Intelligent reputation reduction evaluation system based on large language model

Abstract

The invention relates to the technical field of large model application, and discloses an intelligent reputation value reduction evaluation system based on a large language model, and the system comprises a data obtaining module which is used for obtaining an evaluation material; the model auxiliary module is used for carrying out primary review and recalculation in sequence according to the evaluation material to obtain a to-be-evaluated financial index as a recalculation manuscript, and generating a problem list according to the recalculation manuscript; the report generation module is used for generating an evaluation report according to the question list; wherein during primary examination, the model auxiliary module is used for identifying the effectiveness of the evaluation material and screening out the evaluation material meeting a preset condition; during recalculation, the model auxiliary module is used for extracting the financial indexes and the positions of the financial indexes in the original text to obtain the recalculation manuscript; according to the method, the automation and the interpretability are improved based on the large model, and meanwhile, the accurate extraction of the evaluation indexes under the specific structured data is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of large language model applications, and more particularly to an intelligent goodwill impairment assessment system based on a large language model. Background Art

[0002] Currently, goodwill impairment testing is a core method used in corporate financial management systems to measure changes in a company's true value following mergers and acquisitions. It is crucial for recognizing asset impairment losses and ensuring the fairness of financial statements. The testing process requires comprehensive consideration of multiple factors, including the future cash flows of the acquired assets, market fluctuations, and the setting of discount rates. The calculations involved are complex and often involve considerable subjective judgment.

[0003] Traditional goodwill impairment assessment mainly relies on manual analysis of assessment reports and tabular data, which is not only time-consuming and labor-intensive, but also highly subjective. In addition, faced with the complex financial structures and industry scenarios of different companies, it lacks a stable universal method to support it, making it prone to information omissions, misjudgments and other problems.

[0004] As large language models (LLMs) improve their capabilities in natural language understanding and tabular information processing, some research is attempting to transfer them to data analysis tasks in the accounting field. However, relying solely on the "semantic transfer" of large models to domain tasks still faces numerous challenges in scenarios such as goodwill impairment that require high precision and strong interpretability. For example, the tabular information required for evaluation has strict structural features such as sheet naming, row and column positioning, and financial indicators. Without the guidance of structural annotations, general models struggle to accurately extract it.

[0005] Therefore, how to deeply integrate the semantic understanding ability of large language models with the table structured features required for goodwill impairment assessment to build an intelligent assessment system with high-precision data extraction, automated cross-validation and strong explanatory report generation capabilities is a technical problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0006] In view of this, the present invention provides an intelligent goodwill impairment assessment system based on a large language model, which realizes collaborative assessment of preliminary review and re-review with the assistance of the model, improves automation and interpretability through preliminary review validity screening, recalculation draft generation, problem list generation, and assessment report output. At the same time, financial indicators and original text positioning are extracted in the recalculation stage, and the association annotation of structured information and text semantics is realized, which solves the challenge of structured features.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] An intelligent goodwill impairment assessment system based on a large language model, including

[0009] A data acquisition module, used to obtain evaluation materials;

[0010] A model auxiliary module is used to perform preliminary review and recalculation based on the evaluation materials, obtain the financial indicators to be evaluated as recalculation drafts, and generate a problem list based on the recalculation drafts; and generate an evaluation report based on the problem list;

[0011] Among them, during the initial review, the model auxiliary module is used to identify the validity of the evaluation materials and screen out the evaluation materials that meet the preset conditions; during the recalculation, the model auxiliary module is used to extract financial indicators and their positions in the original text to obtain the recalculation manuscript.

[0012] Preferably, the model auxiliary module includes a preliminary review submodule;

[0013] The preliminary review submodule realizes basic verification of the validity of the materials through the collaborative processing of the bidirectional coding language model and the large language model;

[0014] The bidirectional encoding language model is used to classify according to the tab page names and table context information in the evaluation materials, and to filter according to preset category conditions;

[0015] The large language model extracts key fields based on the table contents in the evaluation materials and performs screening according to preset corresponding field conditions.

