structure, helping the system to understand and match user questions more accurately and comprehensively, thereby generating more precise and comprehensive answers.

[0041] (2) The present invention can query the keyword knowledge graph and the annual report knowledge graph according to the matching results of the user's questions to locate the relevant annual report text fragments, and can quickly, accurately and comprehensively locate the relevant text fragments according to the user's questions, realize the mining of multi-dimensional connections between the annual report text and keywords, and can help identify potential connections between enterprises, providing new ideas for the prediction of technology development trends and the analysis of the competitive landscape.

[0042] (3) The present invention generates answers by combining the text fragments of the annual report and its subject tags and sentiment tags as well as the graph-related information and the user's question text, thereby ensuring the accuracy and traceability of the answers and further ensuring the robustness of intelligent question answering, so that it can be better applied to scientific research, information science, and corporate evaluation. It can illuminate the algorithm black box to a certain extent, improve the interpretability and reproducibility of the results, and reduce algorithmic risks. BRIEF DESCRIPTION OF THE DRAWINGS

[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, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 It is a flowchart of the intelligent question-answering method for annual reports of enterprises based on a generative large model provided by an embodiment of the present invention;

[0045] Figure 2 It is a schematic diagram of a keyword knowledge graph and an annual report knowledge graph provided by an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of matching and expanding a user question with a keyword library in question example 1 provided in an embodiment of the present invention. Figure 1 ;

[0047] Figure 4 This is a schematic diagram of matching and expanding a user question with a keyword library in question example 1 provided in an embodiment of the present invention. Figure 2 ;

[0048] Figure 5 This is a schematic diagram of matching and expanding a user question with a keyword library in question example 1 provided in an embodiment of the present invention. Figure 3 ;

[0049] Figure 6 This is a schematic diagram of matching and expanding a user question with a keyword library in question example 1 provided in an embodiment of the present invention. Figure 4 ;

[0050] Figure 7 is a schematic diagram of matching and expanding a user question with a keyword library in question example 2 provided by an embodiment of the present invention;

[0051] Figure 8 is a schematic diagram of matching and expanding a user question with a keyword library in question example 3 provided by an embodiment of the present invention;

[0052] Fig. 9 It is a structural diagram of the intelligent question-answering system for annual reports of enterprises based on a generative big model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0054] The inventive concept of the present invention is: in response to the problems of incomplete, inaccurate and inefficient information extraction in the knowledge question and answer of corporate annual reports in the prior art, the embodiments of the present invention provide a method and system for intelligent question and answer of corporate annual reports based on a generative big model, by respectively constructing an annual report knowledge graph and a keyword knowledge graph for the corporate annual report text, when facing user questions, it is possible to directly perform queries and locate text fragments based on the keyword knowledge graph and the annual report knowledge graph, thereby achieving efficient, accurate and comprehensive intelligent question and answer of corporate annual reports, and supporting annual report data analysis in multiple fields.

[0055] Figure 1 Schematic diagram of the process of intelligent question-answering method for annual reports based on generative big model provided by an embodiment of the present invention. Figure 1 As shown, the embodiment provides an intelligent question-answering method for annual reports of enterprises based on a generative large model, comprising the following steps:

[0056] S1, obtain the annual report text and preprocess it, and build topic tags, sentiment tags and annual report knowledge graph based on the preprocessed annual report text.

[0057] S1.1, Preprocessing of corporate annual report text.

[0058] Obtain the text of the company's annual report, convert it into a machine-readable standard text, clean the standard text and detect outliers, including identifying and correcting rare word combinations or expressions, eliminating content that does not conform to the subject or field, detecting and correcting anomalies in the data format or structure, and finally storing the pre-processed annual report text in the database.

[0059] S1.2, construct topic tags for the pre-processed annual report text.

[0060] Latent Dirichlet Allocation (LDA) is used to perform topic modeling on the preprocessed annual report text, classify different annual report topics, select texts with adjacent clusters for separate analysis, slice single texts to obtain text fragments, and generate corresponding topic labels for each text fragment. The topic labels include financial status, operating performance, R&D innovation, human resources, strategy and risk management, environment and social responsibility, law and compliance, shareholder information and competition analysis.

[0061] S1.3, construct sentiment labels for the preprocessed annual report text.

