Report analysis method, device and electronic equipment

Abstract

The application relates to a report analysis method and device and electronic equipment, comprising: obtaining a report to be analyzed; converting the report to be analyzed into text information and layout information corresponding to the text information; determining a feature vector of the report to be analyzed in at least one dimension according to the text information and the layout information; obtaining at least one analysis strategy for analyzing the report to be analyzed; determining a strategy weight corresponding to the analysis strategy according to the feature vector; analyzing the report to be analyzed according to the analysis strategy and the strategy weight corresponding to the analysis strategy respectively to obtain a preliminary analysis result corresponding to each analysis strategy; and fusing the preliminary analysis results to obtain a target analysis result of the report to be analyzed. The application can automatically analyze a report such as a financial report that needs to be analyzed, and can assign a corresponding strategy weight to the analysis strategy according to different reports to adapt to the analysis of different reports, thereby improving the analysis efficiency and quality of the report.

Description

Technical Field

[0001] This application belongs to the field of data processing, specifically relating to a report parsing method, apparatus, and electronic device. Background Technology

[0002] In daily investment research, it is often necessary to study and analyze financial reports or other complex reports disclosed by complex financial groups and large corporations.

[0003] Taking financial reports as an example, annual / semi-annual reports are large and complex in format, with topics spanning multiple pages and diverse table structures. The report indicators are scattered and difficult to integrate. Manually extracting relevant indicators is time-consuming, prone to omissions, and not reusable. The data quality of OCR (Optical Character Recognition) indicator analysis tools on the market cannot be guaranteed. When researchers conduct financial report indicator analysis, they lack efficient automated tools, and relying solely on manual research of massive amounts of corporate financial report indicators is inefficient. Summary of the Invention

[0004] The purpose of this application is to provide a report parsing method, apparatus, and electronic device to overcome or at least partially solve the above-mentioned problems.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: A report parsing method, the method comprising: Retrieve the input report to be parsed; The report to be parsed is converted into text information and the layout information corresponding to the text information; Based on the text information and the layout information, determine the feature vector of the report to be parsed in at least one dimension; Obtain at least one parsing strategy for parsing the report to be parsed; The strategy weights corresponding to the parsing strategy are determined based on the feature vectors. The report to be parsed is parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, respectively, to obtain the preliminary parsing results corresponding to each parsing strategy; The target analysis result of the report to be analyzed is obtained by combining the preliminary analysis results.

[0006] In one embodiment of this application, the feature vectors of different dimensions include at least table density, text structure entropy, numerical density, format complexity, semantic relevance, and topic coherence; the parsing strategy includes at least table extraction strategy, regular expression matching strategy, and large language model semantic understanding strategy.

[0007] In one embodiment of this application, the step of parsing the report to be parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy to obtain the preliminary parsing result corresponding to each parsing strategy includes: The report parsing content options are provided to users through a human-computer interaction interface; the report parsing content options include at least the company name and indicator template, and the indicator template is used to indicate the indicator data that the parsing strategy needs to extract from the report to be parsed; In response to the user's operation on the report parsing content option, the target report parsing content is determined; The report to be parsed is parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, based on the parsing content of the target report, to obtain the preliminary parsing results corresponding to each parsing strategy.

[0008] In one embodiment of this application, determining the policy weight corresponding to the parsing policy based on the feature vector includes: Retrieve historical cases that match the feature vector from the historical knowledge base; the historical cases include at least the historical feature vector corresponding to the historical report, the strategy weight corresponding to the parsing strategy used when parsing the historical report, the historical target parsing result, and the historical quality assessment result; The feature vector and the historical cases are input into the Bayesian inference engine to obtain the policy weights corresponding to the parsing strategies. The Bayesian inference engine is used to determine the prior probability of each parsing strategy determined by the historical cases, to determine the likelihood probability of each parsing strategy based on the matching degree between the feature vector and the historical feature vector of the historical cases, to determine the posterior probability of each parsing strategy based on the prior probability and the likelihood probability, to determine the preliminary policy weights based on the posterior probabilities, to normalize the preliminary policy weights, and to obtain the policy weights corresponding to the parsing strategies based on the normalized posterior probabilities and the preset exploration factor.

[0009] In one embodiment of this application, the preliminary analysis result has a corresponding confidence level, and the process of fusing the preliminary analysis result to obtain the target analysis result of the report to be analyzed includes: When there is a conflict in the preliminary analysis results, the preliminary analysis results are verified according to the preset verification rules to obtain the semantic ambiguity. The comprehensive score corresponding to the preliminary analysis result is determined based on the strategy weight, the confidence level, and the semantic ambiguity level corresponding to the preliminary analysis result. The target analysis result of the report to be analyzed is obtained by combining the comprehensive score with the preliminary analysis result.

[0010] In one embodiment of this application, obtaining the target analysis result of the report to be analyzed by fusing the preliminary analysis result with the comprehensive score includes: If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is less than a preset threshold, then the conflicting preliminary analysis results are averaged to obtain an average preliminary analysis result, and the average preliminary analysis result is merged to obtain the target analysis result of the report to be analyzed. If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is greater than or equal to a preset threshold, then the preliminary analysis result with the highest comprehensive score is merged to obtain the target analysis result of the report to be analyzed.

[0011] In one embodiment of this application, after fusing the preliminary analysis results to obtain the target analysis result of the report to be analyzed, the method further includes: The target parsing results are subjected to hierarchical quality verification, and the quality assessment result of the target parsing results is determined based on the results of the hierarchical quality verification; the hierarchical quality verification includes at least data integrity verification, semantic consistency verification, and logical cross-validation. If the quality assessment result does not meet the expected quality requirements, a self-iterative optimization is triggered to generate a new target parsing result for the report to be parsed; the self-iterative optimization includes at least adjusting the weights of the parsing strategy; When the self-iterative optimization meets the preset termination condition, the self-iterative optimization is stopped; wherein, the preset termination condition includes at least the number of iterations of the self-iterative optimization reaching the maximum number of iterations, or the quality assessment result of the new target analysis result of the report to be parsed meeting the expected quality requirements; The strategy weights that meet the expected quality requirements, the target parsing results, and the quality assessment results are saved as cases to the historical knowledge base.

[0012] A report parsing apparatus, the apparatus comprising: The module for obtaining the report to be parsed is used to obtain the input report to be parsed. The report parsing module is used to convert the report to be parsed into text information and layout information corresponding to the text information; The feature vector determination module is used to determine the feature vector of the report to be parsed in at least one dimension based on the text information and the layout information. A parsing strategy determination module is used to obtain at least one parsing strategy for parsing the report to be parsed; The strategy weight determination module is used to determine the strategy weight corresponding to the parsing strategy based on the feature vector. The preliminary analysis result determination module is used to parse the report to be parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, respectively, to obtain the preliminary analysis result corresponding to each parsing strategy; The target parsing result determination module is used to integrate the preliminary parsing results to obtain the target parsing result of the report to be parsed.

