Financial question and answer long text analysis method and system based on role recognition and dialogue structure analysis

Abstract

This invention discloses a long-text analysis method and system for financial question-and-answer based on role recognition and dialogue structure parsing, relating to the field of financial text processing. This invention constructs a role mapping table based on the opening information of the meeting transcript itself, supports dynamic updates, and accurately identifies multi-turn dialogues and complex interaction structures based on a dual mechanism of vector similarity and large language model understanding, ensuring the semantic integrity of the analysis units. It comprehensively calculates the importance score of the units based on semantic cohesion, keyword density, discussion time ratio, and integrity verification results. It uses a large language model to score units based on attitude gap, uncertainty gap, and semantic alignment, and employs a dual-model cross-validation mechanism for verification. Based on the multi-dimensional scoring, it calculates the comprehensive gap score of the units and combines it with the importance score to obtain a globally weighted gap index, generating the final evaluation information and presenting it visually. This provides a unified and standardized means for financial text trust monitoring and risk warning.

Description

Technical Field

[0001] This invention relates to the field of long text processing of financial question-and-answer questions based on natural language processing, and in particular to a method and system for analyzing long text of financial question-and-answer questions based on role recognition and dialogue structure parsing. Background Technology

[0002] Existing text analysis technologies have the following limitations when processing long text financial Q&A in scenarios such as earnings calls, performance briefings, and investor surveys:

[0003] (1) Lack of structured modeling capabilities for two-way question-and-answer interaction. Traditional sentiment analysis techniques treat management's statements as independent text fragments, failing to model the semantic relationship between the content of the statements and the analyst's questions they respond to, and making it difficult to quantify the semantic gap between "the sharpness of the question" and "the degree of evasion in the answer". For example, when faced with a sharp question about the decline of core business, if management adopts a reserved and evasive statement, traditional methods may misjudge it as neutral or positive sentiment, failing to identify semantic deviations and insufficient information disclosure.

[0004] (2) Speaker role identification relies on external data sources and has poor adaptability. Existing methods usually rely on manually maintained participant lists or external databases to identify speakers. When meeting record formats are inconsistent, there are temporary participants, or there is a lack of external data support, the accuracy of role identification drops significantly, affecting the reliability of subsequent analysis.

[0005] (3) Topic boundary identification lacks semantic understanding mechanism. Existing methods mostly use fixed rules (such as simple "question and answer" pairing) to divide topics, which is difficult to effectively handle complex dialogue structures such as multi-turn dialogue, follow-up questions, and cross-discussion, and is prone to damage to semantic integrity.

[0006] (4) The consistency assessment has a single dimension and lacks quantitative standards. Existing technologies have not established a multi-dimensional semantic gap quantification model between "expected questions" and "answer statements", nor do they have a technical solution to aggregate discrete gap scores into a unified information disclosure transparency index that can be compared across companies.

[0007] (5) The difference in information value between question and answer segments is not considered. Existing technology assigns the same weight to all question and answer segments, failing to distinguish between core business discussions and casual pleasantries, resulting in the dilution of key signals.

[0008] Therefore, there is an urgent need for a technical solution that can automatically identify question-and-answer roles, intelligently divide semantically complete question-and-answer topic units, quantify the semantic gap between questions and answers in multiple dimensions, and ultimately output a unified information disclosure transparency index. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for analyzing long financial Q&A texts based on role recognition and dialogue structure parsing. This solution is based on a deep analysis framework of "automatic role recognition → intelligent segmentation of semantic topic units → multi-dimensional gap quantification of Q&A → transparency index aggregation," which enables structured modeling of the two-way interactive characteristics of Q&A scenarios and objective quantification of trust levels.

[0010] This invention provides a method for analyzing long texts of financial question-and-answer questions based on role recognition and dialogue structure parsing, comprising the following steps:

[0011] Step 1, Identifying the speaker's role:

[0012] The system extracts the preceding character range from the long text of financial Q&A for the target object and locates the opening paragraph containing the participant introduction sentence through regular expression pattern matching; it then calls a large language model to perform structured parsing of the opening paragraph and extracts the list of participants to construct a role mapping table.

[0013] Based on the constructed role mapping table, a three-level role matching is performed on each speech segment of the long financial Q&A text, and the role label of each speech segment is determined based on the matching results; if there is a role inference in the matching results, the role mapping table is updated based on the role inference.

[0014] Step 2, Question and Answer Topic Unit Recognition:

[0015] Semantic vector representations of each speech segment are generated using a semantic vector model in the financial domain, and the semantic similarity between adjacent speech segments is calculated.

[0016] Identifying topic boundary candidate points based on semantic similarity, explicit connection signals, and role switching patterns;

[0017] The Large Language Model (LLM) is invoked to verify the semantic coherence of candidate points at the topic boundaries, and semantically coherent speech segments are aggregated into question-answering topic units (UNIT).

[0018] Step 3, scoring the importance of the Q&A topic unit:

[0019] Perform structural integrity verification on the question-and-answer topic units to obtain the structural integrity verification results of the question-and-answer topic units;

[0020] Keyword density is calculated based on a pre-set financial information keyword database, as well as the semantic cohesion of the speech vectors in the question-and-answer topic units and the proportion of discussion time.

[0021] The importance score of each question-and-answer topic unit is calculated by comprehensively considering its semantic cohesion, keyword density, discussion time percentage, and structural integrity verification results.

[0022] Step 4: Multidimensional quantification of the semantic gap between question and answer;

[0023] For each question-answer pair in a question-answer topic unit, a large language model is used to score them from three dimensions: attitude gap, uncertainty gap, and semantic alignment, to obtain the question-answer semantic gap vector of the question-answer topic unit.

[0024] A dual-model cross-validation mechanism is used to validate the question-answer semantic gap vector, and the final question-answer semantic gap vector for each question-answer topic unit is determined based on the validation results.

[0025] Step 5: Calculation of the weighted aggregated Trust & Transparency Generation Index (TGI):

[0026] The semantic alignment in the final question-and-answer semantic gap vector is converted into a gap metric to obtain the alignment gap; then, the comprehensive gap score of each question-and-answer topic unit is obtained by weighting the attitude gap, uncertainty gap and alignment gap in the final question-and-answer semantic gap vector.

[0027] The importance score of each Q&A topic unit is normalized and used as the weight of its comprehensive gap score. The global weighted gap index of financial Q&A long text is obtained by weighting the comprehensive gap scores of all Q&A topic units.

[0028] The globally weighted gap index is transformed into a TGI by using an exponential decay mapping function, and the evaluation information of the long text of financial Q&A is visualized and output based on the TGI.

[0029] Furthermore, in step 1, the three-tier role matching specifically involves:

[0030] Level 1 exact match: Extract the speaker identifier field and perform a string-based exact match with the names registered in the role mapping table. If the match is successful, the corresponding role tag is directly assigned; if the match fails, the level 2 fuzzy match is executed.

[0031] Second-level fuzzy matching: Calculate the edit distance between the speaker identifier and each name in the role mapping table. If the minimum edit distance is less than a preset threshold and is unique, then assign the role label corresponding to the minimum edit distance to the current speech segment; otherwise, perform third-level semantic inference.