[0062] Use pre-trained large language models (including GPT-4 or BERT, etc.) to perform sentiment analysis on the preprocessed annual report text, and generate sentiment labels including sentiment polarity and sentiment intensity for each annual report text fragment, where sentiment intensity includes positive, neutral and negative, and sentiment intensity includes low intensity, medium intensity and high intensity. Then, adjust and correct the sentiment analysis results based on the contextual information of the text fragment.

[0063] S1.4, construct an annual report knowledge graph for the preprocessed annual report text.

[0064] The named entity recognition (NER) model based on BERT (Bidirectional Encoder Representations from Transformers) is used to identify entities (such as company names, product names, and technology names, etc.) in the preprocessed annual report text. The relation extraction (RE) method based on dependency syntactic analysis is used to extract the relationship between entities from the preprocessed annual report text. The processed triples are stored in the library, and the annual report knowledge graph is constructed with entities as nodes and the relationship between entities as edges.

[0065] S2, extract and expand the keywords of the pre-processed annual report text and store them in the keyword library, classify and layer the keywords in the keyword library and construct a keyword knowledge graph.

[0066] S2.1, keyword extraction and expansion.

[0067] S2.1.1, use statistical methods to extract keywords. Use the term frequency-inverse document frequency (TF-IDF) method to calculate the TF-IDF value of each word in the preprocessed annual report text, and extract multiple high-frequency words with TF-IDF higher than the set threshold as candidate keywords.

[0068] S2.1.2, using natural language processing (NLP) methods to extract keywords, including using named entity recognition (NER) methods, question and answer (Q&A) systems or sentiment analysis methods, etc. In the embodiment, the NER method is used to identify entities (such as company names, product names, and technology names, etc.) in the pre-processed annual report text as candidate keywords.

[0069] S2.1.3, use a large language model (LLM) method to extract keywords, including using a BERT or ChatGPT model, etc. In the embodiment, a BERT-base pre-trained large language model is used to encode the preprocessed corporate annual report text, extract contextual semantic information, and extract representative words from the encoded text as candidate keywords through a clustering algorithm.

[0070] S2.1.4, use a generative large model (including ChatGPT, Claude or Gemini, etc.) to perform semantic expansion (including synonym replacement, semantic expansion and other data enhancement operations) on all candidate keywords, and store all candidate keywords and semantically expanded keywords in the keyword library. In order to skip the initial manual labeling, the zero-shot learning or few-shot learning capabilities of the generative large model can be used to extract keywords and semantically expand keywords with little or no manual labeling, so as to reduce labor costs.

[0071] S2.2, construct keyword knowledge graph.

[0072] S2.2.1, input the keywords in the keyword library into the pre-trained large language model (including GPT-4 or BERT, etc.) and convert them into vector representations. Use cosine similarity to calculate the semantic similarity between vector representations to obtain a semantic similarity matrix. Based on semantic similarity, use the K-means clustering algorithm to cluster the keywords, find potential categories, and achieve preliminary classification. The classification includes corporate entities, technical field dimensions, industry field dimensions, application field dimensions, time dimensions, financial dimensions, and product dimensions. Based on the semantic similarity and clustering results of the keywords, a hierarchical clustering algorithm is used to generate a hierarchical structure of keywords, and the keywords in the keyword library are further hierarchically classified from broad to specific to form a tree structure.

[0073] S2.2.2, according to the hierarchically classified keywords, each keyword can have different relationships and connections in different dimensions. The relationships and connections of each keyword in different dimensions including functions, technologies and application scenarios are further used to establish the association relationship between keywords. For example, "intelligent manufacturing" is not only related to "data analysis", but also to technical dimensions such as "real-time monitoring" and "production efficiency". Finally, a complex, multi-level and multi-dimensional keyword library is formed, and the keyword knowledge graph is constructed with keywords as nodes and the association relationships between multi-level and multi-dimensional keywords as edges.