[0013] An electronic device includes: a processor; and a memory for storing processor-executable instructions. The processor is configured to execute the instructions to implement the report parsing method described above.

[0014] A computer-readable storage medium, when the instructions in the storage medium are executed by the processor of a mobile terminal, enables the mobile terminal to perform the above-described report parsing method.

[0015] The embodiments of this application have at least the following beneficial effects: In this embodiment, an input report to be parsed is obtained, and the report is converted into text information and corresponding layout information. Based on the text information and layout information, a feature vector in at least one dimension of the report to be parsed is determined. At least one parsing strategy for parsing the report is obtained, and the strategy weights corresponding to the parsing strategies are determined based on the feature vectors. Subsequently, the report to be parsed is parsed according to the parsing strategy and its corresponding strategy weight, respectively, to obtain preliminary parsing results for each parsing strategy. Finally, the preliminary parsing results are merged to obtain the target parsing result for the report to be parsed. This embodiment can automatically parse reports such as financial reports that require parsing, and can assign corresponding strategy weights to different reports to adapt to the parsing of different reports. Thus, it can improve both the efficiency and quality of report parsing. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a report parsing method provided in an embodiment of this application; Figure 2 This is a schematic diagram of a system architecture provided in an embodiment of this application; Figure 3 This is a schematic diagram of a business process provided in an embodiment of this application; Figure 4 This is a schematic diagram of a data flow provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a report parsing device provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] The embodiments of this application will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be understood that the preferred embodiments are only for illustrating this application and are not intended to limit the scope of protection of this application.

[0018] It should be noted that the embodiments of this application may involve the use of user data. In practical applications, user-specific personal data may be used in the scheme described herein within the scope permitted by applicable laws and regulations, provided that it complies with the applicable laws and regulations of the country (e.g., with the user's explicit consent, with the user being properly notified, etc.).

[0019] Currently, the main report indicator analysis solutions available in the market and industry for financial reports or other complex reports include the following: (1) Rule-based template extraction scheme: The extraction rules and location templates for each type of financial report are predefined and implemented through a large number of if-else (conditional judgment) rules and regular expressions. The drawbacks of this approach are that it can only handle known formats, new document structures need to be redeveloped, maintenance costs are high, rule conflicts are frequent, it is difficult to handle format fine-tuning, and its adaptability is not strong.

[0020] (2) OCR and keyword search solutions: Converting financial reports in PDF (Portable Document Format) format into images and using OCR technology for recognition and keyword matching has drawbacks. It is easy to lose the document structure information of the financial report, cannot understand the table relationships, and the error rate of recognition using OCR technology is easy to accumulate, with an accuracy rate of less than 70%. It is also difficult to handle cross-page content and often interprets information out of context.

[0021] (3) Extraction using deep learning algorithms: Using pre-trained models such as BERT (Bidirectional Encoder Representations from Transformers, a pre-trained language model based on Transformer encoders) for named entity recognition, and using fine-tuned pre-trained models to identify financial indicators, this approach has the drawbacks of requiring a large amount of labeled training data for model training, having poor generalization ability, failing when the industry or report type changes, and having a black box model that cannot be explained or debugged.

[0022] (4) RPA (Robotic Process Automation) solution: The solution involves recording the manual operation process and using RPA tools such as UiPath to record the extraction process of indicators from financial reports. However, this approach is extremely fragile, failing even with minor interface adjustments. It also cannot handle new formats and lacks intelligent extraction capabilities. Essentially, this solution is still based on rules and does not achieve a truly effective extraction solution.

[0023] It is evident that most report indicator parsing solutions currently on the market are static and cannot adapt to new formats, exhibiting poor adaptability. To address the aforementioned issues, the technical solution of this application proposes a report parsing method that adaptively matches and extracts reports with structured, semi-structured, and unstructured data. Furthermore, based on a memory-driven self-iterative mechanism, the extraction results are automatically verified and their quality is evaluated, enabling continuous learning and optimization. This solves the problem of traditional algorithms lacking learning capabilities when extracting different types of reports.

[0024] Reference Figure 1 The diagram illustrates a flowchart of a report parsing method provided in an embodiment of this application, which specifically includes the following steps: Step 101: Obtain the input report to be parsed.

[0025] Step 102: Convert the report to be parsed into text information and the layout information corresponding to the text information.

[0026] In its specific implementation, the embodiment of this application uses a general account investment asset allocation system (hereinafter referred to as the system). The system can serve investment-related researchers to complete in-depth corporate research and analysis needs, such as asset allocation research, listed company financial analysis, corporate research analysis, peer business analysis, industry trend research, corporate financial report analysis, and financial due diligence of investment targets. In addition, it can also be applied to the analysis of other types of corporate announcements, audit reports, regulatory reports, statistical bulletins, and policy document indicators. Users of other platforms or sections who want to analyze financial reports and other reports (documents) can use the system through page references or API (Application Programming Interface) calls.

[0027] In this embodiment of the application, when investment-related researchers need to parse reports such as financial reports, they can input the report to be parsed into the system. The report to be parsed can be a PDF document. After receiving the report to be parsed, the system will convert the report into processable text information and layout information corresponding to the text information. For example, the system's parser will identify the text information and layout information in the report to be parsed, such as text blocks, tables, fonts, positions and other elements.

[0028] Step 103: Based on the text information and the layout information, determine the feature vector of the report to be parsed in at least one dimension.

[0029] In this embodiment of the application, after the system extracts text information and layout information from the report to be parsed, it can determine the feature vector of the report to be parsed in at least one dimension based on the text information and layout information. The feature vectors of different dimensions can include at least table density, text structure entropy, numerical density, format complexity, semantic relevance and topic coherence. The feature vectors of different dimensions can be used to describe the structural and content characteristics of the report to be parsed.

[0030] Specifically, feature vectors of different dimensions can be calculated in the following way: (1) Table density = Σ(table area) / Σ(page area); (2) Text structure entropy = -Σ(p_i×log(p_i)), where p_i is the probability of format i; (3) Numeric density = Number of numeric characters / Total number of characters; (4) Format complexity = number of unique formats / number of text blocks; (5) Semantic relevance = co-occurrence frequency of keywords; (6) Thematic coherence = number of cross-page citations / total number of pages.

[0031] Step 104: Obtain at least one parsing strategy for parsing the report to be parsed.

[0032] In specific implementation, the system is pre-set with a variety of parsing strategies (extraction strategies). These parsing strategies may include at least table extraction strategies, regular expression matching strategies, and large language model semantic understanding strategies (LLM (Large Language Model) semantic understanding strategies). These parsing strategies can work together, or they can work independently as needed. This application embodiment does not impose any restrictions on this.