[0032] The third level of semantic inference involves calling a large language model to analyze the language features of the current speech segment (including the frequency of interrogative sentences, the usage pattern of first-person plural pronouns, and the types of professional terms), performing role inference based on the language features, and recording the role inference results in a role mapping table.

[0033] Furthermore, in step 1, a role mapping table is constructed using a two-layer classifier based on rules and semantic understanding. The two-layer classifier includes:

[0034] Rule layer: Based on the job keyword database, preliminary classification is carried out, including: management, analysts and moderators; for example, personnel whose job descriptions contain keywords such as CEO, CFO, chairman, general manager, etc. are classified as management; personnel whose institutional descriptions contain keywords such as securities, research, fund, brokerage are classified as analysts; and personnel whose job descriptions contain keywords such as moderator, meeting organizer, etc. are classified as moderators.

[0035] Semantic Validation Layer: Performs semantic validation on the preliminary classification results of the rule layer, calls the large language model to make a comprehensive judgment based on the full text description of the job title and the background of the organization, corrects or supplements the uncertain classifications generated by the rule layer, and finally forms a complete role mapping table.

[0036] Furthermore, in step 2, an entity-enhanced hybrid similarity algorithm is used to calculate the semantic similarity between adjacent speech segments. The calculation formula is as follows:

[0037]

[0038] in, The semantic similarity between adjacent speech segments. Number the speech segment. For the speaking segment semantic vectors, For the speaking segment The semantic vector; the first term of this semantic similarity is the cosine similarity based on the semantic vector, which is used to capture coherence at the general language level; and There are two weighting coefficients, and they satisfy... + =1; the second term is the Jaccard similarity based on financial named entities, and its calculation formula is: ,in, , From the speech segment and speech segment The set of key financial entities extracted from this data. This set of key financial entities can be extracted using financial-specific Named Entity Recognition (NER) tools, including financial indicators, time periods, and specific business objects;

[0039] The entity-enhanced hybrid similarity algorithm used in this invention solves the problem of insufficient differentiation between similar-looking but different financial terms such as "year-on-year" and "month-on-month", "first quarter" and "second quarter" by introducing entity overlap, and significantly improves the business logic accuracy of topic segmentation.

[0040] Furthermore, in step 2, the candidate points for identifying topic boundaries based on semantic similarity, explicit connection signals, and role-switching patterns include:

[0041] Similarity drop detection: When the semantic similarity of adjacent speech segments ( When the value falls below a set dynamic threshold, the adjacent positions of adjacent speech segments are recorded as candidate points for topic boundaries. This dynamic threshold is determined by analyzing the distribution characteristics of the semantic similarity of all adjacent speech segments in a long financial Q&A text; for example, for adjacent speech segments... The corresponding topic boundary candidate points are the speech segments. End of speech or speech paragraph The starting point; the dynamic threshold can be set to the 25th percentile of the semantic similarity distribution of all adjacent speech segments.

[0042] Explicit transition signal detection: Explicit transition words in the speech segment are matched using regular expressions, and the positions of the matched explicit transition words are marked as candidate points for topic boundaries; among them, explicit transition words include but are not limited to "next question", "in addition", "change topic", etc.

[0043] Role switching pattern detection: When the speaker's role sequence shows a switching pattern from analyst to management and back to analyst, the position where the analyst reappears and speaks in this switching pattern is marked as a topic boundary candidate point. That is, the second switching position of the switching pattern is a topic boundary candidate point, which is used to handle multiple rounds of follow-up questions from the same analyst.

[0044] Furthermore, in step 2, the semantic coherence verification of the topic boundary candidate points includes:

[0045] For each candidate topic boundary, a text segment of fixed window size (e.g., 3 statements before and after) is taken before and after each candidate topic boundary to construct verification prompts. The large language model is required to evaluate whether the candidate topic boundary constitutes a real topic transition. The evaluation of whether it constitutes a real topic transition includes: whether the topic of the text segments before and after the transition changes, whether the subject of discussion changes, and whether there is a difference in the questioning intent.

[0046] When the large language model determines that a topic boundary candidate point does not constitute a real topic transition, it removes the topic boundary candidate point, thereby avoiding the incorrect segmentation of semantically coherent multi-turn dialogues.

[0047] Furthermore, in step 3, performing structural integrity verification on the question-and-answer topic units includes:

[0048] Test 1: Check whether the Q&A topic unit contains at least one statement from an analyst and one statement from a manager;

[0049] Test 2: Check whether the speaking sequence in the question-and-answer topic unit conforms to the logical order of questions and answers;

[0050] If a question-and-answer topic unit satisfies both Detection 1 and Detection 2, then the question-and-answer topic unit is marked as structurally complete; otherwise, it is marked as structurally incomplete.

[0051] In this invention, based on whether the question-and-answer topic unit has a complete structure, a corresponding penalty coefficient can be applied when calculating the importance score. If the structure is complete, the corresponding structural integrity verification coefficient is applied. Set to 1, otherwise in Settings The value at this time It can be regarded as the corresponding penalty coefficient.

[0052] Furthermore, in step 3, the semantic cohesion is calculated as follows:

[0053] Extract the semantic vector set of all speech segments in the question-and-answer topic unit. And calculate its centroid vector. ; These are the semantic vectors for speech segments 1 through m; calculate each semantic vector. With the centroid vector The cosine similarity is calculated, and the average value is taken as the semantic cohesion:

[0054]

[0055] in, For the first The semantic cohesion of each question-and-answer topic unit The number of speaking segments in the question-and-answer topic unit. Indicates the speech segment semantic vectors With the centroid vector The cosine similarity.

[0056] Furthermore, in step 3, the keyword density is calculated as follows:

[0057]

[0058] in, For the first Keyword density of each question and answer unit, For the first The complete text content of each question and answer unit (including the question text and the answer text). Financial keyword database vocabulary in exist The total number of times it appears in For the first The total number of characters in each question-and-answer unit, including spaces.

[0059] Furthermore, in step 3, the calculation method for the proportion of discussion time is as follows:

[0060]

[0061] in, For the first The percentage of discussion time for each Q&A session. For the first The number of speech segments contained in each question-and-answer unit. This refers to the total number of segments in the long text of the financial Q&A section.

[0062] Furthermore, in step 3, the formula for calculating the importance score is:

[0063]

[0064] in, , , The first The semantic cohesion, keyword density, and discussion duration of each Q&A topic unit; , , These are the weighting coefficients for semantic cohesion, keyword density, and discussion duration, respectively, and they satisfy the following conditions: ; The structural integrity verification coefficient is set to the following value: if the question-and-answer topic unit structure is complete, then... Set to 1; if the question-and-answer topic unit structure is incomplete, then .