[0074] Specifically, the hierarchical structure generated in step S2.2.1 contains the subordinate relationships between the keywords. First, these hierarchical relationships are converted into nodes and edges in the graph to construct a preliminary knowledge graph, such as Figure 2 As shown in the figure, "technological transformation" is the first-level classification, "automated production", "digital transformation", "green technology" and "intelligent manufacturing" are the second-level classification, and "energy-saving technology" and "Internet of Things technology" are the third-level classification. Secondly, based on the classification results of step S2.2.1, the association of each keyword in different dimensions is further analyzed. In addition to the hierarchical relationship, semantic similarity, functional association, etc. are also considered to add more relationships. For example, "intelligent manufacturing" is not only related to "automated production", but also to technologies such as "sensor technology". Thirdly, the semantic similarity matrix and clustering results calculated in step S2.2.1 help identify more potential relationships, and use this information to construct association nodes between keywords to ensure that the relationship of each keyword in the knowledge graph is more comprehensive and accurate. For example, in the functional dimension, the keyword of technological transformation is associated with business process optimization, market competitiveness improvement and customer satisfaction improvement; in the technical dimension, it is associated with artificial intelligence, big data analysis, cloud computing, and the Internet of Things (IoT); in the application scenario dimension, it is associated with manufacturing (such as production line optimization and supply chain management), retail (such as intelligent inventory management and customer behavior analysis), and finance (such as risk assessment and automated trading).

[0075] S2.2.3, if Figure 2 As shown, a direct association is established between the entities in the annual report knowledge graph and the same keywords in the keyword knowledge graph, and an attribute association is established between multiple keywords in the keyword knowledge graph corresponding to multiple attributes of the entities in the annual report knowledge graph.

[0076] S3, after parsing the question text input by the user, matches it with the keyword library, and queries the keyword knowledge graph and annual report knowledge graph based on the matching results to locate relevant corporate annual report text fragments.

[0077] S3.1, query the graph to locate the text fragments of the company's annual report.

[0078] S3.1.1, use natural language processing (NLP) technology to perform grammatical analysis on the question text entered by the user and extract the core intent and initial keywords of the question.

[0079] S3.1.2, match the extracted initial keywords with the keyword library to obtain the same initial keywords and expanded related keywords, and use the graph traversal algorithm to query the keyword nodes corresponding to the initial keywords and related keywords in the keyword knowledge graph.

[0080] S3.1.3, further based on the direct association and attribute association between the keyword knowledge graph and the annual report knowledge graph, query the entity node corresponding to the keyword node in the annual report knowledge graph, and locate the relevant corporate annual report text fragment according to the entity node. In the embodiment, firstly, the matched keywords are queried in the keyword knowledge graph to obtain the keywords and relationships associated with them, and then the vector representation of the keywords (initial keywords and related keywords) is matched with the vector representation of the entities in the annual report knowledge graph, matching the entities with the highest cosine similarity, confirming the positions of these entities in the annual report knowledge graph and other associated entities and relationships, and finally querying these entities in the annual report knowledge graph to locate the text fragments where the entities are located.

[0081] S4, based on the located annual report text fragments and their topic tags and sentiment tags, generates the final answer corresponding to the question text input by the user by calling the generative big model.

[0082] S4.1, process the located annual report text fragments.

[0083] For the relevant annual report text fragments located, the frequency of the initial keywords and the expanded related keywords in the question text entered by the user is counted through the word frequency statistics algorithm. The word frequency statistics results are used to measure the importance of keywords in the relevant text fragments.

[0084] S4.2, Generate answers using generative large models.

[0085] Generative large models (including ChatGPT, Claude or Gemini, etc.) generate the final answer including the frequency of initial keywords and related keywords based on the annual report text fragments and contextual information, topic tags, sentiment tags, and word frequency. In addition, it also includes optimizing and verifying the generated answers to ensure accuracy and relevance, and providing relevant background information and explanations for the answers.

[0086] Based on the method for intelligent question and answering of annual reports of enterprises based on a generative large model provided in the above-mentioned embodiment, the following provides example questions for intelligent question and answering in annual reports of different enterprises.

[0087] Example 1:

[0088] User question: How will Company A perform in terms of environment, social and governance (ESG) in 2023?

[0089] (1) Question text analysis:

[0090] Extract the core keywords in the question: "Company A", "2023", "environment", "society", "governance", "ESG", "specific measures", and "performance".

[0091] (2) Keyword matching and expansion:

[0092] According to the parsed question, the keyword library is matched and other nodes related to the keyword are expanded. The result is as follows: Figure 3 , Figure 4 , Figure 5 and Figure 6 shown.

