Financial text logicality evaluation method and system based on AI Agent

Abstract

The invention belongs to the technical field of large language models, and provides a financial text logicality evaluation method and system based on an AI Agent. The method comprises the following steps: performing digital conversion and preprocessing on an original financial text, segmenting according to a logic structure, and extracting key information to obtain structured standard data; performing fine tuning and initialization on the pre-trained large language model, and constructing a reasoning Agent and an reflecting Agent; based on a chain thinking prompt and a logic knowledge base, a reasoning Agent is guided to analyze the structured standard data, and logic vulnerabilities are identified; the method comprises the following steps of: performing rechecking through a reflection Agent, performing classification integration on logic vulnerabilities through a topic modeling technology, outputting a logic error report, performing visual display, and performing user-defined format export. According to the method, deep context expression and complex logic errors in the financial text can be accurately recognized, and financial text analysis requirements under different business scenes are met.

Description

Technical Field

[0001] The present invention relates to the technical field of large language models, and in particular to an AI agent-based financial text logic evaluation method and system. Background Art

[0002] In the financial sector, financial documents such as earnings conference calls (ECCs) and investment reports often contain a wealth of critical information. Accurately identifying potential logical errors in financial documents is crucial for investment decision-making and market stability. Existing identification methods primarily rely on manual analysis or keyword-based recognition, which suffers from the following drawbacks: First, existing recognition methods can usually only capture the surface language patterns of financial texts and are unable to accurately identify deep contextual expressions and complex logical errors; Second, existing large language models require extensive manual configuration and maintenance to adapt to the specific analysis needs of financial texts when addressing specialized problem recognition, which is time-consuming, labor-intensive, and inefficient. 3. Existing large language models are prone to hallucinations when processing multi-source data, and often generate inaccurate or factually unfounded content; at the same time, existing large language models lack logical reasoning capabilities and are prone to reasoning jumps or logical breaks, resulting in one-sided or erroneous conclusions and a lack of credibility.

[0003] Therefore, we need to develop an AI-agent-based financial text logic evaluation method and system that can reduce dependence on manual labor, accurately identify deep contextual expressions and complex logical errors in financial texts, improve recognition efficiency, and quickly adapt to different financial text analysis needs. Summary of the Invention

[0004] The purpose of the present invention is to provide an AI Agent-based financial text logic evaluation method and system to solve the problems mentioned in the above background technology that the existing recognition methods rely on manual labor, have errors and omissions, and are difficult to accurately identify deep-seated logical errors.

[0005] To achieve the above object, the present invention adopts the following technical solutions: According to one aspect of the present invention, a method for evaluating the logic of financial text based on an AI agent is provided, the method comprising the following steps: Digitally convert and pre-process original financial texts, segment them according to logical structures, extract key information, and obtain structured standard data; Based on the training dataset in the financial field, we fine-tune and initialize the pre-trained large language model and build the reasoning agent and the reflection agent. Based on chain thinking prompts and a logic knowledge base, guiding the reasoning agent to analyze the structured standard data, identify logic loopholes and output logic loophole analysis results; The logic vulnerability analysis results are reviewed by the reflection agent, and the logic vulnerabilities are classified and integrated by the topic modeling technology to output a logic error report; The logical error report is visually displayed and exported in a custom format.

[0006] According to another aspect of the present invention, a financial text logic evaluation system based on AI Agent is provided, which includes: a data preprocessing module, a model initialization module, a logic vulnerability identification module, a logic vulnerability review module, and a visualization display module. The data preprocessing module is used to digitally convert and preprocess the original financial text, segment it according to the logical structure, and extract key information to obtain structured standard data; The model initialization module is used to fine-tune and initialize the pre-trained large language model based on the financial training dataset, and to build the reasoning agent and the reflection agent. The logic vulnerability identification module is used to guide the reasoning agent to analyze the structured standard data based on chain thinking prompts and a logic knowledge base, identify logic vulnerabilities, and output logic vulnerability analysis results; The logic vulnerability review module is used to review the logic vulnerability analysis results through the reflection agent, classify and integrate the logic vulnerabilities through topic modeling technology, and output a logic error report; The above-mentioned visual display module is used to visually display the logical error report and export it in a custom format.