[0033] Specifically, the table extraction strategy focuses on extracting data from the tables in the report to be parsed, intelligently identifying the table structure and target columns. The regular expression matching strategy uses a predefined pattern library to perform text matching in the report to be parsed and leverages context to enhance matching accuracy. The LLM semantic understanding strategy uses a large language model to understand the semantics of the text in the report to be parsed and extracts metric values.

[0034] For example, the parsing process of each parsing strategy for the parsing report can be as follows: (1) Table extraction strategy: detect table → parse structure → find corresponding year data column → extract value; (2) Regular expression matching strategy: apply pattern library → text matching → context expansion → unit recognition; (3) LLM semantic understanding strategy: build prompts → call API → parse response → verify result.

[0035] Step 105: Determine the strategy weight corresponding to the parsing strategy based on the feature vector.

[0036] In this embodiment of the application, the system can determine the strategy weight of each parsing strategy based on multi-dimensional feature vectors. The strategy weight is used to determine the allocation of computing resources and the order of calling the parsing strategy when parsing the report. For example, the parsing strategy with a higher strategy weight can be executed first and can be used to process the core data area in the report to be parsed. The parsing strategy with a lower strategy weight can be executed after the parsing strategy with a higher strategy weight has been executed, in order to supplement coverage or cross-validation, etc.

[0037] Step 106: Parse the report to be parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy to obtain the preliminary parsing results corresponding to each parsing strategy.

[0038] Step 107: Combine the preliminary analysis results to obtain the target analysis result of the report to be analyzed.

[0039] In this embodiment of the application, after determining the strategy weights of each parsing strategy, the system can parse the report to be parsed according to the parsing strategy and the strategy weights corresponding to the parsing strategy, thereby obtaining the preliminary parsing results of each parsing strategy. Subsequently, the preliminary parsing results can be merged to obtain the target parsing result of the report to be parsed, thus completing the parsing of the report. Subsequently, investment-related researchers (users) can use the target parsing result of the report to be parsed to conduct in-depth research and analysis.

[0040] In the above report parsing method, the input report to be parsed is obtained, and the report to be parsed is converted into text information and corresponding layout information. Based on the text information and layout information, the feature vector of the report to be parsed in at least one dimension is determined. At least one parsing strategy for parsing the report to be parsed is obtained, and the strategy weight corresponding to the parsing strategy is determined based on the feature vector. Subsequently, the report to be parsed can be parsed according to the parsing strategy and its corresponding strategy weight, respectively, to obtain preliminary parsing results for each parsing strategy. Finally, the preliminary parsing results are merged to obtain the target parsing result of the report to be parsed. This embodiment of the application can automatically parse reports that require parsing, such as financial reports, and can assign corresponding strategy weights to different parsing strategies to adapt to the parsing of different reports. Thus, it can improve both the efficiency and quality of report parsing.

[0041] In one embodiment of this application, the step of parsing the report to be parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy to obtain the preliminary parsing result corresponding to each parsing strategy may include: The report parsing content options are provided to users through a human-computer interaction interface; the report parsing content options include at least the company name and indicator template, and the indicator template is used to indicate the indicator data that the parsing strategy needs to extract from the report to be parsed; In response to the user's operation on the report parsing content option, the target report parsing content is determined; The report to be parsed is parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, based on the parsing content of the target report, to obtain the preliminary parsing results corresponding to each parsing strategy.

[0042] In this embodiment, the system can provide a human-computer interaction interface for the user. This interface includes options for reporting and parsing content. Specifically, these options may include at least the company name and indicator templates. The company name refers to the name of the company the user wants to analyze, and the indicator templates refer to the key indicator names, indicator classification definitions, and other information to be extracted from the report, listed in a table format. It should be noted that the system can pre-provide some company names and indicator templates for the user to choose from, or the user can input them themselves; this embodiment does not impose any restrictions on this.

[0043] Users can perform operations such as selection and input on the human-computer interaction interface to select report parsing options. In response to the user's operation, the system can determine the target report parsing content. Subsequently, based on the target report parsing content, such as the target company name and target indicator template, the system can quickly query and parse all the indicator data required by the target company name and target indicator template from complex and cumbersome financial reports or other reports to be parsed, which may be hundreds of pages long, as preliminary parsing results. In addition, it can also provide the theme of the indicator, the meaning and scope of the indicator, the values ​​for different years, the units of the values, and indicate the page number of the original text where the indicator is located, the original text segment, the search method, and the large model test results of the indicator, so as to facilitate the user to verify the preliminary parsing results.

[0044] In the above embodiments, the system allows users to determine the target report parsing content, which is used to parse out the indicator data and other data corresponding to the report to be parsed as preliminary parsing results, so that the target parsing results of the report to be parsed obtained in the end meet the user's needs and improve the user experience.

[0045] In one embodiment of this application, determining the policy weight corresponding to the parsing policy based on the feature vector may include: Retrieve historical cases that match the feature vector from the historical knowledge base; the historical cases include at least the historical feature vector corresponding to the historical report, the strategy weight corresponding to the parsing strategy used when parsing the historical report, the historical target parsing result, and the historical quality assessment result; The feature vector and the historical cases are input into the Bayesian inference engine to obtain the policy weights corresponding to the parsing strategies. The Bayesian inference engine is used to determine the prior probability of each parsing strategy determined by the historical cases, to determine the likelihood probability of each parsing strategy based on the matching degree between the feature vector and the historical feature vector of the historical cases, to determine the posterior probability of each parsing strategy based on the prior probability and the likelihood probability, to determine the preliminary policy weights based on the posterior probabilities, to normalize the preliminary policy weights, and to obtain the policy weights corresponding to the parsing strategies based on the normalized posterior probabilities and the preset exploration factor.

[0046] In this embodiment of the application, a pre-trained Bayesian inference engine can be deployed or invoked in the system. In addition, a historical knowledge base can be maintained in the system. The historical knowledge base can include high-quality historical cases. Specifically, historical cases can include historical feature vectors corresponding to historical reports (reports that the system has parsed in the past), strategy weights corresponding to the parsing strategy used when parsing historical reports, historical target parsing results, and historical quality evaluation results.

[0047] The system utilizes a historical knowledge base and feature vectors of the reports to be parsed, and performs Bayesian inference through a Bayesian inference engine to calculate the policy weights corresponding to parsing strategies such as table extraction, regular expression matching, and LLM semantic understanding. The Bayesian inference engine will output the policy weights corresponding to each parsing strategy to guide the subsequent parallel extraction of the reports to be parsed.

[0048] Specifically, the process of Bayesian inference performed by the Bayesian inference engine can be described as follows: (1) Similar case retrieval: Search the historical feature vectors in the historical knowledge base that are close to the feature vector of the report to be parsed, and take the historical case corresponding to the historical feature vector as the matching historical case; if there are many matching historical cases, the historical cases ranked in the TOP-10 (top 10) closest to the matching historical cases can be selected. (2) Calculation of posterior probability: P(strategy i|F) = P(F|strategy i) × P(strategy i) / Σ(P(F|strategy j) × P(strategy j)) Here, policy i and policy j refer to any two analytical policies, P(policy i|F) represents the posterior probability, P(policy i) and P(policy j) represent the prior probability, P(F|policy i) and P(F|policy j) represent the likelihood probability, and F represents the feature vector.