[0093] (3) Text fragment positioning:

[0094] The initial keywords and the expanded related keywords are matched with the annual report knowledge graph through cosine similarity calculation. Taking the matched entity node as the center, all graph data connected to it with relationship edges are retrieved to lock the relevant text fragments.

[0095] (4) Text fragment processing and analysis:

[0096] Count the number of times the keywords and related keywords parsed from the user's question text appear in the text fragment.

[0097] (5) Answer generation and optimization:

[0098] Generate answers: Invoke the generative big model to generate preliminary answers based on the located text snippets, topic tags, sentiment tags, and word frequencies.

[0099] Answer optimization and verification: The generated answers are optimized and verified to ensure their accuracy and relevance.

[0100] Final answer generation: Output the final answer including the frequency of keywords and their related words.

[0101] (6) Example of final answer:

[0102] Basic information of Company A, including industry, scale, geographical location, etc.

[0103] The specific ESG measures and performance of Company A in its 2023 annual report are as follows:

[0104] Environmental Protection:

[0105] Reducing emissions: Company A actively took a number of measures to reduce carbon emissions in 2023, including introducing clean energy and improving production processes. The annual report mentioned "carbon emissions" 15 times, reflecting the company's efforts in environmental protection.

[0106] Resource Management: The company has implemented a number of water management and energy efficiency improvement programs to reduce resource waste and environmental impact.

[0107] Ecological protection: The annual report mentions ecological protection-related measures, such as protecting natural resources and reducing environmental pollution, a total of 8 times.

[0108] society:

[0109] Employee benefits: Company A focuses on the career development and welfare benefits of its employees. The term "employee benefits" appears 12 times in the annual report, with detailed descriptions of employee training, career development plans, and welfare systems.

[0110] Community Involvement: The company actively participates in community building. The annual report mentions "community involvement" five times, describing the company's activities in volunteer services and community donations.

[0111] Governance:

[0112] Anti-corruption measures: Company A has adopted strict internal control and audit measures to ensure the transparency and fairness of the company's operations. "Anti-corruption measures" appeared 7 times in the annual report, demonstrating the company's specific actions in governance.

[0113] Board Governance: The annual report describes in detail the governance structure and working mechanism of the board of directors to ensure the scientific and compliance of the company's decision-making.

[0114] Example 2:

[0115] User question: In 2023, to what extent will Company A apply generative AI?

[0116] (1) Question text analysis:

[0117] Use NLP technology to parse user questions and extract core intent and keywords.

[0118] (2) Keyword matching and expansion:

[0119] Match the extracted keywords with the keyword library to find related extended keywords, such as Figure 7 The following are some examples of keywords.

[0120] (3) Text fragment positioning:

[0121] The initial keywords and the expanded related keywords are matched with the annual report graph through cosine similarity calculation. Taking the matched entity node as the center, all graph data connected with the relationship edges are retrieved to lock the relevant text fragments.

[0122] (4) Text fragment processing and analysis:

[0123] Count the number of times the keywords and related keywords parsed from the user's question text appear in the text fragment.

[0124] (5) Answer generation and optimization:

[0125] Generate answers: Invoke the generative big model to generate preliminary answers based on the located text snippets, topic tags, sentiment tags, and word frequencies.

[0126] Answer optimization and verification: The generated answers are optimized and verified to ensure their accuracy and relevance.

[0127] Final answer generation: Output the final answer including the frequency of keywords and their related words.

[0128] (6) Example of final answer:

[0129] Keywords: Generative AI, generative model, deep learning, text generation, technology application, benefits.

[0130] Word frequency statistics: generative AI (10 times), generative models (7 times), deep learning (5 times), and text generation (3 times).

[0131] In the 2023 annual report of Company A, generative AI was mentioned 10 times, generative models 7 times, deep learning 5 times, and text generation 3 times. The report details the company's investment and technical implementation in generative AI, the significant business results and improvements brought to Company A by the application of generative AI, and its future plans and strategic direction for generative AI.

[0132] Enterprise investment and technology implementation in generative AI:

[0133] Investment: Company A has made significant investments in generative AI technology, including setting up a dedicated R&D department, purchasing advanced computing equipment, and increasing its budget for related technologies.