[0007] Based on the above scheme, the structured standard data is obtained, specifically including: Digitally converting the original financial text in a non-digital format using OCR technology or a structured parsing tool to obtain first converted data; Cleaning and denoising the first conversion data to obtain second conversion data; Based on a text segmentation algorithm, segmenting the second converted data according to a logical structure and marking key information; Performing word segmentation on the segmented text based on specialized word segmentation rules in the financial field, and performing grammatical and semantic analysis, extracting the key information using named entity recognition technology, and constructing the structured standard data; The structured standard data is quality checked, and system parameters are tuned based on the check results.

[0008] Based on the above solution, the pre-trained large language model is fine-tuned and initialized as follows: Load a pre-trained large language model as the basic model of the system; wherein the pre-trained large language model includes but is not limited to any one of ChatGPT, Gemini, Claude, Llama, and DeepSeek R1 models; Constructing a training dataset for the financial field based on the structured standard data and a specialized database for the financial field, and fine-tuning the pre-trained large language model based on the training dataset for the financial field; According to the requirements of the financial text analysis task, the initial parameters of the fine-tuned large language model are set; wherein the initial parameters include the inference step size, judgment threshold, number of chain reasoning layers, and evaluation frequency of the reflection mechanism.

[0009] Based on the above solution, the financial field-specific database includes a financial field knowledge base and a logic knowledge base; the logic knowledge base includes at least one logic vulnerability classification of inconsistency, ambiguity, insufficient evidence, unreasonable assumptions and insufficient relevance.

[0010] Based on the above solution, the output logic error report specifically includes: The reflection agent detects each logic vulnerability description in the logic vulnerability analysis result to determine whether it meets the preset qualification standard; If it has been achieved, there is no need to improve it. The current logic vulnerability description is reviewed and the loop starts to detect the next logic vulnerability description; If it is not achieved, it needs to be improved. The reflection agent supplements the current logic vulnerability description with detailed description based on the preset correction algorithm and gives correction suggestions; Correct the logic vulnerability description that needs to be improved according to the correction suggestions, and use a cyclic detection mechanism to repeatedly verify the corrected logic vulnerability description until the logic vulnerability description meets the preset qualification standard and the review is completed; When all logic vulnerability descriptions in the logic vulnerability analysis results have been reviewed, the logic vulnerabilities are clustered using topic modeling technology, the classification results are integrated and the logic error report is output; wherein the topic modeling technology is SeedLDA topic modeling technology.

[0011] Based on the above solution, the visual display of the logic error report specifically includes: Parsing the logic error report based on a data parser, and classifying and summarizing various logic errors according to time, text location and error type; Highlight text paragraphs with logical loopholes; different types of logical errors are displayed in different colors; Use graphical display technology to generate statistical charts, including the overall situation of logical loopholes in financial texts, the distribution of various logical errors, and the frequency of occurrence of various logical errors; Supports dynamic interactive visualization, and you can view detailed error descriptions and correction suggestions by clicking on the highlighted area.

[0012] Based on the above solution, when exporting in custom formats, including but not limited to PDF, Excel, CSV, and JSON formats.

[0013] It can be seen from the above technical solutions that compared with the prior art, the present invention has at least the following advantages and positive effects: (1) The present invention performs a series of preprocessing on the original financial text, obtains structured standard data, performs data quality detection, and reversely tunes the system parameters based on the detection results. This not only reduces the dependence on manual labor and realizes system adaptive adjustment, but also improves the data quality and analysis accuracy in the subsequent reasoning and reflection stages.

[0014] (2) Based on chain thinking prompts and a logical knowledge base, the present invention guides the reasoning agent to analyze the financial text step by step, and reviews the results of the logical vulnerability analysis through the reflection agent. It can not only identify the potential logical errors in the financial text at multiple levels and steps, maximizing the capture of potential defects, but also ensure that the identification, description and correction of each logical error are reasonable and rigorous.