[0065] Furthermore, in step 4, the attitude gap, uncertainty gap, and semantic alignment are set as follows:

[0066] The attitude gap is set as follows: ,in, and These are the sentiment scores for management's responses and analysts' questions, respectively, obtained based on a large language model. , To embellish the penalty coefficient;

[0067] The uncertainty gap is set as follows: ,in, and These represent the degree of uncertainty in analyst questions and the degree of uncertainty in management responses, respectively, obtained through a large language model; and ;

[0068] Semantic alignment The setting is: the semantic alignment obtained by semantic matching analysis of question-answer pairs in question-answer topic units using a large language model, and .

[0069] In this invention, considering that management may have a defensive spin tendency when facing questions, that is, covering up problems with overly positive statements, this invention utilizes... and Calculate the attitude gap .in, and The specific calculation method is as follows: A sentiment quantification cue word based on a large language model is constructed. This cue word contains a preset sentiment baseline. The question or answer text is input into the large language model, which identifies sentiment verbs, adverbs, and context in the text, and finally maps them to a continuous numerical range of [-1, 1]. Attitude gap The calculation logic is as follows: when When this occurs, it indicates that management's answer is significantly more optimistic than the question asked, suggesting a risk of "downplaying the issue" and automatically amplifying the weight of this discrepancy; when This indicates that management is cautious or in line with analysts, and only the standard absolute difference is calculated. This calculation mechanism can more accurately capture potential trust risks.

[0070] In calculating the uncertainty gap hour, and The evaluation can be based on the frequency of use and semantic strength of vague words (such as "maybe", "probably", "perhaps"), with a value range of [0,1]. When the value is positive, it indicates that management used more vague language in their responses compared to analysts' questions; when... When the value is close to 0, it indicates that the degree of certainty for both parties is roughly equal.

[0071] Semantic alignment A large language model is used to perform semantic matching analysis on question-answer pairs, with values ​​ranging from [0,1]. If the answer directly covers the core keywords in the question and is logically consistent, it is judged as having high alignment. The value range is (0.8, 1.0]; if the answer is relevant but avoids the core question, it is judged as center alignment, and the corresponding... The value range is (0.4, 0.8]; if the answer completely deviates from the question's topic, it is judged as low alignment, and the corresponding... The range of its value is [0, 0.4].

[0072] Furthermore, in step 4, the dual-model cross-validation mechanism is used to validate the question-answer semantic gap vector, specifically including:

[0073] For each question-answer pair within a question-answer topic unit, a large language model is used to score them from three dimensions: attitude gap, uncertainty gap, and semantic alignment, to obtain the question-answer semantic gap vector of the question-answer topic unit.

[0074] A dual-model cross-validation mechanism is used to validate the question-answer semantic gap vector, and the final question-answer semantic gap vector for each question-answer topic unit is determined based on the validation results.

[0075] definition , and These represent the attitude gap, uncertainty gap, and semantic alignment, respectively.

[0076] By calling two large language models with different architectures or different training parameters in parallel to score the same question-answering topic unit, two sets of question-answering semantic difference vectors for that topic unit are obtained: and ;

[0077] Calculate the Euclidean distance between the two sets of question-answer semantic gap vectors as a consistency metric:

[0078]

[0079] in, , and The superscript is used to distinguish different large language models;

[0080] If consistency index Less than or equal to the preset threshold If the mean of the two sets of question-and-answer semantic gap vectors is used, the final question-and-answer semantic gap vector of the current question-and-answer topic unit is obtained; otherwise, a third-party arbitration mechanism or manual review process is triggered to determine the final question-and-answer semantic gap vector of the current question-and-answer topic unit to ensure the reliability of the score.

[0081] Furthermore, step 4 also includes a reverse verification mechanism, specifically:

[0082] definition Indicates the first The question and answer text for each question and answer topic unit;

[0083] After obtaining the question-answer semantic gap vector of the question-answer topic unit using a large language model, reverse verification prompts are constructed, and the textual representation of the question-answer pair of the current question-answer topic unit is regenerated using the large language model based on the obtained question-answer semantic gap vector. ;

[0084] Then and Perform semantic similarity comparison. If the semantic similarity is lower than or equal to a preset threshold, then re-apply the large language model to obtain the next semantic similarity. The semantic gap vector between the question and answer topics in each question-and-answer unit, or the ... can be re-determined through manual review. The semantic gap vector between questions and answers for each question-and-answer topic unit.

[0085] Furthermore, in step 5, the overall gap score is:

[0086]

[0087] in, For the first The overall gap score for each question and answer topic unit. , , The first The attitude gap, uncertainty gap, and semantic alignment of each question-and-answer topic unit. It is the absolute value symbol. Indicates the first The alignment gap between individual question and answer topic units (i.e., the lower the semantic alignment, the greater the alignment gap). , , These are the weighting coefficients for the attitude gap, uncertainty gap, and alignment gap, respectively, and they satisfy the following conditions: .

[0088] Furthermore, in step 5, the formula for calculating TGI is:

[0089]

[0090] in, This represents the natural exponential function. These are preset sensitivity adjustment parameters used to control the sensitivity of TGI to changes in the gap. This is a globally weighted difference index. The calculation formula ensures that the TGI value falls within the interval (0,1]. The smaller the value, the closer the TGI is to 1, indicating higher transparency.

[0091] Another aspect of the present invention provides a financial question-and-answer long text analysis system based on role recognition and dialogue structure parsing, which includes a role recognition module, a question-and-answer topic unit recognition module, a quality control module, a gap quantification module, an index aggregation module, and a display module;

[0092] in,

[0093] Role recognition module: Extracts the preceding character range from the long text of financial Q&A for the target object, and locates the opening paragraph containing the participant introduction sentence through regular expression pattern matching; calls a large language model to perform structured parsing of the opening paragraph, extracts the list of participants to construct a role mapping table;

[0094] Based on the constructed role mapping table, a three-level role matching is performed on each speech segment of the long text of financial Q&A. The role label of each speech segment is determined based on the matching results, and the role label of each speech segment is output to the topic unit recognition module. If there is role inference in the matching results, the role mapping table is updated based on the role inference.

[0095] The question-and-answer topic unit identification module generates semantic vector representations of each speech segment using a semantic vector model in the financial field and calculates the semantic similarity between adjacent speech segments; it identifies candidate points for topic boundaries based on semantic similarity, explicit cohesion signals, and role switching patterns; it calls a large language model to verify the semantic coherence of the candidate points for topic boundaries, and aggregates semantically coherent speech segments into question-and-answer topic units; it outputs a structured list of question-and-answer topic units to the quality control module, the gap quantification module, and the display module, respectively.

[0096] The quality control module performs structural integrity verification on each Q&A topic unit in the Q&A topic unit list, obtaining the structural integrity verification result; calculates keyword density based on a preset financial information keyword library, and calculates the semantic cohesion of the speech vectors and the proportion of discussion time for each Q&A topic unit; comprehensively calculates the importance score of each Q&A topic unit based on the semantic cohesion, keyword density, proportion of discussion time, and structural integrity verification result; and outputs the importance score of each Q&A topic unit to the index aggregation module and the display module.