[0049] (3) After determining the initial policy weights based on the posterior probability, normalize the initial policy weights: W=[w_table,w_regex,w_llm], Σ(w_i)=1 Where W represents the set of initial strategy weights for all parsing strategies, w_table is the initial strategy weight for table extraction strategies, w_regex is the initial strategy weight for regular expression matching strategies, and w_llm is the initial strategy weight for LLM semantic understanding strategies.

[0050] (4) Exploration factor addition: W'=(1-ε)×W+ε×[1 / 3,1 / 3,1 / 3], ε=0.1 Where ε represents the exploration factor, and W' represents the set of final strategy weights obtained by applying the exploration factor to the initial strategy weights of all analytical strategies.

[0051] (5) Output: The strategy weight W' corresponding to each parsing strategy. The strategy weight W' is then passed to the system's multi-strategy extraction engine to parse the report according to the parsing strategy and the strategy weight corresponding to the parsing strategy.

[0052] In the above embodiments, the system can combine successful historical cases from the historical knowledge base with the multi-dimensional feature vectors of the report to be parsed to determine the strategy weights corresponding to the parsing strategy. In this way, the parsing results obtained based on the parsing strategy and strategy weights will be more accurate, and users can make more accurate decisions based on the parsing results.

[0053] In one embodiment of this application, the preliminary analysis result has a corresponding confidence level, and the process of fusing the preliminary analysis result to obtain the target analysis result of the report to be analyzed may include: When there is a conflict in the preliminary analysis results, the preliminary analysis results are verified according to the preset verification rules to obtain the semantic ambiguity. The comprehensive score corresponding to the preliminary analysis result is determined based on the strategy weight, the confidence level, and the semantic ambiguity level corresponding to the preliminary analysis result. The target analysis result of the report to be analyzed is obtained by combining the comprehensive score with the preliminary analysis result.

[0054] In the specific implementation, the final target parsing result can be obtained by fusing the preliminary parsing results extracted by multiple parsing strategies. During the fusion process, the strategy weight of each parsing strategy and the confidence level of each preliminary parsing result are considered. If there is a conflict between the preliminary parsing results of different parsing strategies, the system will resolve the conflict according to preset verification rules such as consistency check, source credibility and context verification, and generate the final target parsing result.

[0055] Specifically, the system can verify the preliminary parsing results according to preset verification rules to obtain the semantic ambiguity. Then, it can determine the comprehensive score corresponding to the preliminary parsing results based on the strategy weight, confidence level, and semantic ambiguity. The higher the comprehensive score, the more reliable the preliminary parsing results are. Then, the preliminary parsing results can be fused according to the comprehensive score to obtain the final target parsing result of the report to be parsed. For example, when there is a conflict, the preliminary parsing result with the highest comprehensive score can be selected for fusion to obtain the target parsing result.

[0056] For example, the formula for calculating the overall score can be as follows: Score(r)=w_strategy×confidence(r)×(1-ambiguity(r)) Final = argmax(Score(r_i)) Where r refers to the preliminary parsing result extracted based on the parsing strategy; w_strategy refers to the global reliability weight (strategy weight) of the parsing strategy corresponding to the preliminary parsing result r in the current document environment, calculated by the Bayesian inference engine; confidence(r) refers to the internal certainty score (i.e., confidence) of the preliminary parsing result. Different parsing strategies have their own corresponding confidence calculation algorithms. For example, the table extraction strategy can calculate the confidence based on the matching accuracy between the table header and the target indicator and the alignment of the data column with the required year. The regular expression matching strategy can calculate the confidence based on the matching pattern and the year and unit of the keywords in the context. The LLM semantic understanding strategy... (The confidence score of the internally generated results can be directly output). 1-ambiguity(r) is used as semantic clarity, where ambiguity(r) refers to semantic ambiguity. For example, when there are negative expressions such as "does not contain" or multiple years around the numerical value, the semantic ambiguity score increases, while the semantic clarity decreases. Semantic clarity is used to exclude "high-risk" preliminary parsing results that, although matched, have ambiguous context and are prone to misinterpretation. Final represents the preliminary parsing result corresponding to the highest comprehensive score. argmax is a function that evaluates the parameters (set) of a function. Score(r_i) represents the comprehensive score corresponding to the i-th preliminary parsing result.

[0057] In the above embodiments, when there are conflicts in the preliminary parsing results obtained according to different parsing strategies, the comprehensive score corresponding to the preliminary parsing results can be determined based on multiple data such as the strategy weight, confidence level, and semantic ambiguity level of the preliminary parsing results to fuse the preliminary parsing results and obtain the target parsing result. This can avoid the deviation from reality that can be easily achieved by relying on a single data point and ensure the accuracy of the fused target parsing result.

[0058] In one embodiment of this application, obtaining the target analysis result of the report to be analyzed by fusing the preliminary analysis result with the comprehensive score may include: If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is less than a preset threshold, then the conflicting preliminary analysis results are averaged to obtain an average preliminary analysis result, and the average preliminary analysis result is merged to obtain the target analysis result of the report to be analyzed. If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is greater than or equal to a preset threshold, then the preliminary analysis result with the highest comprehensive score is merged to obtain the target analysis result of the report to be analyzed.

[0059] In this embodiment of the application, after determining the comprehensive score corresponding to each preliminary analysis result, the difference between the preliminary analysis result with the highest comprehensive score and other preliminary analysis results can be compared. If the difference is less than a preset threshold, for example, the relative difference between the values ​​extracted for the same indicator is less than 5%, it indicates that the numerical difference between the indicators of the preliminary analysis result with the highest comprehensive score and other preliminary analysis results is small. In this case, the conflicting preliminary analysis results can be averaged to obtain an average preliminary analysis result. Then, the average preliminary analysis result is fused to obtain the target analysis result of the report to be analyzed. If the difference is greater than or equal to a preset threshold, for example, the relative difference between the values ​​extracted for the same indicator is greater than or equal to 5%, the preliminary analysis result with the highest comprehensive score is fused to obtain the target analysis result of the report to be analyzed.

[0060] In the above embodiments, the solution to the conflict can be determined based on the difference between the preliminary analysis result with the highest comprehensive score and other preliminary analysis results, which ensures the accuracy of the final target analysis result and conforms to the actual data patterns and market rules.