[0016] Preferably, the large language auxiliary module includes a complex operator module;

[0017] The recalculation submodule locates the indicators of the tables in the evaluation materials through the fine-tuned large language model, obtains various evaluation indicators and corresponding tab pages and row numbers, and forms a standard structured list as the recalculation manuscript.

[0018] Preferably, the large language auxiliary module includes an evaluation submodule, and the evaluation submodule is used to generate an evaluation report based on the recalculation manuscript and preset evaluation prompts.

[0019] Preferably, it also includes a manual interactive review module, which obtains the recalculation manuscript of the model auxiliary module in the recalculation stage and the final evaluation report, and visualizes them.

[0020] Preferably, the manual interaction review module is further used to mark low-confidence results in the recalculation manuscript according to the analysis results of the model auxiliary module in the recalculation stage and to visualize them.

[0021] A large language model fine-tuning method for intelligent goodwill impairment assessment includes the following steps:

[0022] Construct a first special data set and fine-tune the preset model using the first special data set; use the fine-tuned model to conduct a preliminary review and screen out evaluation materials that meet specific conditions.

[0023] Preferably, in the first special data set, the tab page name and the table context are taken as a sample, and the semantic category corresponding to the tab page name is taken as the sample category label.

[0024] A large language model fine-tuning method for intelligent goodwill impairment assessment includes the following steps:

[0025] Construct a second special data set, and use the second special data set to fine-tune the preset model; use the fine-tuned model to recalculate, locate the original text of the evaluation indicators in the evaluation materials, and extract them to obtain the recalculation manuscript.

[0026] Preferably, the second special data set includes question-answer pairs extracted from financial indicators, in which the target indicators are used as question points, and the corresponding table positions and original text contents are used as standard answers.

[0027] It can be seen from the above technical solution that compared with the existing technology, the present invention discloses an intelligent goodwill impairment assessment system based on a large language model, which realizes the collaborative assessment of preliminary review and re-review based on the model assistance, improves automation and interpretability through preliminary review validity screening, recalculation draft generation, generation of problem lists, and output of assessment reports. At the same time, financial indicators and original text positioning are extracted in the recalculation stage, and the association annotation of structured information and text semantics is realized, which solves the challenge of structured features. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] 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 are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0029] Figure 1 A schematic diagram of the structure of an intelligent goodwill impairment assessment system based on a large language model provided in an embodiment of the present invention.

[0030] Figure 2 The figure is a schematic diagram of the evaluation process of an intelligent goodwill impairment evaluation system based on a large language model in an embodiment of the present invention.

[0031] Figure 3 Schematic diagram of the training method for fine-tuning the model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0033] Example 1

[0034] like Figure 1 and Figure 2 The embodiment of the present invention discloses an intelligent goodwill impairment assessment system based on a large language model, including a data acquisition module and a model auxiliary module.

[0035] The data acquisition module is used to obtain evaluation materials, such as financial statements. The reports have strict structural features such as sheet page naming, row and column positions, and financial indicators.

[0036] The model auxiliary module is used to conduct preliminary review and recalculation based on the evaluation materials in sequence, obtain the financial indicators to be evaluated as the recalculation draft, and generate a list of questions based on the recalculation draft; and is used to generate an evaluation report based on the list of questions.

[0037] Among them, during the initial review, the model auxiliary module is used to identify the validity of the evaluation materials and screen out the evaluation materials that meet the preset conditions; during the recalculation, the model auxiliary module is used to extract financial indicators and their position in the original text to obtain the recalculation draft.

[0038] In this embodiment, the present invention constructs an intelligent assessment system that covers the entire process from initial review to recalculation and problem list generation. By introducing hierarchical module division and human-computer collaboration, this process significantly improves assessment efficiency and automation while ensuring assessment accuracy and consistency.

[0039] To further implement the above technical solution, during the preliminary review phase, the model-assisted module utilizes a preliminary review submodule to effectively verify the assessment materials. Manually configured assessment task parameters, such as the data type and validity period, are required for the goodwill impairment assessment. The preliminary review submodule then conducts a preliminary assessment of the assessment materials and selects those that meet the assessment task parameters for subsequent evaluation.