[0134] Technology cooperation: The company has established partnerships with a number of leading AI technology companies and research institutions to jointly develop and apply generative AI technology;

[0135] Rulemaking: Participated in the formulation of AI-related standards

[0136] The application of generative AI has brought significant business results and improvements to Company A:

[0137] Improved efficiency: Through the use of automated processes and intelligent tools, production efficiency has increased by 20% and customer service response time has been reduced by 50%;

[0138] Increased revenue: The company sees sales increase by 15% in 2023 due to product innovation and increased go-to-market efficiency;

[0139] Customer experience: Customer satisfaction indicators increased by 10%, and customer retention rates increased significantly.

[0140] Future plans and strategic directions for generative AI:

[0141] Continued investment: In the next three years, Company A plans to continue to increase its investment in generative AI, especially in the research and development of cutting-edge technologies;

[0142] Expanded applications: Expand the application of generative AI to more business areas, such as supply chain management, financial analysis, and human resource management;

[0143] Talent cultivation: Establish an internal AI training program to cultivate more professionals with generative AI technical capabilities.

[0144] Example 3:

[0145] User question: What is the resilience of Company A’s supply chain in 2023?

[0146] (1) Question text analysis:

[0147] Use NLP technology to parse user questions and extract core intent and keywords.

[0148] (2) Keyword matching and expansion:

[0149] Match the extracted keywords with the keyword library to find related extended keywords, such as Figure 8 The following are some examples of keywords.

[0150] (3) Text fragment positioning:

[0151] The initial keywords and the expanded keywords are matched with the annual report graph through cosine similarity calculation. Taking the matched entity node as the center, all graph data connected with the relationship edges are retrieved to lock the relevant text fragments.

[0152] (4) Text fragment processing and analysis:

[0153] Count the number of times the keywords and related keywords parsed from the user's question text appear in the text fragment.

[0154] (5) Answer generation and optimization:

[0155] Generate answers: Invoke the generative big model to generate preliminary answers based on the located text snippets, topic tags, sentiment tags, and word frequencies.

[0156] Answer optimization and verification: The generated answers are optimized and verified to ensure their accuracy and relevance.

[0157] Final answer generation: Output the final answer including the frequency of keywords and their related words.

[0158] (6) Example of final answer:

[0159] In Company A's 2023 annual report, supply chain resilience was mentioned 12 times, risk management 8 times, supply chain diversification 6 times, inventory management 4 times, supply chain management system 5 times, data analysis tools 3 times, supply chain visibility 2 times, response speed 3 times, adaptability 3 times, supply chain digitization 4 times, intelligence 3 times, geopolitics 1 time, natural disasters 2 times, supply chain risk management 2 times, optimization capabilities 2 times, and supply chain partners 2 times.

[0160] In 2023, Company A has made significant progress in supply chain resilience. The company has enhanced the flexibility and responsiveness of its supply chain by implementing a number of measures, including supplier diversification, inventory management optimization, digital technology applications, risk management strategies, and partnerships. These measures include:

[0161] (a) Increasing the number of suppliers of key raw materials and components, reducing reliance on a single supplier, and reducing the risk of supply chain disruptions;

[0162] (b) Introducing advanced inventory management systems and adopting dynamic inventory strategies to ensure that production and delivery can be maintained in the event of supply chain disruptions;

[0163] (c) Applying big data analysis and Internet of Things technologies to achieve visualization and real-time monitoring of the supply chain, improving the ability to respond quickly to market changes;

[0164] (d) Develop a detailed supply chain risk management plan, including risk identification, assessment and response measures, to quickly respond to emergencies;

[0165] (e) Strengthen collaboration and communication with partners at all links of the supply chain, share information and resources, and jointly improve the overall resilience of the supply chain.