[0061] In one embodiment of this application, after fusing the preliminary analysis results to obtain the target analysis result of the report to be analyzed, the method may further include: The target parsing results are subjected to hierarchical quality verification, and the quality assessment result of the target parsing results is determined based on the results of the hierarchical quality verification; the hierarchical quality verification includes at least data integrity verification, semantic consistency verification, and logical cross-validation. If the quality assessment result does not meet the expected quality requirements, a self-iterative optimization is triggered to generate a new target parsing result for the report to be parsed; the self-iterative optimization includes at least adjusting the weights of the parsing strategy; When the self-iterative optimization meets the preset termination condition, the self-iterative optimization is stopped; wherein, the preset termination condition includes at least the number of iterations of the self-iterative optimization reaching the maximum number of iterations, or the quality assessment result of the new target analysis result of the report to be parsed meeting the expected quality requirements; The strategy weights that meet the expected quality requirements, the target parsing results, and the quality assessment results are saved as cases to the historical knowledge base.

[0062] In this embodiment of the application, the fused target parsing results are subjected to hierarchical quality verification to obtain the quality assessment results of the target parsing results. The hierarchical quality verification may include at least data integrity verification, semantic consistency verification and logical cross-validation. Specifically, data integrity verification is used to ensure the integrity of necessary fields, semantic consistency verification is used to ensure the standardization of indicator names and the rationality of numerical values, logical cross-validation is used to check the logical relationship through multi-source comparison, and the quality assessment results are used to evaluate the parsing quality of the parsing report to be parsed according to the parsing strategy and strategy weight.

[0063] The system performs knowledge synthesis based on the quality assessment results. If the quality assessment results do not meet the expected quality requirements, such as a matching rate below 90% or the presence of high-value missing indicators, it triggers self-iterative optimization. This self-iterative optimization adjusts parameters such as confidence thresholds and strategy weights, and re-parses the report to be parsed based on the parsing strategy and strategy weights to obtain new target parsing results.

[0064] Specifically, when the system decides to continue its iterative optimization, it diagnoses the quality shortcomings of the previous round (completeness, accuracy, coverage, etc.) and adjusts the hyperparameters for the next round accordingly. For example, if the completeness is low, the system considers the current parsing effort insufficient and needs to expand the search scope. It will increase the number of target metrics for a certain topic by 30%, forcing the regular expression matching strategy and the LLM semantic understanding strategy to mine more fine-grained metrics in these areas. If the accuracy (one of the metrics for quality assessment results) does not meet the expected quality requirements, it means that there are few high-confidence parsing results. The system can choose to lower the threshold to expand the candidate pool by lowering the confidence threshold of the candidate pool, for example, by 0.15, with a minimum of 0.6. Alternatively, if the coverage of the target topic is low (one of the metrics for quality verification results), the system will identify weak topics (e.g., topics with less than 10 extracted metrics) and focus on them in the next round of parsing, while increasing the strategy weight of the LLM semantic understanding strategy, etc. This is because LLM is good at handling unstructured or sparse data.

[0065] When the self-iterative optimization meets the preset termination conditions, such as the number of iterations of the self-iterative optimization reaching the maximum number of iterations, or the quality assessment result of the new target parsing result of the report to be parsed meets the expected quality requirements, the self-iterative optimization can be stopped. Subsequently, the strategy weights, target parsing results and quality assessment results that meet the expected quality requirements can be saved as cases to the historical knowledge base for the parsing of other reports in the future.

[0066] In the above embodiments, the system can perform hierarchical quality verification on the target parsing results. When the hierarchical quality verification does not meet the expected quality requirements, it can trigger self-iterative optimization to generate new target parsing results. In this way, it can be ensured that the final target parsing results of the report to be parsed meet the expected quality requirements, so that users can conduct research based on accurate target parsing results, thereby ensuring the accuracy of research results.

[0067] In summary, the embodiments of this application solve the problem that current static rule-based report index analysis schemes cannot adapt to new formats. By using Bayesian inference to achieve dynamic parsing strategy selection without the need for preset rules, it also solves the problem of single strategy. It proposes a multi-strategy collaborative algorithm that combines a structured extraction strategy (table extraction strategy) + a regular expression matching strategy + an LLM semantic understanding strategy. This algorithm can adaptively match parsing strategies for reports including structured, semi-structured, and unstructured data. Furthermore, based on a memory-driven self-iterative mechanism, it automatically verifies and evaluates the quality of the target parsing results, enabling continuous learning and optimization. This addresses the problem that traditional algorithms lack learning capabilities when extracting reports from different types of documents.

[0068] Reference Figure 2 This is a schematic diagram of a system architecture provided in an embodiment of this application. The system may include an input layer, a processing layer, and an output layer, specifically: Input layer: Used for inputting data such as PDF document streams (reports), target indicator templates, and historical knowledge bases; Processing layer and output layer: used to process input data and output; input data processing and output can be divided into the following stages: Phase 1: Deep Structure Analysis (Deep Structure Analysis and Topic Identification): The system extracts spatial features (position / distance / alignment), text features (paragraphs / sentences / vocabularies), and visual features (paragraphs / sentences / vocabularies) from the input data, thereby extracting text information and layout information. It then performs structural recognition on the text information and layout information, such as topic boundary detection, hierarchical structure construction, and content classification and labeling. Finally, it obtains feature vectors of different dimensions based on the structural recognition results.

[0069] Phase 2: Multi-strategy parallel extraction (adaptive multi-strategy extraction): Content type analysis and strategy matching are performed on feature vectors of different dimensions to determine the strategy weights of table extraction strategy, regular expression matching strategy, and LLM semantic understanding strategy, which are used to parse the report and obtain preliminary analysis results. Phase 3: Cross-validation and evaluation (stratified quality validation): The preliminary analysis results are subjected to hierarchical quality verification (quality assessment) including numerical reasonableness verification, contextual consistency verification, temporal logic verification, and cross-reference verification. The confidence level is calculated to obtain a comprehensive score, namely the accuracy score and the completeness score, which are used for quality assessment. Phase 4: Knowledge Synthesis and Optimization (Knowledge Synthesis and Self-Iteration) The initial analysis results are deduplicated to obtain structured indicator data. This process involves updating the historical knowledge base, identifying missing items, and adaptively optimizing the parameters of strategies such as regular expression matching. New target analysis results are then generated based on the optimized strategies until the target is met (the expected quality requirements are satisfied). Once the target is met, the structured indicator data, i.e., the final output target analysis result, is obtained. If the target is not met, the process can return to Phase 2.

[0070] Reference Figure 3 This application provides a schematic diagram of a business process, specifically, the business processing flow at each stage may include: Phase 1: Deep structural analysis and topic recognition. The system uses a document parser to parse PDFs and extract text, a layout analyzer to divide regions and understand layouts, a feature extractor to perform six-dimensional feature analysis and numerical calculations, and a topic recognizer to perform hierarchical recognition and topic clustering, resulting in a multi-dimensional feature vector F{f1,f2,f3,f4,f5,f6}.