[0015] (3) The present invention uses SeedLDA topic modeling technology to perform cluster analysis on the identified logical vulnerabilities, which can deeply explore deeper logical errors in financial texts and meet the financial text analysis needs in different business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below only relate to some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 A flowchart of a method for logical evaluation of financial texts based on an AI agent provided by an embodiment of the present invention is shown; Figure 2 A flow chart of a method for converting original financial text into structured standard data provided by an embodiment of the present invention is shown; Figure 3 A flow chart of a method for reviewing logic vulnerability analysis results based on a reflection mechanism provided by an embodiment of the present invention is shown; Figure 4 The figure shows a schematic diagram of the structure of a financial text logic evaluation system based on AI Agent provided by an embodiment of the present invention; in, Figure 4 The reference numerals in the figures are described as follows: 400 - A financial text logic evaluation system based on AI Agent; 401 - Data preprocessing module, 4011 - Basic processing unit, 4012 - Parameter tuning unit; 402 - Model initialization module, 4021 - Model configuration unit, 4022 - Agent construction unit; 403 - Logical vulnerability identification module, 4031 - Reasoning analysis unit, 4032 - Error identification unit; 404 - Logical vulnerability review module, 4041 - Result review unit, 4042 - Classification and integration unit; 405 - Visualization display module, 4051 - Visualization display unit, 4052 - Custom export unit. DETAILED DESCRIPTION

[0018] In order to more clearly illustrate the purpose, technical solutions and advantages of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The example implementation methods can be implemented in various forms and should not be understood as being limited to the examples described herein. On the contrary, these implementation methods are provided to make the present invention more comprehensive and complete, and to fully convey the concepts of the example implementation methods to those skilled in the art.

[0019] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, it will be appreciated by those skilled in the art that the technical solutions of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0020] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0021] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0022] The present invention will be described in detail below with reference to specific embodiments: Example 1

[0023] Before describing each step of the present invention in detail, a brief introduction to the relevant technical terms involved in the technical solution of the present invention is given.

[0024] An AI agent (artificial intelligence agent) is an intelligent system that makes autonomous decisions and executes tasks. Typically based on a large language model, it can combine tools, environmental interactions, and feedback mechanisms to accomplish complex tasks. AI agents possess the ability to perceive, reason, and act, and can be used in applications such as automated analysis, intelligent question-answering, and information retrieval.

[0025] The Transformer is the core architecture of modern natural language processing (NLP). Compared to traditional RNNs and LSTMs, the Transformer abandons the recurrent structure and adopts a self-attention mechanism. By calculating the relevance of each word to other words, the model can understand long-range dependencies and improve text comprehension.

[0026] The Large Language Model (LLM) is an AI model based on deep learning and natural language processing that can understand, generate, and process complex text data. Its core technology relies on the Transformer architecture, which uses a self-attention mechanism to learn semantic relationships within text. It also combines large-scale pre-training and fine-tuning to enhance its adaptability to diverse tasks. The LLM is capable of performing a variety of tasks, including text generation, machine translation, automatic summarization, code generation, and dialogue systems. Compared to traditional methods, it outperforms traditional approaches in contextual understanding, reasoning, and information integration. Modern LLMs also employ reinforcement learning to optimize human-computer interaction and improve answer quality, and they integrate multimodal learning to expand into areas such as images and audio.

[0027] Chain of Thought (CoT) is a prompt engineering technology that improves the reasoning capabilities of large language models. It guides the model to gradually break down complex problems, simulating the human thought process. Traditional LLMs often generate only a single conclusion when answering complex questions, lacking transparent reasoning steps. CoT, however, enhances the arithmetic, common sense, and reasoning capabilities of large models by requiring the model to explicitly output the intermediate reasoning steps before outputting the final answer, resulting in a clearer reasoning process and more reliable results.

[0028] The Reflection Mechanism is an optimization method used to enhance the quality of LLM output. It allows the model to self-evaluate and adjust its own reasoning and conclusions. It guides the model to review its own reasoning steps and detect possible errors, redundancies, or contradictions, thereby performing self-correction.

[0029] Topic modeling is an unsupervised machine learning technique used to automatically discover and extract potential topics from large amounts of text data. It analyzes word frequency and co-occurrence patterns in text using statistical methods or deep learning algorithms to classify text content into distinct topic categories. For example, the Seeded Latent Dirichlet Allocation (SeedLDA) algorithm employed in this paper is an optimized topic modeling algorithm that uses predefined seed words for more precise topic clustering.