[0040] Specifically, the preliminary review sub-module realizes basic verification of the validity of materials through the collaborative processing of the bidirectional coding language model and the large language model. After fine-tuning and training with a specific special data set, the bidirectional coding language model is used to judge the validity of the data type. It can identify the label page category of the report and determine whether it is a goodwill impairment-related indicator page; and the large language model extracts keywords based on RAG (Retrieval-Augmented Generation) and judges the fields corresponding to other evaluation parameters, such as base date judgment and discount rate matching. It automatically scans the report content through a coarse-grained search strategy, extracts the fields where the base date and discount rate are located, and performs matching judgment under rule constraints (such as whether it contains numbers).

[0041] Among them, when performing classification, the bidirectional encoding language model not only obtains the tab name, but also deeply considers the table context to achieve dynamic classification.

[0042] In addition, the coarse-grained search strategy can be based on pages. After completing field extraction and matching judgment, the page where the current field is located is output.

[0043] In order to further implement the above technical solution, in the recalculation stage, the model auxiliary module uses the recalculation sub-module to realize information extraction of effective evaluation materials and further extract specific evaluation indicators therein.

[0044] Specifically, the complex operator module loads a large language model that has been fine-tuned with the financial indicator question-and-answer dataset, locates and extracts the evaluation indicators, namely financial indicators, in the evaluation materials, obtains their line numbers in the original text through positioning, and finally generates a standard structured list.

[0045] When generating a structured list, the preset question template can be automatically loaded and input into the fine-tuned large language model, so that it can organize the data based on the recalculation results, for example: "Please extract all the project names and original texts about operating income on this page."

[0046] In this embodiment, the generated recalculation manuscript includes the financial indicators required for each assessment and their positions in the original text. This structured data list helps facilitate manual search and comparison during manual review.

[0047] Furthermore, the preset structured prompt template is input into the large language model to screen the extraction results. For extraction results with complex structures or insufficient model confidence, the system automatically marks them as "low-confidence tasks" and hands them over to manual review for additional confirmation, which greatly reduces the pressure of manual review and improves accuracy.

[0048] In order to further implement the above technical solution, the model auxiliary module uses the evaluation sub-module to realize the final auxiliary evaluation. Based on the recalculation draft, it combines predefined question prompts, evaluation reports and evaluation instructions to generate a list of questions for goodwill impairment evaluation under the predefined question prompts as evaluation considerations, and further gives preliminary responses to obtain an evaluation report.

[0049] In this embodiment, the evaluation submodule generates a list of questions in QA format based on the tabs, line numbers and key fields identified in the preliminary review and review stages, combined with a preset rule library, and calls a large language model to generate answers.

[0050] Furthermore, to prevent large language models from generating hallucinatory answers, the list of questions can be screened before generating answers, removing structure-bound question tasks (such as those requiring item judgment tasks that combine label pages, line numbers, and evaluation reports). These removed questions can be fed into a bidirectional encoding language model to generate corresponding answers.

[0051] During the manual review stage, humans only need to spot-check and correct low-confidence items in the list to further improve the accuracy of overall assessment quality control.

[0052] In order to further implement the above technical solution, the system also includes a manual interactive review module, which obtains the recalculation draft and final evaluation report of the model auxiliary module in the recalculation stage and visualizes them; it is also used to mark the low-confidence results in the recalculation draft according to the analysis results of the model auxiliary module in the recalculation stage and visualize them.

[0053] Finally, to support the system's engineering deployment requirements in high-concurrency scenarios, this paper designed an efficient task scheduling architecture based on multi-process and asynchronous queue mechanisms. Document processing tasks are parallelized at the process level, while checklist generation and problem reasoning are asynchronously scheduled at the thread level. This ensures high responsiveness and low latency even when running concurrently across multiple enterprises and projects.