[0071] Phase 2: Adaptive Multi-Strategy Extraction: The Bayesian decision engine (Bayesian inference engine) performs historical knowledge base queries, strategy weight calculations, weight normalization, exploration factor additions, and posterior probability updates to obtain the strategy weight vector (strategy weight) W; preliminary analysis results are obtained based on table extraction, regular expression matching, LLM semantic understanding (LLM semantic extraction) strategies, and strategy weight parsing reports; the preliminary analysis results are then subjected to result fusion and conflict resolution processing, which may include weighted fusion, result selection, and result deduplication to obtain the target analysis result.

[0072] Phase 3: Layered Quality Verification: The target parsing results are subjected to layered quality verification, including completeness verification (necessary fields, format specifications, unit conversion), consistency verification (indicator standardization, numerical rationality, logical relationships), cross-validation (multi-source comparison, time series, correlation indicators), and confidence assessment (confidence score, source assessment, confidence selection). Based on the results of the layered quality verification, a quality score is given to determine whether self-iterative optimization is triggered. Phase 4: Knowledge Synthesis and Self-Iteration: The final target analysis results are aggregated (intelligent deduplication, source merging, result output), updated (pattern recording, experience storage, knowledge transfer), iterated (iteration conditions, parameter adjustment, learning rate control), and evaluated (quality indicators, effect evaluation, performance evaluation) for the final target analysis results.

[0073] Reference Figure 4 This is a schematic diagram of a data flow provided in an embodiment of this application, which may specifically include the following steps: Step 1: Input and Document Parsing & Structure Analysis: The system receives PDF documents of enterprise reports as input and converts them into processable text and layout information. The parser identifies elements such as text blocks, tables, fonts, and positions within the document.

[0074] Step 2: Six-dimensional feature extraction: Extract six-dimensional feature vectors from the parsed PDF document, including table density, text structure entropy, numerical density, format complexity, semantic relevance, and topic coherence. These feature vectors describe the document's structural and content characteristics.

[0075] Step 3: Bayesian Decision Engine: Utilizing the historical knowledge base and the feature vectors of the current PDF document, the Bayesian decision engine calculates the policy weights of three parsing strategies (table extraction, regular expression matching, and LLM semantic understanding) through Bayesian inference. The Bayesian decision engine outputs the policy weights of each parsing strategy to guide subsequent parallel extraction of indicator data.

[0076] Step 4: Parallel Extraction with Multiple Strategies: Based on the strategy weights assigned by the Bayesian decision engine, three parsing strategies are activated simultaneously: Table Extraction Strategy: Focuses on extracting data from tables, intelligently identifying table structure and target columns. Regular Expression Matching Strategy: Uses a predefined pattern library for text matching and leverages context to enhance matching accuracy. LLM Semantic Understanding Strategy: Understands text semantics through a large language model and extracts indicator values.

[0077] Step 5: Result Fusion and Conflict Resolution. The preliminary parsing results extracted by the three parsing strategies are fused. During the fusion process, the strategy weight of each parsing strategy and the confidence level of each preliminary parsing result are considered. If there are conflicts in the preliminary parsing results of different strategies, the system will resolve the conflicts according to rules such as consistency checks, source credibility, and context verification, generating the final target parsing result.

[0078] Step 6: Layered quality verification: Perform three-layer verification on the fused target parsing results: data integrity verification (ensuring the completeness of necessary fields), semantic consistency verification (standardization of indicator names and reasonableness of values), and cross-validation (through multi-source comparison and logical relationship checks). The verification results are used to evaluate the quality of the extracted target parsing results.

[0079] Step 7: Knowledge Synthesis and Self-Iteration: Based on the results of hierarchical quality verification, the system performs knowledge synthesis. If the quality does not meet the expected requirements, such as a matching rate below 90% or the presence of high-value missing indicators, self-iterative optimization is triggered. Self-iterative optimization adjusts parameters, such as confidence thresholds and strategy weights, and re-executes the extraction process, for example, starting from step four or five. The self-iterative optimization continues until termination conditions are met, such as reaching the maximum number of iterations or achieving the required quality.

[0080] Step 8: Output and Knowledge Update: Output the final structured result (objective parsing result). Simultaneously, the feature vectors, policy weights, objective parsing results, and quality assessment results based on the large model are saved as new cases in the historical knowledge base for use in subsequent task decisions. Specifically, the new cases will be stored in the large model as part of the historical knowledge base content for reporting and parsing private domain tasks.

[0081] Step 9: Closed-Loop Feedback: The processing results and knowledge updates of the entire system form a closed loop. The next time a similar document is processed, the system will use information from the historical knowledge base to make more accurate decisions, thus continuously evolving.

[0082] The data flow described above ensures that the system can adapt to report documents with different structures, continuously improve extraction accuracy through multi-strategy collaboration and self-iterative optimization, and achieve continuous learning through knowledge accumulation.

[0083] To enable those skilled in the art to better understand the embodiments of this application, specific examples are used for illustration below.

[0084] Example 1: An investment research analyst needs to analyze a company's 2024 annual report (425 pages in total), extract 168 core financial indicators for peer analysis, and also extract all indicators from the financial report.

[0085] 1) First, the investment research analyst uploads a PDF file through the web interface. The system parses out 425 pages of content and identifies 1235 text blocks, 89 tables, and 156 charts. 2) The system completes feature vector calculations for different dimensions: Table density: 0.68 (high, financial data is concentrated in the table); Text structure entropy: 3.2 (medium, with some formatting variations); Numerical density: 0.42 (high, large amount of financial figures); Format complexity: 0.35 (medium); Semantic relevance: 0.78 (high, concentrated financial terminology); Thematic coherence: 0.85 (high, clear chapter division).

[0086] 3) The system completes the strategy weight calculation: The system finds successful historical cases from the company's 2023 annual report in the historical knowledge base: Bayesian inference yields the optimal policy weights: {Table: 0.65, Regularization: 0.25, LLM: 0.10}. Due to the high table density, the table extraction strategy dominates.

[0087] 4) First round of extraction: Table extraction strategy + regular expression matching strategy + LLM semantic understanding strategy.

[0088] 5) Iterative optimization: The system found 35 missing indicators, including "ceded reinsurance contract assets". The confidence threshold was lowered from 0.8 to 0.7, and the search scope was expanded to two adjacent pages. In the second round, 17 indicators were successfully added, with a matching rate of 93.6%. In the third round, 9 indicators were successfully added, with a final matching rate of 95.6%.

[0089] 6) Output of target analysis results: Generate an Excel report containing 823 deduplicated and intelligently integrated full indicators, displayed in separate sheets according to different themes, as well as 156 core indicators extracted according to the indicator template requirements. Each indicator is labeled with: its theme, meaning and scope, values ​​for different years, unit of measurement, page number of the original text, the relevant excerpt, search method, and large model test results. Eight indicators that were not found are labeled as "not disclosed in the financial report". Total time taken: 93 seconds.

[0090] Example 2: The Risk Management Department needs to extract risk indicators from the Q1 2024 reports of 5 listed insurance companies within one day for industry risk assessment.