[0030] like Figure 1 As shown, the embodiment of the present invention provides a method for evaluating the logic of financial text based on AI Agent. The specific steps of the method are as follows: S1: Digitally convert and preprocess the original financial text, segment it according to the logical structure, extract key information, and obtain structured standard data; In this embodiment, the purpose of step S1 is to improve the data quality and analysis accuracy of the subsequent reasoning and reflection stages; specifically, Figure 2 As shown, the specific steps for obtaining structured standard data include: S201: Digitally converting the original financial text in a non-digital format using OCR technology or a structured parsing tool to obtain first converted data; In this embodiment, non-digital original financial documents include, but are not limited to, paper or PDF formats. If the original financial documents are already in electronic format, they can be directly imported into the system for subsequent pre-processing. During the digital conversion, high-precision OCR models such as Tesseract, Google Cloud Vision, and PaddleOCR can be used to ensure the integrity, accuracy, and readability of the converted text.

[0031] S202: Cleaning and denoising the first converted data to obtain second converted data; After the original financial text is digitized, it undergoes cleaning and noise reduction. This involves using regular expressions and string matching algorithms to perform a preliminary scan of the converted data, removing irrelevant symbols, garbled characters, extraneous spaces, advertising information, and other distracting content. To ensure clean and coherent text, a specific noise reduction algorithm is employed to uniformly process punctuation and special characters.

[0032] S203: Based on a text segmentation algorithm, segment the second converted data according to a logical structure and mark key information; Furthermore, based on the text segmentation algorithm, the second converted data is segmented according to the logical structure, including but not limited to dividing the long text into multiple sentences or paragraphs, and recursively segmenting the multi-level structure existing in the text to ensure that the information at each level accurately corresponds; further, the segmented text is annotated with key information, including but not limited to company name, time, spokesperson, questions and answers, numerical indicators and other information.

[0033] S204: performing word segmentation processing on the segmented text based on specialized word segmentation rules in the financial field, performing grammatical and semantic analysis, extracting the key information using named entity recognition technology, and constructing the structured standard data; Furthermore, targeting the specialized terminology and expressions in financial text, a word segmentation tool is used to segment the segmented text, ensuring the subsequent correct recognition of key data such as specialized terminology, abbreviations, and numbers. Combining a part-of-speech tagging algorithm, grammatical and semantic analysis is performed on each segmented word, providing basic grammatical information for subsequent logical reasoning. Furthermore, after completing word segmentation, grammatical, and semantic analysis, named entity recognition (NER) technology is applied to extract the key information annotated in step S203, forming structured standard data.

[0034] S205: Performing quality inspection on the structured standard data, and optimizing system parameters based on the inspection results.

[0035] Furthermore, the structured standard data is quality-checked to verify the integrity of key entities and structural information, ensuring no information is missed or misidentified. Simultaneously, the test results are fed back to the system's data preprocessing module for algorithm parameter tuning and rule optimization, enabling system adaptive adjustment.

[0036] The original financial text may be an unprocessed meeting record and may be modified according to actual conditions and specific requirements, which is not limited in this application.

[0037] S2: Based on the training dataset in the financial field, fine-tune and initialize the pre-trained large language model, and build the reasoning agent and the reflection agent; In this embodiment, the purpose of step S2 is to build an intelligent system suitable for financial text analysis, which can achieve efficient analysis and evaluation of the logic of financial text by loading a pre-trained large language model, domain fine-tuning, initializing parameter settings, and creating a dedicated agent.

[0038] Specifically, the financial training dataset is constructed using the structured standard data obtained in step S1 and a specialized knowledge base for the financial field. The specialized knowledge base for the financial field includes a financial knowledge base and a logical knowledge base. The financial knowledge base is used to enhance the LLM's understanding of financial terminology, market patterns, and financial analysis methods. Data sources can include financial textbooks, securities analysis reports, and financial news. The logical knowledge base is primarily used to enhance the LLM's logical reasoning ability. Data sources can include classical logic literature. The logical knowledge base includes at least one logical vulnerability classification from the following: inconsistency, ambiguity, insufficient evidence, unreasonable assumptions, and insufficient relevance.

