A large model-based report generation method, device, medium and equipment

Abstract

The present application relates to the technical field of report generation, and in particular to a report generation method and device based on a large model, medium and equipment, on the basis of collecting full-quantity interaction round interaction data, the time sequence semantic reasoning ability of the preset large model is used to judge the problem deep evolution direction, so that the interaction report can reveal from the dynamic perspective whether the user demand is deepening or generalizing, and the round order is converted into the information adoption degree according to the problem deep trend, so that the information value evaluation conforms to the user real demand evolution logic, avoids the core information being submerged caused by the traditional average allocation weight, finally through the deep fusion of the interaction data, the information adoption degree and the preset report template, the content generation and structure arrangement are carried out by using the preset large model, so that the interaction report can not only retain the original details of user interaction, but also highlight the key information value, and at the same time realize the format standardization and insight deepening, obtain high-quality interaction report with high pertinence and high professionalism.

Description

Technical Field

[0001] This invention relates to the field of report generation technology, and in particular to a method, apparatus, medium and device for generating reports based on large models. Background Technology

[0002] With the maturation of large model technology in the field of natural language generation, report generation methods based on large models have been widely applied in scenarios such as customer service dialogue summaries, medical consultation records, and corporate meeting minutes.

[0003] Existing large-model-based report generation methods typically employ a single-input generation model. For example, in patent CN119203966A, entitled "A Large-Model Intelligent Report Generation Method, Device, Electronic Device, and Medium," a structured query statement is generated based on the natural language dialogue of the target object, using a large language model combined with metadata. An intelligent report request is then generated based on the structured query statement, and a template engine generates the report text based on a preset template. Because this approach only processes single-turn queries or isolated text, it fails to capture the semantic evolution logic in multi-turn interactions. Furthermore, it relies heavily on fixed templates to fill content, cannot adjust the content structure according to the depth and trend of user interaction, and lacks quantitative analysis of the information value during the interaction process, resulting in insufficient professionalism and relevance in the reports.

[0004] Therefore, when generating reports based on large models, improving the relevance and professionalism of the reports in summarizing user interactions has become an urgent problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, the present invention provides a report generation method based on a large model, which includes the following steps: S100, acquire the interaction data between the target client and the query platform for several interaction rounds, wherein the interaction data for each interaction round includes the target query statement input by the target client into the pre-trained preset large model in the query platform, and the target analysis result text of the preset large model for the target query statement.

[0006] S200: Based on all target query statements corresponding to all interaction rounds, obtain the question depth trend corresponding to the target client, where the question depth trend is either positive or negative.

[0007] S300 maps the information adoption rate for each interaction round based on the question depth trend corresponding to the target client.

[0008] S400 inputs the interaction data, the information adoption rate for each interaction round, and the preset report template into the preset large model to obtain the interaction report between the target client and the query platform.

[0009] The present invention also provides a report generation apparatus based on a large model, the report generation apparatus based on a large model comprising: The data interaction module is used to acquire the interaction data between the target client and the query platform for several interaction rounds. The interaction data for each interaction round includes the target query statement input by the target client into the pre-trained preset large model in the query platform, and the target analysis result text of the preset large model for the target query statement.

[0010] The trend analysis module is used to obtain the question depth trend of the target client based on all target query statements corresponding to all interaction rounds. The question depth trend can be positive or negative.

[0011] The degree mapping module is used to map the degree of information adoption for each interaction round based on the question depth trend of the target client.

[0012] The report generation module is used to input interaction data, the degree of information adoption for each interaction round, and a preset report template into a preset large model to obtain an interaction report between the target client and the query platform.

[0013] The present invention also provides a non-transitory computer-readable storage medium storing at least one instruction or at least one program, wherein the at least one instruction or at least one program is loaded and executed by a processor to implement any of the above-mentioned large model-based report generation methods.

[0014] The present invention also provides an electronic device, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0015] This invention has at least the following beneficial effects: By collecting user query and platform response data from all interaction rounds, the interaction data retains the consistency of user identity while dynamically adapting to changes in the problem domain, providing high-quality interaction data for analyzing the dynamic changes in user needs. Based on the target query statements of all interaction rounds, the temporal semantic reasoning capability of the preset large model is used to determine the direction of problem depth evolution, enabling the interaction report to reveal from a dynamic perspective whether user needs are deepening or generalizing. Furthermore, the round order is converted into information adoption degree based on the problem depth trend, ensuring that information value assessment aligns with the evolution logic of users' actual needs, avoiding the burying of core information due to traditional average weighting. Finally, by deeply integrating interaction data, information adoption degree, and the preset report template, and using the preset large model for content generation and structural arrangement, the interaction report retains the original details of user interactions while highlighting the value of key information, achieving both format standardization and in-depth insight, providing highly targeted and professional high-quality interaction reports for the target client. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of a report generation method based on a large model provided in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of a report generation device based on a large model provided in Embodiment 2 of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It is understood that, where appropriate, the terms used to distinguish similar objects can be interchanged so that the invention can also be implemented in other embodiments besides the illustrated or described embodiments. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0020] Example 1 like Figure 1 As shown in the figure, this embodiment provides a report generation method based on a large model, which includes the following steps: S100, acquire the interaction data between the target client and the query platform for several interaction rounds, wherein the interaction data for each interaction round includes the target query statement input by the target client into the pre-trained preset large model in the query platform, and the target analysis result text of the preset large model for the target query statement.