[0054] In this embodiment,

[0055] The recalculation draft stores the values of various financial indicators previously extracted using the bidirectional coding language model and the large language model. The question list includes questions and answers to the questions, such as "What is the discount rate for 2023-2024?" and "Is the scope of the asset group in this report consistent with the scope of the asset group in the previous year's report?" Some questions are about the financial indicators in the recalculation draft, and some questions need to find answers from the appraisal report and appraisal explanation. The large language model is responsible for finding answers from the appraisal report, appraisal explanation, and recalculation draft. The appraisal report, appraisal explanation, and financial statements are the original materials, and the recalculation draft also extracts financial indicators from these. The predefined question prompts are mainly for the questions that need to be answered in the question list. They are questions written to ask the large language model. Because directly asking the large language model may not get a satisfactory answer, it is necessary to design question prompts. With the predefined question prompts and recalculation drafts, valuation reports, and valuation instructions, the large language model can organize the required material paragraphs from them, and the program reorganizes them into a completed goodwill impairment valuation report according to the rules. The rules here are mainly the structure, sequence, and contents of each part of the goodwill impairment valuation report. The rules are set (the large language model does not need to participate), and the large language model is mainly responsible for generating a question list (questions + answers).

[0056] Example 2

[0057] Based on the same inventive concept, an embodiment of the present invention discloses a large language model fine-tuning method for intelligent goodwill impairment assessment. This fine-tuning method can be used in the assessment system of Example 1, enabling it to achieve more accurate classification in the preliminary review stage, thereby accurately screening assessment materials that meet the assessment task conditions. The steps include:

[0058] Construct a first specialized dataset and fine-tune the pre-set model using it. Use the fine-tuned model for preliminary review, selecting evaluation materials that meet specific criteria. In the first specialized dataset, the tab name and table context are considered a sample, and the semantic category corresponding to the tab name is used as the sample category label.

[0059] In this embodiment, by collecting a large number of structurally differentiated financial statements (including annual reports, evaluation instructions, financial analysis tables, etc.), the semantic categories of each tab page are manually annotated. The annotation system classifies the tab page according to the header content, context text and table structure characteristics, such as "main business income", "administrative expenses", "depreciation and amortization", etc., and finally generates a multi-category tab page classification sample under a unified label set. This subset is used as the training input of the model, which can effectively improve the model's tab recognition ability in multi-source tables and provide a page-level entry judgment basis for subsequent indicator extraction.

[0060] In this embodiment, in response to the problems of diverse tabs and irregular naming in financial statements, the present invention designs a joint modeling dataset generation method. Traditional methods usually rely solely on keyword matching of tab names, which makes it difficult to cope with situations such as large differences in industry templates and ambiguous header semantics. To this end, the present invention uses the tab name and the first few rows of table content in the page as model inputs, and through the multi-channel joint encoding method of "label name + table context", the model can simultaneously perceive static labels and dynamic content semantics, and output classification results, such as "income category", "expense category" or "asset indicator category". This context-based modeling method breaks through the previous static classification logic, making the model more adaptable when dealing with complex and changeable corporate report structures.

[0061] Example 3

[0062] Based on the same inventive concept, an embodiment of the present invention discloses a large language model fine-tuning method for intelligent goodwill impairment assessment. This method can be used in the assessment system of Example 1 to locate and extract specific financial indicators during the recalculation phase. The steps include:

[0063] Construct a second special data set, and use the second special data set to fine-tune the preset model; use the fine-tuned model to recalculate, locate the original text of the evaluation indicators in the evaluation materials and extract them to obtain a recalculation draft; wherein, the second special data set includes question-answer pairs extracted from financial indicators, in which the target indicators are used as question points, and the corresponding table positions and original text content are used as standard answers.

[0064] For example, we construct questions based on target indicators in financial statements (e.g., "Please extract the row and textual content of the operating income item"), annotating the corresponding table row number and textual content as the standard answer. This question-answer data not only includes basic indicators (such as operating income, operating costs, and sales expenses), but also covers multiple summary items and structured group items. This enables the model to learn to recognize modified expressions and contextual variations of the target indicator under different templates, significantly improving extraction accuracy and robustness.