[0039] Furthermore, the pre-trained large language model is fine-tuned based on the constructed training dataset in the financial field; wherein, the pre-trained large language model can adopt ChatGPT, Gemini, Claude, Llama, DeepSeek R1 or any other model under legal circumstances; further, after the model is fine-tuned, the initial parameters of the fine-tuned large language model are set according to the requirements of the financial text analysis task, including the inference step size, judgment threshold, number of chain reasoning layers and evaluation frequency of the reflection mechanism.

[0040] S3: Based on the chain thinking prompt and the logic knowledge base, guide the reasoning agent to analyze the structured standard data, identify logical loopholes and output the logical loophole analysis results; In this embodiment, the purpose of step S3 is to achieve multi-level and multi-step identification of potential logical errors in the text and to maximize the capture of potential defects.

[0041] For example, the chain thinking prompt may be: You are FinWise, an agent specializing in financial text analysis, possessing deep reasoning and self-reflection capabilities. Your task is to analyze the logic of questions and answers given by listed company management during earnings conference calls, assessing their completeness, accuracy, and potential risks, and conducting a systematic analysis based on your professional financial knowledge. You will employ chain reasoning, first gradually breaking down management's responses to determine whether they address the questions positively and provide sufficient logical support. You will then analyze their rationality based on industry context and historical data. You will then utilize reflection to self-verify your reasoning, identifying logical loopholes, redundant information, or potential misleading points, and optimizing your analysis. Finally, you will generate a structured logical error analysis report, including the question context, answer analysis, reasoning chain, reflection adjustments, and final conclusion. Ensure that the output is logically clear, rigorous, and credible, providing effective decision-making guidance for investors.

[0042] Specifically, the reasoning agent, supported by pre-set chain-thinking prompts and a logical knowledge base, gradually conducts a logical analysis of each sentence in the structured standard data. For example, consider a sentence like "Profits increased significantly this quarter, but revenue decreased." The reasoning agent's analysis involves first extracting the two pieces of information: "Profits increased significantly" and "revenue decreased." Next, it compares the conventional relationship between the two pieces of information to identify logical contradictions (normally, profit growth should be accompanied by revenue growth, but in this sentence, profit growth is accompanied by revenue decline, thus determining a logical contradiction). Finally, it matches the sentence with the logical vulnerability classification in the logical knowledge base (this sentence is inconsistent), confirming the presence of a logical vulnerability and the corresponding logical vulnerability category.

[0043] Furthermore, the logic vulnerability analysis results are output; wherein, the logic vulnerability analysis results are a detailed reasoning report, which includes the reasoning process of each step, the matched logic vulnerability category, and the preliminary error description and annotation. For example, the above analysis statement can be output as: {Logic Error ID: ERR001; Error Type: Inconsistency; Error Description: There is a contradiction between profit growth and revenue decline; Original Text Snippet: Profits increased significantly this quarter, but revenue declined.}

[0044] S4: reviewing the logic vulnerability analysis results through the reflection agent, classifying and integrating the logic vulnerabilities through topic modeling technology, and outputting a logic error report; In this embodiment, the purpose of step S4 is to perform a secondary review and optimization on the logic vulnerability analysis results generated by the reasoning agent to ensure that the identification, description and correction of each logic error are reasonable and rigorous.

[0045] Specifically, if Figure 3As shown, the logical vulnerability analysis results are reviewed based on the reflection mechanism, including the following steps: S301: The reflection agent detects each logic vulnerability description in the logic vulnerability analysis result to determine whether it meets the preset qualification standard; In this embodiment, the preset qualification criteria include but are not limited to: no semantic ambiguity or no information omission, etc.; specific qualification criteria can be preset according to actual conditions and specific needs, and the present invention does not impose any restrictions.