[0166] In summary, an intelligent question-answering method for annual reports based on a generative large model provided by an embodiment of the present invention expands and classifies and hierarchically processes keywords in the process of establishing a keyword knowledge graph, which helps to cover all relevant keywords and provide a clear hierarchical structure, helping the system to understand and match user questions more accurately and comprehensively, thereby generating more accurate and comprehensive answers; it can query the keyword knowledge graph and the annual report knowledge graph according to the matching results of the user's questions to locate relevant annual report text fragments, and can quickly, accurately and comprehensively locate relevant text fragments according to user questions, realize the mining of multi-dimensional connections between annual report texts and keywords, and can help identify potential connections between enterprises, providing new ideas for technology development trend prediction and competition landscape analysis; by combining annual report text fragments and their topic tags and sentiment tags as well as graph-related information and user question text to generate answers, the accuracy and traceability of the answers are guaranteed, and the robustness of intelligent question-answering is further guaranteed, so that it can be better applied to scientific research, information science, enterprise evaluation and other fields, and can illuminate the algorithm black box to a certain extent, improve the interpretability and reproducibility of the results, and reduce algorithm risks.

[0167] Based on the same inventive concept, Fig. 9 As shown, an embodiment of the present invention also provides an intelligent question-answering system 900 for annual reports of enterprises based on a generative big model, including: an annual report knowledge graph construction module 910, a keyword knowledge graph construction module 920, a question analysis and query positioning module 930 and a model answer generation module 940.

[0168] Among them, the annual report knowledge graph construction module 910 is used to obtain and preprocess the company's annual report text, and construct topic tags, sentiment tags and annual report knowledge graphs based on the preprocessed company annual report text.

[0169] The keyword knowledge graph construction module 920 is used to extract and expand the keywords of the pre-processed annual report text and store them in the keyword library, classify and hierarchize the keywords in the keyword library and construct a keyword knowledge graph.

[0170] The question analysis and query positioning module 930 is used to analyze the question text input by the user and match it with the keyword library, and query the keyword knowledge graph and the annual report knowledge graph according to the matching results to locate the relevant corporate annual report text fragments.

[0171] The model answer generation module 940 is used to generate a final answer corresponding to the question text input by the user by calling the generative large model based on the located corporate annual report text fragments and their topic tags and sentiment tags.

[0172] It should be noted that the intelligent question and answer system for annual reports based on the generative big model provided in the above-mentioned embodiment belongs to the same inventive concept as the intelligent question and answer method for annual reports based on the generative big model. The specific implementation process is detailed in the embodiment of the intelligent question and answer method for annual reports based on the generative big model, which will not be repeated here.

[0173] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent question-answering method for corporate annual reports based on a generative big model. It is characterized in that The following steps are involved: Obtain and preprocess the company's annual report text, and build topic tags, sentiment tags, and annual report knowledge graphs based on the preprocessed company's annual report text; Extract and expand keywords from the pre-processed annual report text and store them in the keyword library. Classify and layer the keywords in the keyword library and build a keyword knowledge graph. After parsing the question text entered by the user, it is matched with the keyword library, and the keyword knowledge graph and annual report knowledge graph are queried based on the matching results to locate relevant corporate annual report text fragments; Based on the located corporate annual report text fragments and their topic tags and sentiment tags, the final answer corresponding to the question text entered by the user is generated by calling the generative big model.

2. According to the method for intelligent question answering of annual reports of enterprises based on generative big model in claim 1, It is characterized in that Construct topic labels based on the preprocessed annual report text, including: The latent Dirichlet distribution is used to perform topic modeling on the preprocessed annual report text, and corresponding topic labels are generated for each annual report text fragment. The topic labels include financial status, operating performance, R&D innovation, human resources, strategy and risk management, environmental and social responsibility, law and compliance, shareholder information and competitive analysis.

3. The intelligent question-answering method for annual reports based on a generative big model according to claim 1, It is characterized in that Construct sentiment labels based on the preprocessed annual report text, including: A pre-trained large language model is used to perform sentiment analysis on the pre-processed annual report text, and a sentiment label including sentiment polarity and sentiment intensity is generated for each annual report text fragment, where sentiment intensity includes positive, neutral and negative, and sentiment intensity includes low intensity, medium intensity and high intensity. The sentiment analysis results are then adjusted and corrected based on the contextual information of the text fragment.

4. The intelligent question-answering method for annual reports based on a generative big model according to claim 1, It is characterized in that Construct an annual report knowledge graph based on the preprocessed annual report text, including: A BERT-based named entity recognition model is used to identify entities in the preprocessed annual report text. A relation extraction method based on dependency syntactic analysis is used to extract the relationships between entities from the preprocessed annual report text. An annual report knowledge graph is constructed with entities as nodes and relationships between entities as edges.