[0065] Furthermore, to ensure consistency and effectiveness of data training, the system performs a unified preprocessing process after dataset construction, including deduplication, missing data handling, field standardization, and format conversion. All samples are divided into training, validation, and test sets. This ensures sufficient generalization of the model during training and allows for quantitative evaluation of its structural understanding and data extraction performance during validation and testing.

[0066] In this embodiment, the present invention designs standardized question templates and answer structures so that the model can understand instructions such as "Please extract the original text lines of all main business income related items on this page" and return standard answers in the form of ['main business income', 'other business income'].

[0067] Compared to conventional keyword extraction or rule-based row number matching, this method combines task-driven questioning with semantic reasoning, enabling the model to dynamically adapt to various table layouts and accurately identify fine-grained information. Furthermore, the constructed dataset not only supports basic indicator extraction tasks but can also be extended to complex reasoning requirements such as summary item identification and inter-row dependency judgment, demonstrating excellent transfer and scalability.

[0068] Example 4

[0069] like Figure 3 Based on the same inventive concept, an embodiment of the present invention provides a model fine-tuning method for the bidirectional encoding model and large language model involved in the above implementation, focusing on solving the problems of pre-trained language models being susceptible to label noise interference and poor evaluation stability during vertical field fine-tuning, ensuring that the model performs more robustly and reliably in goodwill impairment tasks.

[0070] The overall process of this mechanism is divided into two stages: (1) Linear Probing (LP); (2) Cleaning + Full Fine-Tuning (FFT). To reduce the noise interference of the model, this embodiment obtains confidence through linear layer training and performs screening based on the confidence. In the second stage, the network is dynamically updated by clustering the prediction loss.

[0071] In the first stage, the system freezes the main parameters of the bidirectional encoding language model, trains its terminal linear layer to output confidence, and performs preliminary screening of samples based on the confidence threshold. Similarly, the large language model generates confidence under structured prompt words and performs preliminary screening of samples based on the confidence threshold. In turn, relatively poor samples are screened out, reducing the impact of noise labels on the feature classifier. This embodiment estimates the confidence of the training samples and generates an initial label confidence score. By adopting the generalized cross entropy (GCE) loss function, the impact of noise labels on the model learning process can be reduced, and the pre-trained feature extractor can be ensured not to be contaminated.

[0072] In the second phase, the system performs Gaussian mixture modeling (GMM) on the samples based on the prediction loss from the previous phase, cleans the training set, and extracts a clean subset for further training. A full-scale fine-tuning (FFT) is then performed on this subset, updating all model parameters and improving its adaptability to the target task distribution. This entire FFT process can be iteratively executed, and as model quality improves, the recognition rate of clean samples further increases.

[0073] In this embodiment, the model trained in the previous stage will give a loss for different samples during inference. If this loss is relatively large, it means that there is a difference between the label and the sample itself, that is, it may be a noise sample, so it needs to be eliminated. GMM is a clustering method. In a set, the loss calculated for each sample is evenly sampled in each class, and two Gaussian distributions can be clustered and fitted. One Gaussian distribution represents a cleaner sample with a relatively low loss, and the other represents a relatively high sample with possible noise. By clustering into two distributions, cleaner samples can be screened out and clean subsets can be extracted. It should be noted that the full fine-tuning here still adds a small amount of noise samples, because more than half of the noise samples will make the model worse, but about 10% of the noise samples will improve the model effect, making the model not easy to overfit and more robust.

[0074] In order to further implement the above technical solution, the present invention also designs a "label source marking + confidence level assessment" joint system on the data side, introduces credibility assessment identifiers in manual annotations or weakly supervised labels, and dynamically adjusts the loss weight during model training, guiding the model to focus more on high-quality sample semantics and improve training stability.