[0021] The target client is the terminal where the user initiates the query, such as a mobile app or web interface. It serves as the entry point for human-computer interaction, converting the natural language input by the user into a processable digital signal.

[0022] The target query statement is the specific question text submitted by the user to the query platform through the target client, and it is the raw input for semantic parsing of the pre-set large model.

[0023] The query platform is a system carrier that integrates large models, tag libraries, and knowledge bases. It provides the hardware and software environment for query processing and coordinates the collaborative operation of various task modules such as tag generation and knowledge base matching.

[0024] The pre-trained large-scale model can be an artificial intelligence model trained on text generation, capable of understanding natural language instructions and generating text that conforms to formatting requirements. Those skilled in the art will recognize that any existing large-scale language model and its training method fall within the protection scope of this invention, such as the GPT series, ERNIE, etc., and will not be elaborated upon here.

[0025] The target analysis result text is a structured answer generated by the query platform's preset large model for the target query statement. It is the direct output of the query platform in response to user needs.

[0026] An interaction round is a continuous question-and-answer unit between a user and the query platform through the target client. It begins when the user initiates the first query and ends when the user stops interacting, and includes at least one set of target query statements and corresponding target analysis result text. As the smallest sequential unit of report analysis, the interaction round reflects the dynamic changes in user needs through the content relationships between rounds, avoiding the one-sidedness caused by interaction reports being based solely on isolated question-and-answer responses.

[0027] As described above, by recording the question-and-answer content of all interaction rounds, the information gaps caused by relying solely on single-round question-and-answer reports are avoided, providing high-quality interaction data for analyzing the dynamic changes in user needs.

[0028] In one specific embodiment, S100 includes the following steps: S110, input the first target query statement of the target client into the pre-trained large model in the query platform, and obtain the first label and the second label corresponding to the first target query statement. The first label is used to represent the identity domain corresponding to the first target query statement, and the second label is used to represent the question domain corresponding to the first target query statement.

[0029] S120, the first tag is matched with the preset tag library in the query platform to obtain the third tag corresponding to the first target query statement. The preset tag library includes the mapping relationship between the first tag and the third tag, and each third tag corresponds to several query knowledge bases in the query platform.

[0030] S130, input the first target query statement and the first label, second label and third label corresponding to the first target query statement into the preset large model, and obtain the first target analysis result text corresponding to the first target query statement.

[0031] S140, Obtain the (j+1)th target query statement from the target client for the analysis result text of the j-th target, where j > 0.

[0032] S150, input the (j+1)th target query statement into the preset large model, and obtain the second label corresponding to the (j+1)th target query statement.

[0033] S160, input the (j+1)th target query statement, the second label corresponding to the (j+1)th target query statement, and the first and third labels corresponding to the first target query statement into the preset large model to obtain the (j+1)th target analysis result text corresponding to the (j+1)th target query statement.

[0034] S170, update j=j+1, repeat step S140 until the target client ends its interaction with the query platform, and determine the target query statement and target analysis result text between the target client and the query platform for each interaction round as the interaction data between the target client and the query platform for each interaction round.

[0035] In particular, since the focus of attention for the same problem, similar problems, or the same problem may differ significantly due to different user identities, the pre-set large model needs to provide targeted answers based on the identity information of the user who puts out the target query, in order to improve the quality of information interaction between the user and the large model.

[0036] Therefore, a pre-defined large model is used to perform dual-channel analysis of the deep semantics of the target query statement, enabling simultaneous identification of user identity and question type. Specifically, the pre-defined large model extracts identity identifiers from the target query statement using an identity feature library learned through training on massive amounts of text. At the same time, relying on question domain classification capabilities, it identifies the core theme of the query and finally outputs structured first and second labels, laying the foundation for accurate matching with the subsequent knowledge base.