[0091] 1) Batch task creation: Submit 5 PDF documents (reports) simultaneously using the batch upload tool, selecting the "Insurance Industry Risk Indicator Template" (45 indicators); 2) Parallel processing (the system automatically allocates 5 processing threads): Company 1: 87 pages, matching rate 95.6%; Company Name 2: 393 pages, matching rate 97.8%; Company 3: 76 pages, matching rate 93.3%; Company 4: 69 pages, matching rate 93.1%; Company 5: 82 pages, matching rate 94.4%; 3. Knowledge transfer effect: When processing the first company (Company 1), a general strategy (general parsing strategy and strategy weights) was used. When processing Company 2, the successful pattern of Company 1 was reused, resulting in a 15% speed improvement. Subsequently, by the time of the fifth company (Company 5), the system had accumulated industry characteristics, and the first-round matching rate reached 94%+.

[0092] 4. Summary Output: Generate industry comparison reports to facilitate cross-industry analysis.

[0093] Total time: 25 minutes (5 working days required for manual processing).

[0094] In the above example, users such as investment researchers can be freed from tedious document parsing and analysis. There is no need for repeated training for new types of cross-industry unstructured reports, and they can be quickly adapted and extracted. It supports flexible adjustment of analysis framework indicator templates. When new indicators need to be added, there is no need to redevelop the entire system. Only the indicator template needs to be updated. The system can quickly adapt through self-iteration and build an efficient automated enterprise report indicator parsing system. It can rapidly increase the number of enterprise analysis and tracking by several times, forming unique industry knowledge assets. Users can focus on in-depth analysis and value judgment, which greatly enhances the sensitivity of investment strategy response.

[0095] This application addresses the challenges faced by investment research teams at the business level when manually extracting relevant indicators for peer analysis, industry trend research, financial report analysis, and due diligence of investment targets. These challenges include time-consuming manual extraction of indicators, potential omissions, lack of reusability, inconsistent data quality, and a lack of efficient automation tools. Specifically, this application provides a self-iterative multi-analysis strategy indicator retrieval algorithm engine based on Bayesian inference. This algorithm engine consists of multiple dynamic parsing strategies, including dynamically fused table extraction strategies, regular expression matching strategies, and LLM semantic understanding strategies. It can intelligently adjust the combination of extraction strategies based on document characteristics without manual intervention. Furthermore, it introduces an agent-like approach driven by memory, featuring self-evaluation, multi-strategy selection, and closed-loop optimization. With five rounds of self-iteration, it eliminates the need for pre-training and debugging for specific report templates, unlike traditional deep learning algorithms. The system learns and evolves after each use, continuously improving extraction accuracy. Ultimately, the embodiments of this application can form a verifiable, reusable, and scalable dataset of enterprise / report metrics, assisting researchers in implementing more comprehensive and in-depth asset allocation strategies and enterprise investment value mining, thereby ultimately improving investment returns.

[0096] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of this application.

[0097] Reference Figure 5 The diagram shows a structural block diagram of a report parsing device provided in an embodiment of this application. The device may specifically include the following modules: The report acquisition module 501 is used to acquire the input report to be parsed; The report parsing module 502 is used to convert the report to be parsed into text information and layout information corresponding to the text information; The feature vector determination module 503 is used to determine the feature vector of the report to be parsed in at least one dimension based on the text information and the layout information. The parsing strategy determination module 504 is used to obtain at least one parsing strategy for parsing the report to be parsed; The strategy weight determination module 505 is used to determine the strategy weight corresponding to the parsing strategy based on the feature vector. The preliminary analysis result determination module 506 is used to parse the report to be parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, respectively, to obtain the preliminary analysis result corresponding to each parsing strategy; The target parsing result determination module 507 is used to fuse the preliminary parsing results to obtain the target parsing result of the report to be parsed.

[0098] In one embodiment of this application, the preliminary analysis result determination module 506 is used for: The report parsing content options are provided to users through a human-computer interaction interface; the report parsing content options include at least the company name and indicator template, and the indicator template is used to indicate the indicator data that the parsing strategy needs to extract from the report to be parsed; In response to the user's operation on the report parsing content option, the target report parsing content is determined; The report to be parsed is parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, based on the parsing content of the target report, to obtain the preliminary parsing results corresponding to each parsing strategy.

[0099] In one embodiment of this application, the strategy weight determination module 505 is used for: Retrieve historical cases that match the feature vector from the historical knowledge base; the historical cases include at least the historical feature vector corresponding to the historical report, the strategy weight corresponding to the parsing strategy used when parsing the historical report, the historical target parsing result, and the historical quality assessment result; The feature vector and the historical cases are input into the Bayesian inference engine to obtain the policy weights corresponding to the parsing strategies. The Bayesian inference engine is used to determine the prior probability of each parsing strategy determined by the historical cases, to determine the likelihood probability of each parsing strategy based on the matching degree between the feature vector and the historical feature vector of the historical cases, to determine the posterior probability of each parsing strategy based on the prior probability and the likelihood probability, to determine the preliminary policy weights based on the posterior probabilities, to normalize the preliminary policy weights, and to obtain the policy weights corresponding to the parsing strategies based on the normalized posterior probabilities and the preset exploration factor.

[0100] In one embodiment of this application, the target parsing result determination module 507: When there is a conflict in the preliminary analysis results, the preliminary analysis results are verified according to the preset verification rules to obtain the semantic ambiguity. The comprehensive score corresponding to the preliminary analysis result is determined based on the strategy weight, the confidence level, and the semantic ambiguity level corresponding to the preliminary analysis result. The target analysis result of the report to be analyzed is obtained by combining the comprehensive score with the preliminary analysis result.

[0101] In one embodiment of this application, the target parsing result determination module 507: If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is less than a preset threshold, then the conflicting preliminary analysis results are averaged to obtain an average preliminary analysis result, and the average preliminary analysis result is merged to obtain the target analysis result of the report to be analyzed. If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is greater than or equal to a preset threshold, then the preliminary analysis result with the highest comprehensive score is merged to obtain the target analysis result of the report to be analyzed.

[0102] In one embodiment of this application, the apparatus further includes: a stratified quality verification module, used for: The target parsing results are subjected to hierarchical quality verification, and the quality assessment result of the target parsing results is determined based on the results of the hierarchical quality verification; the hierarchical quality verification includes at least data integrity verification, semantic consistency verification, and logical cross-validation. If the quality assessment result does not meet the expected quality requirements, a self-iterative optimization is triggered to generate a new target parsing result for the report to be parsed; the self-iterative optimization includes at least adjusting the weights of the parsing strategy; When the self-iterative optimization meets the preset termination condition, the self-iterative optimization is stopped; wherein, the preset termination condition includes at least the number of iterations of the self-iterative optimization reaching the maximum number of iterations, or the quality assessment result of the new target analysis result of the report to be parsed meeting the expected quality requirements; The strategy weights that meet the expected quality requirements, the target parsing results, and the quality assessment results are saved as cases to the historical knowledge base.