[0097] The gap quantification module calls a large language model to perform multi-dimensional scoring of the question-answer pairs for each question-answer topic unit, including attitude gap, uncertainty gap, and semantic alignment, to obtain the question-answer semantic gap vector for the question-answer topic unit. A dual-model cross-validation mechanism is used to validate the question-answer semantic gap vector. Based on the validation results, the final question-answer semantic gap vector for each question-answer topic unit is determined and output to the exponential aggregation module.

[0098] The index aggregation module converts the semantic alignment in the final question-and-answer semantic gap vector into a gap metric to obtain the alignment gap. Then, based on the weighted sum of the attitude gap, uncertainty gap, and alignment gap in the final question-and-answer semantic gap vector, it obtains the comprehensive gap score for each question-and-answer topic unit. The importance score of each question-and-answer topic unit is normalized and used as the weight for its comprehensive gap score. Based on the weighted sum of the comprehensive gap scores of all question-and-answer topic units, it obtains the global weighted gap index for the long financial question-and-answer text. Finally, through an exponential decay mapping function, the global weighted gap index is converted into a Trust and Transparency Index (TGI), and the comprehensive gap score and TGI are output to the display module.

[0099] The display module is used to visualize the processing results to form evaluation information for the long text of financial Q&A to be evaluated, including: displaying the question text, answer text and comprehensive gap score of each individual Q&A topic unit in chronological order; drawing a time series chart of TGI to show the TGI change trend of the same target object in previous meetings; and drawing a bar chart comparing the TGI of different target objects in the same industry.

[0100] The technical solution provided by this invention brings at least the following beneficial effects:

[0101] (1) High-precision role recognition with zero external dependency: The role mapping table can be constructed by relying only on the opening information of the meeting record itself, and supports dynamic updates, which solves the problem of poor robustness caused by relying on external lists.

[0102] (2) Semantic complete question-and-answer topic unit division: Based on the dual mechanism of vector similarity and semantic parsing of large language model, it can accurately identify multi-turn dialogues and complex interaction structures, ensuring the semantic integrity of question-and-answer topic units.

[0103] (3) Breakthrough two-way gap quantification: It quantifies the gap between analysts’ questions and management’s answers in terms of attitude, uncertainty and semantic alignment, and can effectively identify hidden information disclosure problems such as “superficial positive but actual avoidance”.

[0104] (4) Unified comparable transparency metrics: A normalized TGI with clear business meaning has been constructed to support horizontal comparison and vertical monitoring across objects and time.

[0105] (5) Robust quality control: By weighting the importance of question-and-answer topic units and cross-validating multiple models, the weight of key information is ensured and the risk of illusion generated by a single model is reduced. Detailed Implementation

[0106] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments.

[0107] This invention provides a method for analyzing long texts of financial questions and answers based on role recognition and dialogue structure parsing. The specific implementation steps include:

[0108] Step S1: Speaker role identification based on opening information

[0109] This step automatically constructs speaker role mappings using only long texts of financial Q&A (such as earnings calls, performance briefings, investor surveys, etc.) without relying on external databases.

[0110] (1) Opening Information Positioning and Analysis

[0111] The system scans the first 1500 characters of long financial Q&A texts. It then locates the opening paragraph using a predefined set of regular expression patterns, including but not limited to:

[0112] Match paragraph headings containing keywords such as "participants" or "attendees".

[0113] Matches text lines containing the job description pattern, where the job description pattern is in the format "Name, Company Name, Job Title";

[0114] Matches text lines that contain the format of organization name and personal name listed side by side, such as the structure "Name – Organization Name".

[0115] After locating the opening paragraph, the large language model is invoked to perform structured parsing of that paragraph. Constructing prompts requires the large language model to extract three pieces of information for each participant: name, position, and affiliation, and output them in JSON format.

[0116] (2) Construction of the role mapping table

[0117] A two-layer classifier is used to determine the roles of the extracted attendees:

[0118] Rule-based classification: Maintain three keyword databases: a management position keyword database, an analyst / institution keyword database, and a moderator position keyword database. The management position keyword database includes CEO, CFO, Chairman, General Manager, CTO, COO, etc.; the analyst / institution keyword database includes securities, research, investment, funds, brokerage firms, asset management, etc.; and the moderator position keyword database includes moderator, meeting organizer, investor relations, etc. It iterates through the position and institution fields of each participant, and if a keyword matches a corresponding database, categorizes them into the appropriate role.

[0119] Semantic Validation Layer: Performs semantic validation on uncertain classifications generated by the rule layer (e.g., job descriptions are non-standard expressions such as "consultant" or "expert"). It calls a large language model, inputting the full job description and organizational background of the individual. The model is required to make a comprehensive judgment based on its knowledge of the financial industry, outputting the role category and confidence score. If the confidence score is higher than a preset threshold (e.g., 0.8), the classification result is adopted.

[0120] After processing by a two-layer classifier, a RoleMap table is constructed. The data structure is a dictionary, with the key being the name of the participant and the value being the role label (management, analyst, or host).

[0121] (3) Three-level role matching for main text speech

[0122] Perform three-level role matching for each speech segment in the main body of the meeting minutes:

[0123] Level 1 exact match: Extract the speaker identifier field (usually located at the beginning of the speech paragraph, in the format "Name:") and perform a string-by-string match with the names already registered in the RoleMap. If the match is successful, the corresponding role tag is directly assigned to the speech paragraph; if the match fails, the level 2 fuzzy match is performed.

[0124] The second level of fuzzy matching calculates the edit distance between the speaker's identifier and each name in the RoleMap. If a unique minimum edit distance exists and this distance is less than a preset threshold (e.g., 2), the speaker is determined to be the corresponding role; otherwise, the third level of semantic inference is performed. This matching mechanism is used to handle misspelled or abbreviated names, such as matching "Zhang Zong" with "Zhang San".

[0125] Third-level semantic inference: When the first two levels of matching fail, a large language model is invoked to perform linguistic feature analysis on the speech segment. Specifically, the following features are extracted:

[0126] Frequency of interrogative sentences: assessed by counting the number of question marks and the frequency of interrogative words (such as "what", "how", etc.).

[0127] First-person plural pronoun usage patterns: Analysts tend to use "we" to refer to the research team or investor group, while management tends to use "we" to refer to the company.

[0128] Types of technical terms: Analysts may use investment analysis terms such as "valuation", "return on equity", and "cash flow"; management may use operations management terms such as "capacity", "channels", and "R&D investment".

[0129] Based on these features, the large language model outputs role inference results and confidence scores. If the confidence score meets the requirements, the inference result and the new speaker information are dynamically updated in the RoleMap for matching subsequent speeches.

[0130] (4) Dynamic update mechanism

[0131] During the processing of meeting minutes, when a new speaker is identified through third-level semantic inference, the system automatically records their name, inferred role label, and confidence level in the RoleMap. This dynamic update mechanism enables the system to handle temporary attendees or opening information, improving the adaptability of role recognition.

[0132] Step S2: Question-answering topic unit identification based on semantic vector clustering and large language model understanding

[0133] This step transforms linear meeting transcript text into structured question-and-answer topic units.