5. The intelligent question-answering method for annual reports based on a generative big model according to claim 1, It is characterized in that The keyword extraction and expansion of the pre-processed annual report text are then stored in the keyword library, including: The word frequency-inverse document frequency method is used to calculate the high-frequency words in the preprocessed annual report text as candidate keywords. The named entity recognition method is used to identify the entities in the preprocessed annual report text as candidate keywords. The pre-trained large language model is combined with the clustering algorithm to extract representative words in the preprocessed annual report text as candidate keywords. The generative large model is used to semantically expand all candidate keywords, and all candidate keywords and semantically expanded keywords are stored in the keyword library.

6. The intelligent question-answering method for annual reports based on a generative big model according to claim 4, It is characterized in that The method of classifying and stratifying the keywords in the keyword library and constructing a keyword knowledge graph includes: Input the keywords in the keyword library into the pre-trained large language model and convert them into vector representations. Calculate the semantic similarity between the vector representations. Cluster the keywords based on the semantic similarity using a clustering algorithm. Generate a hierarchical structure of keywords based on the semantic similarity and clustering results using a hierarchical clustering algorithm. Classify the keywords in the keyword library hierarchically. According to the hierarchically classified keywords, the relationships and connections between each keyword in different dimensions including functions, technologies and application scenarios are further established to establish the association relationship between keywords, so as to form a multi-level and multi-dimensional keyword library, and the keyword knowledge graph is constructed with keywords as nodes and the association relationship between multi-level and multi-dimensional keywords as edges; Establish a direct association between the entities in the annual report knowledge graph and the same keywords in the keyword knowledge graph, and establish attribute association between multiple keywords in the keyword knowledge graph corresponding to multiple attributes of the entities in the annual report knowledge graph.

7. The method for intelligent question answering of annual reports based on a generative big model according to claim 6, It is characterized in that The question text input by the user is parsed and matched with the keyword library, and the keyword knowledge graph and the annual report knowledge graph are queried according to the matching results to locate the relevant corporate annual report text fragments, including: The question text entered by the user is parsed and the initial keywords are extracted. The initial keywords are matched with the keyword library to obtain the same initial keywords and expanded related keywords. The graph traversal algorithm is used to query the keyword nodes corresponding to the initial keywords and related keywords in the keyword knowledge graph. Further, based on the direct association and attribute association between the keyword knowledge graph and the annual report knowledge graph, the entity nodes corresponding to the keyword nodes are queried in the annual report knowledge graph, and the relevant corporate annual report text fragments are located according to the entity nodes.

8. The intelligent question-answering method for annual reports based on a generative big model according to claim 1, It is characterized in that Before calling the generative model to generate the final answer, it also includes: For the located relevant annual report text fragments, the frequency of the initial keywords and the expanded related keywords in the question text input by the user are counted through a word frequency counting algorithm.

9. The intelligent question-answering method for annual reports of enterprises based on a generative big model according to claim 8, It is characterized in that The generative large model generates the final answer including the word frequency of initial keywords and related keywords based on the company's annual report text fragments and context information, topic tags, sentiment tags, and word frequency.

10. An intelligent question-answering system for corporate annual reports based on a generative big model. It is characterized in that include: Annual report knowledge graph construction module, keyword knowledge graph construction module, question analysis and query positioning module and model answer generation module; The annual report knowledge graph construction module is used to obtain and preprocess the enterprise annual report text, and to construct topic tags, sentiment tags and annual report knowledge graphs based on the preprocessed enterprise annual report text; The keyword knowledge graph construction module is used to extract and expand keywords from the pre-processed annual report text and store them in a keyword library, classify and hierarchically process the keywords in the keyword library and construct a keyword knowledge graph; The question analysis and query positioning module is used to analyze the question text input by the user and match it with the keyword library, and query the keyword knowledge graph and the annual report knowledge graph according to the matching results to locate the relevant corporate annual report text fragments; The model answer generation module is used to generate a final answer corresponding to the question text input by the user by calling the generative large model based on the located corporate annual report text fragments and their theme tags and sentiment tags.

Images