[0103] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0104] This application also provides an electronic device, such as... Figure 6 As shown, it includes a processor 1001, a device interface 1002, a memory 1003, and a bus 1004; Memory 1003 is used to store computer programs; The processor 1001 executes the above steps when executing the program stored in the memory 1003.

[0105] The bus mentioned in the above terminal can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0106] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0107] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0108] This application also provides a storage medium that, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to execute the report parsing method of the foregoing embodiments.

[0109] The algorithms and displays provided herein are not inherently related to any particular computer, virtual device, or other equipment. The structure required to construct such a device is obvious from the above description. Furthermore, this application is not directed to any particular programming language. It should be understood that the content of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of this application.

[0110] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0111] Similarly, it should be understood that, in order to simplify this application and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of this application, various features of this application are sometimes grouped together into a single embodiment, figure, or description thereof. However, this method of disclosure should not be construed as reflecting an intention that the claimed application requires more features than are expressly recited in each claim. Rather, as reflected in the following claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0112] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0113] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components in the sequencing device according to this application. This application can also be implemented as a device or apparatus program for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can take the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0114] It should be noted that the above embodiments are illustrative of this application and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0116] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

[0117] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0118] It should be noted that the various data-related processes in the embodiments of this application are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

Claims

1. A report parsing method, characterized in that, The method includes: Retrieve the input report to be parsed; The report to be parsed is converted into text information and the layout information corresponding to the text information; Based on the text information and the layout information, determine the feature vector of the report to be parsed in at least one dimension; Obtain at least one parsing strategy for parsing the report to be parsed; The strategy weights corresponding to the parsing strategy are determined based on the feature vectors. The report to be parsed is parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, respectively, to obtain the preliminary parsing results corresponding to each parsing strategy; The target analysis result of the report to be analyzed is obtained by combining the preliminary analysis results.

2. The method according to claim 1, characterized in that, The feature vectors of different dimensions include at least table density, text structure entropy, numerical density, format complexity, semantic relevance, and topic coherence; the parsing strategy includes at least table extraction strategy, regular expression matching strategy, and large language model semantic understanding strategy.

3. The method according to claim 2, characterized in that, The step involves parsing the report to be parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, respectively, to obtain preliminary parsing results corresponding to each parsing strategy, including: The report parsing content options are provided to users through a human-computer interaction interface; the report parsing content options include at least the company name and indicator template, and the indicator template is used to indicate the indicator data that the parsing strategy needs to extract from the report to be parsed; In response to the user's operation on the report parsing content option, the target report parsing content is determined; The report to be parsed is parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, based on the parsing content of the target report, to obtain the preliminary parsing results corresponding to each parsing strategy.

4. The method according to claim 1, characterized in that, Determining the policy weights corresponding to the parsing policy based on the feature vector includes: Retrieve historical cases that match the feature vector from the historical knowledge base; the historical cases include at least the historical feature vector corresponding to the historical report, the strategy weight corresponding to the parsing strategy used when parsing the historical report, the historical target parsing result, and the historical quality assessment result; The feature vector and the historical cases are input into the Bayesian inference engine to obtain the policy weights corresponding to the parsing strategies. The Bayesian inference engine is used to determine the prior probability of each parsing strategy determined by the historical cases, to determine the likelihood probability of each parsing strategy based on the matching degree between the feature vector and the historical feature vector of the historical cases, to determine the posterior probability of each parsing strategy based on the prior probability and the likelihood probability, to determine the preliminary policy weights based on the posterior probabilities, to normalize the preliminary policy weights, and to obtain the policy weights corresponding to the parsing strategies based on the normalized posterior probabilities and the preset exploration factor.

5. The method according to claim 1, characterized in that, The preliminary analysis results have corresponding confidence levels, and the process of fusing the preliminary analysis results to obtain the target analysis result for the report to be analyzed includes: When there is a conflict in the preliminary analysis results, the preliminary analysis results are verified according to the preset verification rules to obtain the semantic ambiguity. The comprehensive score corresponding to the preliminary analysis result is determined based on the strategy weight, the confidence level, and the semantic ambiguity level corresponding to the preliminary analysis result. The target analysis result of the report to be analyzed is obtained by combining the comprehensive score with the preliminary analysis result.

6. The method according to claim 5, characterized in that, The process of obtaining the target analysis result of the report to be analyzed by integrating the comprehensive score with the preliminary analysis result includes: If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is less than a preset threshold, then the conflicting preliminary analysis results are averaged to obtain an average preliminary analysis result, and the average preliminary analysis result is merged to obtain the target analysis result of the report to be analyzed. If the difference between the preliminary analysis result with the highest comprehensive score and the other preliminary analysis results is greater than or equal to a preset threshold, then the preliminary analysis result with the highest comprehensive score is merged to obtain the target analysis result of the report to be analyzed.

7. The method according to claim 1, characterized in that, After fusing the preliminary analysis results to obtain the target analysis result of the report to be analyzed, the method further includes: The target parsing results are subjected to hierarchical quality verification, and the quality assessment result of the target parsing results is determined based on the results of the hierarchical quality verification; the hierarchical quality verification includes at least data integrity verification, semantic consistency verification, and logical cross-verification; If the quality assessment result does not meet the expected quality requirements, a self-iterative optimization is triggered to generate a new target parsing result for the report to be parsed; the self-iterative optimization includes at least adjusting the weights of the parsing strategy; When the self-iterative optimization meets the preset termination condition, the self-iterative optimization is stopped; wherein, the preset termination condition includes at least the number of iterations of the self-iterative optimization reaching the maximum number of iterations, or the quality assessment result of the new target analysis result of the report to be parsed meeting the expected quality requirements; The strategy weights that meet the expected quality requirements, the target parsing results, and the quality assessment results are saved as cases to the historical knowledge base.

8. A report parsing device, characterized in that, The device includes: The module for obtaining the report to be parsed is used to obtain the input report to be parsed. The report parsing module is used to convert the report to be parsed into text information and layout information corresponding to the text information; The feature vector determination module is used to determine the feature vector of the report to be parsed in at least one dimension based on the text information and the layout information. A parsing strategy determination module is used to obtain at least one parsing strategy for parsing the report to be parsed; The strategy weight determination module is used to determine the strategy weight corresponding to the parsing strategy based on the feature vector. The preliminary analysis result determination module is used to parse the report to be parsed according to the parsing strategy and the strategy weight corresponding to the parsing strategy, respectively, to obtain the preliminary analysis result corresponding to each parsing strategy; The target parsing result determination module is used to integrate the preliminary parsing results to obtain the target parsing result of the report to be parsed.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to execute the instructions to implement the report parsing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the mobile terminal, the mobile terminal is able to perform the report parsing method as described in any one of claims 1 to 7.

Images