[0134] (1) Semantic vectorization of speech segments

[0135] We utilize a semantic vector model from the financial domain to generate fixed-dimensional semantic vector representations for each speech segment in the meeting transcript. The specific process is as follows:

[0136] The text of each speech segment is preprocessed, including removing format marks and normalizing punctuation.

[0137] The preprocessed text is input into a semantic vector model in the financial domain to obtain its hidden layer representation.

[0138] The average pooled vector of the hidden layer representation is taken as the speech vector of this speech segment. That is, semantic vector.

[0139] In this embodiment, a pre-trained semantic vector model (such as the FinBERT2 model) based on the Bidirectional Encoder Representation Model (BERT) architecture is used for fine-tuning on a massive amount of high-quality Chinese financial corpus (such as financial reports, research reports, financial news, etc.) to enable the model to understand complex financial contexts and transform unstructured meeting transcripts into fixed-dimensional semantic vectors that can be computed by computers.

[0140] (2) Calculation of entity-enhanced hybrid semantic similarity between adjacent speech segments

[0141] For the chronological sequence of speeches in meeting minutes, to address the issue of insufficient differentiation between similar-looking but semantically different financial terms such as "year-on-year / month-on-month" and "first quarter / second quarter" using general semantic vectors, this embodiment employs an entity-enhanced hybrid similarity algorithm to calculate the degree of association between adjacent speech segments, thereby obtaining the semantic similarity between adjacent speech segments. The specific calculation process includes:

[0142] The first step is to calculate the general semantic similarity to obtain the semantic vector. , Subscripts are used to identify speech segment numbers; calculation and Cosine similarity It is used to capture coherence at the level of a common language.

[0143] The second step is to calculate the overlap of entity concerns. Using a pre-built financial sector-specific named entity recognition tool, the overlap is calculated from the first... Section and the Extracting the set of key financial entities from the speech text and .calculate and similarity to Jaccard .

[0144] The third step is weighted fusion.

[0145] Based on the two set weights and For cosine similarity Similarity to Jaccard Weighted fusion is performed, in which, In this embodiment, the following is set: (Emphasis on semantic coherence) (Focusing on the continuity of business entities), the final semantic similarity is calculated. This can also be called hybrid similarity; a similarity sequence is formed based on the calculated semantic similarity.

[0146] (3) Identification of topic boundary candidate points

[0147] Similarity drop detection: Analyze the distribution characteristics of similarity sequences and calculate the 25th percentile of the sequence as a dynamic threshold. When the similarity of a pair of adjacent speech segments At that time, mark the position These serve as candidate points for topic boundaries. This method can adapt to the semantic density distribution of different meetings.

[0148] Explicit transition signal detection: Maintain an explicit transition vocabulary, including common topic transition expressions in Chinese and English, such as "next question", "in addition", "change the topic", "let me think about it", etc.

[0149] These transition words are matched at the beginning of each speech segment using regular expressions. If a match is detected, the start position of that speech segment is marked as a candidate topic boundary.

[0150] Role switching pattern detection: Analyze the speaking role sequence. When a switching pattern of "analyst → management → analyst" occurs, mark the position of the second switching (the position where the analyst reappears and speaks in the switching pattern) as a candidate point for the topic boundary. This mechanism is particularly useful for handling scenarios where the same analyst asks multiple follow-up questions, because follow-up questions usually form a pattern of "analyst asks a question → management answers → analyst asks follow-up questions".

[0151] (4) Semantic coherence verification of topic boundary candidate points

[0152] Not all topic boundary candidate points constitute a true topic transition. To avoid over-segmentation, a large language model can be used to verify the semantic coherence of each topic boundary candidate point.

[0153] For the candidate point of the topic boundary, extract the three paragraphs of speech before and after it. If the candidate point of the topic boundary is close to the beginning or end of the text, adjust the window size accordingly.

[0154] A verification prompt is constructed to ask: "Do the statements before and after the topic boundary candidate point discuss the same topic?". In this embodiment, the judgment criteria for the large language model to perform the query evaluation include: whether the topic of the preceding and following text segments has undergone a substantial change, such as from discussing "revenue growth" to discussing "cost control"; whether the object of discussion has changed, such as from discussing "product A" to discussing "product B"; and whether there is a significant difference in the intent of the question, such as from asking about "performance reasons" to asking about "future plans". If the large language model determines that the topic boundary candidate point does not constitute a real topic change (i.e., the statements before and after are still discussing the same topic), then the topic boundary candidate point is removed.

[0155] (5) Generation of question and answer topic units

[0156] The topic boundary candidate points retained after semantic coherence verification divide the speech sequence into multiple segments. For each segment, combining the role information identified in step S1, the "analyst speech + management speech" are aggregated into a structured question-and-answer topic unit (UNIT). Each UNIT object contains the following fields:

[0157] Question text: A compilation of all analyst comments within this unit;

[0158] Answer text: A compilation of all management statements within this unit;

[0159] Role Sequence: The sequence of role labels for each speech segment within this unit;

[0160] Start and End Positions (StartIndex, EndIndex): The start and end indices of this UNIT in the original speech sequence.

[0161] Step S3: Quality control and importance scoring of question-and-answer units

[0162] To distinguish between core information and ineffective dialogue, an importance score is calculated for each unit. To further distinguish between core information and non-substantive dialogue, quality control is performed on each unit, and an importance score is calculated, specifically including:

[0163] (1) Structural integrity verification

[0164] Structural integrity checks are performed on each unit, and the criteria for judgment include:

[0165] Does this unit contain at least one analyst's statement? If not, it indicates a missing question section.

[0166] Does this unit contain at least one statement from a management role? If not, it indicates a missing response section.

[0167] Does the role sequence of this UNIT conform to the logical order of question and answer? Ideally, it should be "Analyst → Management" or "Analyst → Management → Analyst → Management" (multiple rounds of dialogue). If there is an obvious abnormal sequence (such as multiple management personnel speaking in succession without any analyst asking questions in between), it is judged as structural abnormality.

[0168] For UNITs with incomplete structures, a penalty coefficient of 0.5 is applied when calculating their importance score to reduce their weight.

[0169] (2) Keyword density calculation

[0170] Based on the maintained financial information keyword database, for the k-th unit, the frequency of keyword occurrence in its question and answer texts is calculated to obtain the keyword density. The financial information keyword database covers financial indicators, business indicators, and strategic themes, such as "net profit," "year-on-year growth," "gross profit," "revenue," "cash flow," and "market share."

[0171] (3) Semantic cohesion calculation

[0172] Extract the semantic vector set of all statements for each unit. And calculate its centroid vector. ; This is the speech vector of the current UNIT. For the number of speech vectors; calculate and cosine similarity Then, the semantic cohesion of the current UNIT is obtained by taking the average. .

[0173] (4) Calculation of the proportion of discussion time

[0174] For the k-th unit, calculate the proportion of its speaking segments to the total number of speaking segments in the entire meeting to obtain the proportion of discussion time for the k-th unit. .

[0175] (5) Overall Importance Score

[0176] First, based on the above calculation results, calculate the importance score for each UNIT. :

[0177]

[0178] in, , , These are the weighting coefficients for semantic cohesion, keyword density, and discussion duration, respectively, and they satisfy the following conditions: In this embodiment, the following is set: , , .