[0046] S302: If the requirements are not met, then improvement is required. The reflection agent provides a detailed description of the current logic vulnerability based on a preset correction algorithm and provides correction suggestions. Specifically, taking the logical vulnerability analysis results output in step S3 as an example, after review and correction by the reflection agent, the output can be: {Logical error ID: ERR001; Error type: Inconsistency; Detailed description: There is an obvious contradiction between "profits increased significantly" and "revenue decreased" in the original text. It is recommended to verify whether there is a one-time income impact or data recording error; Correction suggestion: Verify the profit structure and the reasons for the change in revenue; Original text fragment: Profits increased significantly this quarter, but revenue decreased.}

[0047] S303: Based on the correction suggestion, correct the current logic vulnerability description; S304: repeatedly verifying the corrected logic vulnerability description using a cyclic detection mechanism until the logic vulnerability description meets the preset qualification standard and the review is completed; S305: If the requirements have been met, no further improvement is required. The current logic vulnerability description is reviewed and the next logic vulnerability description is checked in a loop. S306: When all logic vulnerability descriptions in the logic vulnerability analysis results have been reviewed, the logic vulnerabilities are clustered using SeedLDA topic modeling technology, the classification results are integrated, and the logic error report is output.

[0048] S5: Visually display the logic error report and export it in a customized format.

[0049] In this embodiment, the purpose of step S5 is to intuitively display the logic error report output by the reflection agent so that the user can quickly identify, understand and use the analysis results.

[0050] Specifically, the logical error report is visually displayed, including: Parsing the logic error report based on a data parser, and classifying and summarizing various logic errors according to time, text location and error type; Highlight text paragraphs with logical loopholes; different types of logical errors are displayed in different colors; Use graphical display technology to generate statistical charts, including the overall situation of logical loopholes in financial texts, the distribution of various logical errors, and the frequency of occurrence of various logical errors; Supports dynamic interactive visualization, and you can view detailed error descriptions and correction suggestions by clicking on the highlighted area.

[0051] Furthermore, when exporting in custom formats, it includes but is not limited to PDF, Excel, CSV, and JSON formats.

[0052] The AI ​​Agent-based financial text logic evaluation method described in the embodiment of the present invention can not only reduce dependence on manual labor, achieve system adaptive adjustment, and improve data quality and analysis accuracy in subsequent reasoning and reflection stages; it can also perform multi-level and multi-step identification of potential logical errors in financial texts, deeply explore deeper logical errors in financial texts, and meet the financial text analysis needs in different business scenarios. Example 2

[0053] like Figure 4 As shown, an embodiment of the present invention provides an AI agent-based financial text logic evaluation system 400, which includes: a data preprocessing module 401, a model initialization module 402, a logic vulnerability identification module 403, a logic vulnerability review module 404, and a visualization display module 405; wherein: The data preprocessing module 401 is used to digitally convert and preprocess the original financial text, segment it according to the logical structure, and extract key information to obtain structured standard data; The data preprocessing module 401 includes: a basic processing unit 4011 and a parameter tuning unit 4012; wherein: The basic processing unit 4011 is configured to digitally convert and pre-process the original financial text, segment it according to the logical structure, extract key information, and obtain structured standard data. Specifically, the steps include: Digitally converting the original financial text in a non-digital format using OCR technology or a structured parsing tool to obtain first converted data; Cleaning and denoising the first conversion data to obtain second conversion data; Based on a text segmentation algorithm, segmenting the second converted data according to a logical structure and marking key information; Extracting the key information through specialized word segmentation and named entity recognition technologies in the financial field, and performing grammatical and semantic analysis on the words in the key information in combination with a part-of-speech tagging algorithm to construct the structured standard data; The parameter tuning unit 4012 is configured to: perform quality monitoring on the structured standard data obtained in the basic processing unit 4011, and perform algorithm parameter tuning and rule optimization on the data preprocessing module 401 based on the detection results to achieve system adaptive adjustment.

[0054] Model initialization module 402, used to fine-tune and initialize the pre-trained large language model based on the training dataset in the financial field, and build the reasoning agent and the reflection agent; The model initialization module 402 includes: a model configuration unit 4021 and an agent construction unit 4022; wherein: The model configuration unit 4021 is configured to: construct a financial domain training dataset, fine-tune the pre-trained large language model, and set initial parameters of the fine-tuned large language model; wherein the initial parameters include the inference step size, judgment threshold, number of chained inference layers, and evaluation frequency of the reflection mechanism; The Agent building unit 4022 is configured to build a reasoning agent and a reflection agent.