[0075] Specifically, when constructing a training set (such as the specialized dataset constructed in Example 2 or Example 3), the labels of each training sample are marked with their source to distinguish between manual annotation and model generation. Different weights are assigned to different source types to represent the credibility of different data sources. During the training process, when calculating the loss or gradient, the weights are applied according to the corresponding weights, and the final result is calculated.

[0076] Finally, during the model inference and system output phase, the present invention establishes a confidence threshold mechanism. When the confidence level of the model output falls below the set threshold, the system automatically transfers the task to manual review, thus preventing erroneous results from directly affecting the evaluation report and achieving system-level controllability of results and business stability.

[0077] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0078] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent goodwill impairment assessment system based on a large language model, characterized by: It includes a data acquisition module for acquiring assessment materials; A model auxiliary module is used to perform preliminary review and recalculation based on the evaluation materials, obtain the financial indicators to be evaluated as a recalculation draft, and generate a problem list based on the recalculation draft; for generating an assessment report based on the list of questions; Among them, during the initial review, the model auxiliary module is used to identify the validity of the evaluation materials and screen out the evaluation materials that meet the preset conditions; during the recalculation, the model auxiliary module is used to extract financial indicators and their positions in the original text to obtain the recalculation manuscript.

2. The intelligent goodwill impairment assessment system based on a large language model according to claim 1 is characterized in that: The model auxiliary module includes a preliminary review submodule; The preliminary review submodule realizes basic verification of the validity of the materials through the collaborative processing of the bidirectional coding language model and the large language model; The bidirectional encoding language model is used to classify according to the tab page names and table context information in the evaluation materials, and to filter according to preset category conditions; The large language model extracts key fields based on the table contents in the evaluation materials and performs screening according to preset corresponding field conditions.

3. The intelligent goodwill impairment assessment system based on a large language model according to claim 1 or 2, characterized in that: The large language auxiliary module includes a complex operator module; The recalculation submodule locates the indicators of the tables in the evaluation materials through the fine-tuned large language model, obtains various evaluation indicators and corresponding tab pages and row numbers, and forms a standard structured list as the recalculation manuscript.

4. The intelligent goodwill impairment assessment system based on a large language model according to claim 3 is characterized in that: The large language auxiliary module includes an evaluation submodule, and the evaluation submodule is used to generate an evaluation report based on the recalculation manuscript and preset evaluation prompts.

5. The intelligent goodwill impairment assessment system based on a large language model according to claim 1 is characterized in that: It also includes a manual interactive review module, which obtains the recalculation manuscript of the model auxiliary module in the recalculation stage and the final evaluation report, and visualizes them.

6. The intelligent goodwill impairment assessment system based on a large language model according to claim 5 is characterized in that: The manual interaction review module is also used to mark the low-confidence results in the recalculation manuscript according to the analysis results of the model auxiliary module in the recalculation stage and visualize them.

7. A large language model fine-tuning method for intelligent goodwill impairment assessment, characterized by: The method is applicable to the evaluation system according to any one of claims 1 to 6, and comprises the following steps: Construct a first special data set and fine-tune the preset model using the first special data set; use the fine-tuned model to conduct a preliminary review and screen out evaluation materials that meet specific conditions.

8. The large language model fine-tuning method for an intelligent goodwill impairment assessment system according to claim 7, characterized in that: In the first special data set, a tab page name and a table context are taken as a sample, and a semantic category corresponding to the tab page name is taken as a sample category label.

9. A large language model fine-tuning method for intelligent goodwill impairment assessment, characterized by: The method is applicable to the evaluation system according to any one of claims 1 to 6, and comprises the following steps: Construct a second special data set, and use the second special data set to fine-tune the preset model; use the fine-tuned model to recalculate, locate the original text of the evaluation indicators in the evaluation materials, and extract them to obtain the recalculation manuscript.

10. A large language model fine-tuning method for intelligent goodwill impairment assessment according to claim 9, characterized in that: The second special data set includes question-answer pairs extracted from financial indicators. In the question-answer pairs, the target indicators are used as question points, and the corresponding table positions and original text contents are used as standard answers.

Images