[0179] If the structural integrity verification of a unit fails (i.e., the unit structure is incomplete), its importance score will be discounted.

[0180]

[0181] in, This represents the final importance score of the k-th UNIT. In this embodiment, the penalty coefficient is... Set as For UNITs that pass structural integrity verification, then .

[0182] Step S4: Multidimensional Quantification of Question-Answer Semantic Gap Based on Large Language Model

[0183] In this step, a large language model is used to score the semantic gap of each UNIT in multiple dimensions to obtain the question-answer semantic gap vector of each UNIT.

[0184] (1) Acquisition of three-dimensional scores

[0185] Regarding the first For each unit, construct prompt words and provide the question text and answer text for that unit to the large language model. The model is required to score the unit based on the following three dimensions:

[0186] Attitude Gap: Calculating the Sentiment Values ​​of Analyst Questions Using a Large Language Model Sentimental bias in management's responses The attitude gap is then calculated based on the size relationship between the two. .

[0187] Uncertainty Gap: Calculating the degree of uncertainty in analyst questions using a large language model. The degree of uncertainty in the management's response Based on and The difference yields the uncertainty gap .

[0188] Semantic alignment: The semantic alignment obtained by performing semantic matching analysis on the question-answer pairs of UNIT using a large language model. .

[0189] When analyzing semantic alignment, this invention uses a large language model to evaluate the degree of direct response of management's answers to analyst questions. In this embodiment, the large language model makes judgments based on the following criteria:

[0190] Does the answer directly address the key theme of the question? For example, if the question asks "Why did the gross profit margin decline?", the answer should directly explain the reason rather than shifting to other topics.

[0191] Does the answer provide the specific information expected in the question? For example, if the question asks for "revenue guidance for the next quarter", the answer should provide specific figures or ranges rather than just qualitative descriptions.

[0192] The answer should address whether there is any attempt to shift the topic or avoid the question, such as asking a negative question but emphasizing other positive aspects in the answer.

[0193] In the specific calculation, if the answer directly covers the core keywords in the question and corresponds logically, it is judged as high alignment, and the corresponding semantic alignment value range is [0.8-1.0]; if the answer is relevant but avoids the core question, it is judged as medium alignment, and the corresponding semantic alignment value range is [0.4-0.8); if the answer completely deviates from the question topic, it is judged as low alignment, and the corresponding semantic alignment value range is [0-0.4].

[0194] Based on the A unit , and Obtain its question-answer semantic gap vector .

[0195] (2) Dual-model cross-validation mechanism

[0196] To ensure the stability and reliability of the scoring, a dual-model cross-validation mechanism is adopted:

[0197] Two large language models with different architectures or training parameters are called in parallel to score the same unit. The question-answer semantic gap vector obtained by one of the models is: The other model yielded the following question-answering semantic gap vector: .in , and The superscript is used to distinguish different large language models.

[0198] Calculate the Euclidean distance between the two sets of question-answer semantic gap vectors as a consistency metric:

[0199]

[0200] If consistency index Then, the arithmetic mean of the two sets of question-answer semantic gap vectors is taken as the final question-answer semantic gap vector of the current UNIT; if If this is not the case, a third-party arbitration mechanism or manual review process will be triggered to ensure the reliability of the score. In this embodiment, as a preset threshold, Set it to 0.3.

[0201] (3) Reverse verification mechanism

[0202] To obtain the question-answer semantic gap vector of a certain unit from a large language model. Then, reverse validation is performed to check whether the model truly understands the question-and-answer content:

[0203] definition Indicates the first A unit's question and answer text;

[0204] Constructing reverse verification prompts requires the large language model to base its work on the currently obtained question-answer semantic gap vector (e.g., , , Regenerate a possible question-and-answer pair text representation, denoted as . The resulting question-answer semantic gap vector is considered the initial score for UNIT.

[0205] Will and The results are converted into semantic vectors, and then the cosine similarity is calculated. If the cosine similarity is lower than a preset threshold (e.g., 0.7), it is determined that the initially obtained question-answer semantic gap vector may have a bias or the model has produced an illusion, and it needs to be re-scored or transferred to the manual review process.

[0206] This reverse verification mechanism reduces the risk of unreliable scores generated by large language models through a closed-loop detection process of "scoring → generation → comparison".

[0207] (4) Calculation of the overall gap score of a single unit

[0208] The first The question-answer semantic gap vector of each UNIT ( , and The weighted aggregate is combined into a single comprehensive gap score. ,in, , and The first Attitude gap, uncertainty gap and semantic alignment of individual UNITs.

[0209] In this embodiment, firstly according to Calculate the first The alignment difference of each unit is then determined based on the set attitude difference weight. Uncertainty gap weight Alignment difference weight Perform weighted aggregation to obtain In this embodiment, the following is set: , , This is to emphasize the impact of vague statements and irrelevant answers on trust levels.

[0210] Step S5: Calculation of a weighted aggregated trust and transparency index

[0211] First, the discrete gap scores are aggregated into a final index:

[0212]

[0213] in, This is a weighted gap indicator for the entire meeting. The total number of question-and-answer topic units identified throughout the entire conference. Score the importance of the k-th question-and-answer topic unit. For the first The overall gap score for each question and answer topic unit. For the first The importance score for each Q&A topic unit.

[0214] Normalized weights This ensures that the sum of the weights of each UNIT is 1, thus... The numerical range is mainly determined by The scope is determined by.

[0215] This weighted aggregation mechanism makes core UNITs with high semantic cohesion, high density of financial keywords, and long discussion time have a greater impact on the final metrics, while non-substantive dialogues (such as small talk) have a smaller impact.

[0216] (2) Mapping calculation of TGI

[0217] Using the exponential decay function to weight the global gap index Convert to TGI:

[0218]

[0219] in, Represents the natural exponential function; This is a sensitivity adjustment parameter used to control the sensitivity of TGI to changes in the gap; in this embodiment, it is set to... The selection of this parameter makes it possible when When varying within the range of [0,1], TGI exhibits moderate sensitivity in its response.

[0220] (3) TGI output and visualization

[0221] The calculated TGI, used as a comprehensive evaluation index of trust and transparency for this meeting, can be applied in the following ways:

[0222] Time series analysis: Tracking the TGI trends of the same company's past meetings to monitor the evolution of information disclosure transparency;

[0223] Horizontal comparison: Compare the TGI levels of different companies in the same industry to identify relative differences in the quality of information disclosure;

[0224] Early warning mechanism: Set a TGI threshold, and trigger a risk warning when the TGI of a meeting falls below the threshold.