[0055] The logic vulnerability identification module 403 is used to guide the reasoning agent to analyze the structured standard data based on chain thinking prompts and a logic knowledge base, identify logic vulnerabilities, and output logic vulnerability analysis results; The logic vulnerability identification module 403 includes: an inference analysis unit 4031 and an error identification unit 4032; wherein: The reasoning analysis unit 4031 is configured to guide the reasoning agent to analyze structured standard data based on chain thinking prompts and a logical knowledge base, wherein the logical knowledge base includes at least one logical vulnerability classification of inconsistency, ambiguity, insufficient evidence, unreasonable assumption, and insufficient relevance; The error identification unit 4032 is configured to: identify logical loopholes in the text based on the analysis results of the reasoning analysis unit 4031 and output the logical loophole analysis results.

[0056] The logic vulnerability review module 404 is used to review the logic vulnerability analysis results through the reflection agent, classify and integrate the logic vulnerabilities through topic modeling technology, and output a logic error report; The logic vulnerability review module 404 includes: a result review unit 4041 and a classification integration unit 4042; wherein: The result review unit 4041 is configured to review the logic vulnerability analysis result output by the error identification unit 4032 through the reflection agent. Specifically, the steps include: The reflection agent detects each logic vulnerability description in the logic vulnerability analysis result to determine whether it meets the preset qualification standard; If it is not achieved, it needs to be improved. The reflection agent supplements the current logic vulnerability description with detailed description based on the preset correction algorithm and gives correction suggestions; Based on the correction suggestions, correct the current logic vulnerability description; A cyclic detection mechanism is used to repeatedly verify the corrected logic vulnerability description until the logic vulnerability description meets the preset qualification standard and the review is completed; If it has been achieved, there is no need to improve it. The current logic vulnerability description is reviewed and the loop starts to detect the next logic vulnerability description until all logic vulnerability descriptions in the logic vulnerability analysis results have been reviewed.

[0057] The classification integration unit 4042 is configured to: cluster the logic vulnerabilities based on the reviewed logic vulnerability descriptions using SeedLDA topic modeling technology, integrate the classification results and output a logic error report.

[0058] The visual display module 405 is used to visually display the logic error report and export it in a custom format.

[0059] The visualization display module 405 includes: a visualization display unit 4051 and a custom export unit 4052; wherein: The visualization display unit 4051 is configured to: visualize the logic error report output by the classification integration unit 4042, specifically including: Parsing the logic error report based on a data parser, and classifying and summarizing various logic errors according to time, text location and error type; Highlight text paragraphs with logical loopholes; different types of logical errors are displayed in different colors; Use graphical display technology to generate statistical charts, including the overall situation of logical loopholes in financial texts, the distribution of various logical errors, and the frequency of occurrence of various logical errors; Supports dynamic interactive visualization, and you can view detailed error descriptions and correction suggestions by clicking on the highlighted area.

[0060] The custom export unit 4052 is configured to export the logic error report in a custom format, including but not limited to PDF, Excel, CSV and JSON formats.

[0061] In this embodiment, the data preprocessing module 401 performs a series of preprocessing on the original financial text, obtains structured standard data, performs data quality detection, and reversely tunes the system parameters based on the detection results, which can reduce dependence on manual work and achieve system adaptive adjustment; the logical vulnerability identification module 403 guides the reasoning agent to analyze the financial text step by step based on chain thinking prompts and logical knowledge base, and conducts review through the logical vulnerability review module 404 and cluster analysis of the identified logical vulnerabilities through SeedLDA topic modeling technology, which can not only ensure that the identification, description and correction of each logical error are reasonable and rigorous, but also deeply explore deeper logical errors in financial texts to meet the financial text analysis needs in different business scenarios.

[0062] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present invention are indicated by the claims. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and that various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. The AI ​​Agent-based financial text logic evaluation method is characterized by: The steps include: Digitally convert and pre-process original financial texts, segment them according to logical structures, extract key information, and obtain structured standard data; Based on the training dataset in the financial field, we fine-tune and initialize the pre-trained large language model and build the reasoning agent and the reflection agent. Based on chain thinking prompts and a logic knowledge base, guiding the reasoning agent to analyze the structured standard data, identify logic loopholes and output logic loophole analysis results; The logic vulnerability analysis results are reviewed by the reflection agent, and the logic vulnerabilities are classified and integrated by the topic modeling technology to output a logic error report; The logical error report is visually displayed and exported in a custom format.