[0225] In one embodiment, the present invention provides a long-text financial question-and-answer system, comprising: a role recognition module, a question-and-answer topic unit recognition module, a quality control module, a gap quantification module, an index aggregation module, and a display module; the specific configuration of each module is as follows:

[0226] Role Recognition Module: Taking the meeting transcript text of the input target object as the long text of financial Q&A to be evaluated, the module extracts the preceding character range of the long text of financial Q&A, locates the opening paragraph containing the participant introduction sentence through regular expression pattern matching, and calls a large language model to perform structured parsing of the opening paragraph to extract the list of participants to construct a role mapping table;

[0227] Based on the constructed role mapping table, a three-level role matching is performed on each speech segment of the long text of financial Q&A. The role label of each speech segment is determined based on the matching results, and the role label of each speech segment is output to the topic unit recognition module. If there is role inference in the matching results, the role mapping table is updated based on the role inference.

[0228] The question-and-answer topic unit identification module generates semantic vector representations of each speech segment using a semantic vector model in the financial field and calculates the semantic similarity between adjacent speech segments; it identifies candidate points for topic boundaries based on semantic similarity, explicit cohesion signals, and role switching patterns; it calls a large language model to verify the semantic coherence of the candidate points for topic boundaries, and aggregates semantically coherent speech segments into question-and-answer topic units; it outputs a structured list of question-and-answer topic units to the quality control module, the gap quantification module, and the display module, respectively.

[0229] The quality control module performs structural integrity verification on each Q&A topic unit in the Q&A topic unit list, obtaining the structural integrity verification result; calculates keyword density based on a preset financial information keyword library, and calculates the semantic cohesion of the speech vectors and the proportion of discussion time for each Q&A topic unit; comprehensively calculates the importance score of each Q&A topic unit based on the semantic cohesion, keyword density, proportion of discussion time, and structural integrity verification result; and outputs the importance score of each Q&A topic unit to the index aggregation module and the display module.

[0230] The gap quantification module calls a large language model to perform multi-dimensional scoring of the question-answer pairs for each question-answer topic unit, including attitude gap, uncertainty gap, and semantic alignment, to obtain the question-answer semantic gap vector for the question-answer topic unit. A dual-model cross-validation mechanism is used to validate the question-answer semantic gap vector. Based on the validation results, the final question-answer semantic gap vector for each question-answer topic unit is determined and output to the exponential aggregation module.

[0231] The index aggregation module converts the semantic alignment in the final question-and-answer semantic gap vector into a gap metric to obtain the alignment gap. Then, based on the weighted sum of the attitude gap, uncertainty gap, and alignment gap in the final question-and-answer semantic gap vector, it obtains the comprehensive gap score for each question-and-answer topic unit. The importance score of each question-and-answer topic unit is normalized and used as the weight for its comprehensive gap score. Based on the weighted sum of the comprehensive gap scores of all question-and-answer topic units, it obtains the global weighted gap index for the long financial question-and-answer text. Finally, through an exponential decay mapping function, the global weighted gap index is converted into a Trust and Transparency Index (TGI), and the comprehensive gap score and TGI are output to the display module.

[0232] The display module is used to visualize the processing results of the module to form evaluation information of the long text of financial Q&A to be evaluated, including: displaying the question text, answer text and comprehensive gap score of each individual Q&A topic unit in chronological order; drawing a time series chart of TGI to show the TGI change trend of the same target object in previous meetings; and drawing a bar chart comparing the TGI of different target objects in the same industry to facilitate horizontal comparison.

[0233] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0234] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A long text analysis method for financial question-and-answer based on role recognition and dialogue structure parsing, characterized in that, Includes the following steps: Step 1, Identifying the speaker's role: The system extracts the preceding character range from the long text of financial Q&A for the target object and locates the opening paragraph containing the participant introduction sentence through regular expression pattern matching; it then calls a large language model to perform structured parsing of the opening paragraph and extracts the list of participants to construct a role mapping table. Based on the constructed role mapping table, a three-level role matching is performed on each speech segment of the long text of financial Q&A, and the role label of each speech segment is determined based on the matching results; If the matching results contain role inferences, then update the role mapping table based on the role inferences; Step 2, Question and Answer Topic Unit Recognition: Semantic vector representations of each speech segment are generated using a semantic vector model in the financial domain, and the semantic similarity between adjacent speech segments is calculated. Identifying topic boundary candidate points based on semantic similarity, explicit connection signals, and role switching patterns; The large language model is invoked to verify the semantic coherence of candidate points at the topic boundaries, and semantically coherent speech segments are aggregated into question-and-answer topic units. Step 3, scoring the importance of the Q&A topic unit: Perform structural integrity verification on the question-and-answer topic units to obtain the structural integrity verification results of the question-and-answer topic units; Keyword density is calculated based on a pre-set financial information keyword database, as well as the semantic cohesion of the speech vectors in the question-and-answer topic units and the proportion of discussion time. The importance score of each question-and-answer topic unit is calculated by comprehensively considering its semantic cohesion, keyword density, discussion time percentage, and structural integrity verification results. Step 4: Multidimensional quantification of the semantic gap between question and answer; For each question-answer pair in a question-answer topic unit, a large language model is used to score them from three dimensions: attitude gap, uncertainty gap, and semantic alignment, to obtain the question-answer semantic gap vector of the question-answer topic unit. A dual-model cross-validation mechanism is used to validate the question-answer semantic gap vector, and the final question-answer semantic gap vector for each question-answer topic unit is determined based on the validation results. Step 5, Calculation of the weighted aggregate Trust and Transparency Index (TGI): The semantic alignment in the final question-answer semantic gap vector is converted into a gap metric to obtain the alignment gap. Then, based on the weighted sum of the attitude gap, uncertainty gap and alignment gap in the final question-and-answer semantic gap vector, the comprehensive gap score of each question-and-answer topic unit is obtained; The importance score of each Q&A topic unit is normalized and used as the weight of its comprehensive gap score. The global weighted gap index of financial Q&A long text is obtained by weighting the comprehensive gap scores of all Q&A topic units. The globally weighted gap index is transformed into a TGI by using an exponential decay mapping function, and the evaluation information of the long text of financial Q&A is visualized and output based on the TGI.

2. The method as described in claim 1, characterized in that, In step 1, the three-level role matching is specifically as follows: Level 1 exact match: Extract the speaker identifier field and perform a string-based exact match with the names registered in the role mapping table. If the match is successful, the corresponding role tag is directly assigned; if the match fails, the level 2 fuzzy match is executed. Second-level fuzzy matching: Calculate the edit distance between the speaker's identifier and each name in the role mapping table. If the minimum edit distance is less than the preset threshold and is unique, then assign the role label corresponding to the minimum edit distance to the current speech segment. Otherwise, perform third-level semantic inference; Level 3 semantic inference: The large language model is called to analyze the language features of the current speech segment, and role inference is performed based on the language features. The role inference results are recorded in the role mapping table.

3. The method as described in claim 1, characterized in that, In step 2, candidate points for topic boundaries are identified based on semantic similarity, explicit connection signals, and role-switching patterns, including: Similarity drop detection: When the semantic similarity between adjacent speech segments is lower than the set dynamic threshold, the adjacent positions of the adjacent speech segments are recorded as candidate points of the topic boundary. The dynamic threshold is determined by analyzing the distribution characteristics of the similarity between all adjacent speech segments in the long text of financial Q&A. Explicit transition signal detection: Explicit transition words in the speech segment are matched using regular expressions, and the positions of the matched explicit transition words are marked as candidate points for topic boundaries; Role transition pattern detection: When the speaker's role sequence shows a transition pattern from analyst to management and back to analyst, mark the position where the analyst reappears and speaks in this transition pattern as a candidate point for the topic boundary.

4. The method as described in claim 1, characterized in that, In step 3, performing structural integrity verification on the question-and-answer topic units includes: Test 1: Check whether the Q&A topic unit contains at least one statement from an analyst and one statement from a manager; Test 2: Check whether the speaking sequence in the question-and-answer topic unit conforms to the logical order of questions and answers; If a question-and-answer topic unit satisfies both Detection 1 and Detection 2, then the question-and-answer topic unit is marked as structurally complete; otherwise, it is marked as structurally incomplete.

5. The method as described in claim 1, characterized in that, In step 3, the formula for calculating the importance score is: ； in, , , The first The semantic cohesion, keyword density, and discussion duration of each Q&A topic unit; , , These are the weighting coefficients for semantic cohesion, keyword density, and discussion duration, respectively, and they satisfy the following conditions: ; The structural integrity verification coefficient is set to the following value: if the question-and-answer topic unit structure is complete, then... Set to 1; if the question-and-answer topic unit structure is incomplete, then .

6. The method as described in claim 1, characterized in that, In step 4, the attitude gap, uncertainty gap, and semantic alignment are set as follows: The attitude gap is set as follows: ,in, and These are the sentiment scores for management's responses and analysts' questions, respectively, obtained based on a large language model. , To embellish the penalty coefficient; The uncertainty gap is set as follows: ,in, and These represent the degree of uncertainty in analyst questions and the degree of uncertainty in management responses, respectively, obtained through a large language model; and ; Semantic alignment The setting is: the semantic alignment obtained by semantic matching analysis of question-answer pairs in question-answer topic units using a large language model, and .

7. The method as described in claim 1, characterized in that, In step 4, the question-answer semantic gap vector is validated using a dual-model cross-validation mechanism, specifically including: definition , and These represent the attitude gap, uncertainty gap, and semantic alignment, respectively. By calling two large language models with different architectures or different training parameters in parallel to score the same question-answering topic unit, two sets of question-answering semantic difference vectors for that topic unit are obtained: and ; Calculate the Euclidean distance between the two sets of question-answer semantic gap vectors as a consistency metric: ； in, , and The superscript is used to distinguish different large language models; If consistency index Less than or equal to the preset threshold If the mean of the two sets of question-and-answer semantic gap vectors is used, the final question-and-answer semantic gap vector of the current question-and-answer topic unit is obtained; otherwise, a third-party arbitration mechanism or manual review process is triggered to determine the final question-and-answer semantic gap vector of the current question-and-answer topic unit.

8. The method as described in claim 1, characterized in that, Step 4 also includes a reverse verification mechanism, specifically: definition Indicates the first The question and answer text for each question and answer topic unit; After obtaining the question-answer semantic gap vector of the question-answer topic unit using a large language model, reverse verification prompts are constructed, and the textual representation of the question-answer pair of the current question-answer topic unit is regenerated using the large language model based on the obtained question-answer semantic gap vector. ; Then and Perform semantic similarity comparison. If the semantic similarity is lower than or equal to a preset threshold, then re-apply the large language model to obtain the next semantic similarity. The semantic gap vector between the question and answer of each question-and-answer topic unit, or the semantic gap vector of the first question-and-answer unit, can be re-determined through manual review. The semantic gap vector between questions and answers for each question-and-answer topic unit.

9. The method as described in claim 1, characterized in that, In step 5, the formula for calculating TGI is: ； in, This represents the natural exponential function. These are preset sensitivity adjustment parameters used to control the sensitivity of TGI to changes in the gap. It is a globally weighted gap indicator.

10. A financial question-and-answer long text analysis system based on role recognition and dialogue structure parsing, characterized in that: It includes a role recognition module, a question and answer topic unit recognition module, a quality control module, a gap quantification module, an index aggregation module, and a display module; in, Role Recognition Module: Extracts the preceding character range from the long financial Q&A text of the target object, and locates the opening paragraph containing participant introduction sentences through regular expression pattern matching; calls a large language model to perform structured parsing of the opening paragraph, extracts the list of participants to construct a role mapping table; performs three-level role matching on each speech segment of the long financial Q&A text based on the constructed role mapping table, determines the role label of each speech segment based on the matching results, and outputs the role label of each speech segment to the topic unit recognition module; if there is role inference in the matching results, the role mapping table is updated based on the role inference; The question-and-answer topic unit identification module generates semantic vector representations of each speech segment using a semantic vector model in the financial field and calculates the semantic similarity between adjacent speech segments; it identifies candidate points for topic boundaries based on semantic similarity, explicit cohesion signals, and role switching patterns; it calls a large language model to verify the semantic coherence of the candidate points for topic boundaries, and aggregates semantically coherent speech segments into question-and-answer topic units; it outputs a structured list of question-and-answer topic units to the quality control module, the gap quantification module, and the display module, respectively. The quality control module performs structural integrity verification on each Q&A topic unit in the Q&A topic unit list, obtaining the structural integrity verification result; calculates keyword density based on a preset financial information keyword library, and calculates the semantic cohesion of the speech vectors and the proportion of discussion time for each Q&A topic unit; comprehensively calculates the importance score of each Q&A topic unit based on the semantic cohesion, keyword density, proportion of discussion time, and structural integrity verification result; and outputs the importance score of each Q&A topic unit to the index aggregation module and the display module. The gap quantification module calls a large language model to perform multi-dimensional scoring of the question-answer pairs for each question-answer topic unit, including attitude gap, uncertainty gap, and semantic alignment, to obtain the question-answer semantic gap vector for the question-answer topic unit. A dual-model cross-validation mechanism is used to validate the question-answer semantic gap vector. Based on the validation results, the final question-answer semantic gap vector for each question-answer topic unit is determined and output to the exponential aggregation module. The index aggregation module converts the semantic alignment in the final question-and-answer semantic gap vector into a gap metric to obtain the alignment gap. Then, based on the weighted sum of the attitude gap, uncertainty gap, and alignment gap in the final question-and-answer semantic gap vector, it obtains the comprehensive gap score for each question-and-answer topic unit. The importance score of each question-and-answer topic unit is normalized and used as the weight for its comprehensive gap score. Based on the weighted sum of the comprehensive gap scores of all question-and-answer topic units, it obtains the global weighted gap index for the long financial question-and-answer text. Finally, through an exponential decay mapping function, the global weighted gap index is converted into a Trust and Transparency Index (TGI), and the comprehensive gap score and TGI are output to the display module. The display module is used to visualize the processing results of the module to form evaluation information of the long text of financial Q&A to be evaluated, including: displaying the question text, answer text and comprehensive gap score of each individual Q&A topic unit in chronological order; drawing a time series chart of TGI to show the TGI change trend of the same target object in past meeting records; and drawing a bar chart comparing the TGI of different target objects in the same industry.

Images