2. The AI ​​Agent-based financial text logic evaluation method according to claim 1 is characterized in that: The structured standard data obtained specifically includes: Digitally converting the original financial text in a non-digital format using OCR technology or a structured parsing tool to obtain first converted data; Cleaning and denoising the first conversion data to obtain second conversion data; Based on a text segmentation algorithm, segmenting the second converted data according to a logical structure and marking key information; Performing word segmentation on the segmented text based on specialized word segmentation rules in the financial field, and performing grammatical and semantic analysis, extracting the key information using named entity recognition technology, and constructing the structured standard data; The structured standard data is quality checked, and system parameters are tuned based on the check results.

3. The AI ​​Agent-based financial text logic evaluation method according to claim 1 is characterized in that: The pre-trained large language model is fine-tuned and initialized as follows: Load the pre-trained large language model as the basic model of the system; Constructing a training dataset for the financial field based on the structured standard data and a specialized database for the financial field, and fine-tuning the pre-trained large language model based on the training dataset for the financial field; According to the requirements of the financial text analysis task, set the initial parameters of the fine-tuned large language model.

4. The AI ​​Agent-based financial text logic evaluation method according to claim 3 is characterized in that: The pre-trained large language model adopts any one of ChatGPT, Gemini, Claude, Llama, and DeepSeek R1 models; the initial parameters include the inference step size, the judgment threshold, the number of chain reasoning layers, and the evaluation frequency of the reflection mechanism.

5. The AI ​​Agent-based financial text logic evaluation method according to claim 1 is characterized in that: The logic knowledge base includes at least one logic vulnerability classification of inconsistency, ambiguity, insufficient evidence, unreasonable assumption and insufficient relevance.

6. The AI ​​Agent-based financial text logic evaluation method according to claim 1 is characterized in that: The topic modeling technology is SeedLDA topic modeling technology.

7. The AI ​​Agent-based financial text logic evaluation method according to claim 1 is characterized in that: The output logic error report specifically includes: The reflection agent detects each logic vulnerability description in the logic vulnerability analysis result to determine whether it meets the preset qualification standard; If it has been achieved, there is no need to improve it. The current logic vulnerability description is reviewed and the loop starts to detect the next logic vulnerability description; If it is not achieved, it needs to be improved. The reflection agent supplements the current logic vulnerability description with detailed description based on the preset correction algorithm and gives correction suggestions; When all logic vulnerability descriptions in the logic vulnerability analysis results have been reviewed, the logic vulnerabilities are clustered using SeedLDA topic modeling technology, the classification results are integrated, and the logic error report is output.

8. The AI ​​Agent-based financial text logic evaluation method according to claim 7 is characterized in that: After correcting the logic vulnerability description that needs to be improved according to the correction suggestion given by the reflection agent, the corrected logic vulnerability description is verified using a cyclic detection mechanism until the logic vulnerability description meets the preset qualification standard and the review is completed.

9. AI Agent-based financial text logic evaluation system, characterized by: include: The data preprocessing module is used to digitize and preprocess the original financial text, segment it according to the logical structure, and extract key information to obtain structured standard data; The model initialization module is used to fine-tune and initialize the pre-trained large language model based on the training dataset in the financial field, and to build the reasoning agent and the reflection agent; A logic vulnerability identification module is used to guide the reasoning agent to analyze the structured standard data based on chain thinking prompts and a logic knowledge base, identify logic vulnerabilities, and output logic vulnerability analysis results; A logic vulnerability review module is used to review the logic vulnerability analysis results through the reflection agent, classify and integrate the logic vulnerabilities through topic modeling technology, and output a logic error report; The visualization display module is used to visualize the logic error report and export it in a custom format.

10. The AI ​​Agent-based financial text logic evaluation system according to claim 9, characterized in that: The structured standard data is quality tested, and the algorithm parameters and rules of the data preprocessing module are tuned based on the test results to achieve system adaptive adjustment.

Images