[0037] The first tag is used to accurately characterize the identity attributes of the user who asked the question, including core features such as occupation, role, and permission level. It is a key basis for identifying the user's background and potential needs. The identity dimension can be used to narrow down the service scope, ensuring that the subsequent response of the pre-set model is consistent with the user's professional background, knowledge base, or permission restrictions. For example, the first tag can be a professional identity such as financial analyst, corporate employee, pediatrician, high school teacher, or software engineer; a role identity such as student, patient, or corporate manager; or a permission identity such as VIP user, internal employee, or visitor.

[0038] The second tag is used to define the subject matter of the user's question, focusing on the professional field, business scope, or knowledge category involved in the question itself. It is the core basis for matching the corresponding knowledge base and professional resources. The subject dimension can be used to lock in the scope of information, ensuring the professionalism and relevance of the subsequent response content of the pre-set model. For example, the second tag can be a subject area such as finance, education, and healthcare; a business area such as human resources, law, and technology; or a scenario area such as product consultation, troubleshooting, policy interpretation, and process handling.

[0039] The default tag library is a database that stores tag mapping rules and knowledge base relationships. It is built using a graph database structure (such as Neo4j) and supports efficient tag association queries. It is used to abstract user identities into knowledge base access paths. The query knowledge base is a collection of structured knowledge units organized by domain and topic. It is stored using a vector database (such as Milvus) and supports semantic retrieval.

[0040] The third tag is an intermediate semantic layer set up by the query platform to connect user identity (first tag) and knowledge resources (query knowledge base). It is used to map the user's abstract identity characteristics to specific knowledge access paths. Specifically, the preset tag library contains a mapping table that maps the first tag to the third tag. For example, financial analyst is mapped to financial research, individual investor is mapped to investment and wealth management, bank loan officer is mapped to credit risk control, patient is mapped to health consultation, and education administrator is mapped to education decision-making.

[0041] A third-party tag corresponds to several knowledge bases within the query platform. For example, "financial research" corresponds to multiple knowledge bases such as the macroeconomic database, industry analysis report database, financial policy and regulation database, and academic journal article database for the financial sector. "Investment and wealth management" corresponds to multiple knowledge bases such as the fund product database, stock market database, and financial planning guide database. "Credit risk control" corresponds to multiple knowledge bases such as the enterprise credit database, industry risk assessment model database, and credit approval case database. "Health consultation" corresponds to multiple knowledge bases such as the disease science database, medication guidance database, and medical process guide database. "Education decision-making" corresponds to multiple knowledge bases such as the education statistics database, policy and regulation database, and school management case database.

[0042] A query knowledge base within a query platform can also be referenced by multiple third-party tags. For example, the basic medical database can correspond to both the clinical diagnosis and treatment and medical education third-party tags.

[0043] Based on the first tag of the target query statement, multiple query knowledge bases can be mapped. By filtering by identity of the first tag, the query scope can be narrowed from the entire knowledge base to a subset strongly related to the user's identity. Furthermore, relevant knowledge resources scattered across multiple databases can be integrated into a set of query knowledge bases oriented towards a specific identity, avoiding blind searching in the entire knowledge base.

[0044] The mapping relationship between the first and third tags can be generated through training on historical interaction data between users and a pre-set large model. The correspondence between the third tag and the query knowledge base can be set by the implementer in the query platform according to the actual situation.

[0045] The target query statement is fused with three types of tags and input into a pre-defined large model. The target query statement provides the basic semantics ("what to query"), the first tag (user identity) constrains the output perspective ("from whose perspective to answer"), the second tag (query topic) limits the domain scope ("within which professional field to answer"), and the third tag (knowledge scenario) anchors the knowledge source ("based on which knowledge bases the content is generated"). This achieves accurate text generation under the four-dimensional constraints of "user specific needs, user identity, query topic, and knowledge scenario resources". Furthermore, the second and subsequent target queries can directly use the first tag representing the user identity and the third tag anchoring the knowledge source corresponding to the first target query statement, thereby eliminating redundant steps and improving data acquisition efficiency.

[0046] The pre-trained large model transforms the four types of inputs mentioned above into structured generation rules through its pre-trained domain adaptation capabilities. Specifically, it prioritizes calling knowledge base resources associated with the third tag, and logically decomposes and deeply analyzes the target query statement from the professional perspective corresponding to the first tag, around the topic boundary of the second tag. The generated target analysis result text simultaneously meets four conditions: "semantic fit to query requirements, perspective matching user identity, content covering the topic range, and arguments sourced from the associated knowledge base." This avoids generating irrelevant or unauthorized information, thereby improving the quality of information interaction between the target client and the large model.

[0047] Correspondingly, the target analysis result text is a structured analysis content that directly responds to the user's query needs, generated by a pre-set large model based on the collaborative input of the target query statement, the first tag, the second tag, and the third tag, after domain knowledge reasoning and multi-dimensional constraints.

[0048] In this embodiment, based on the interaction pattern of users' subsequent queries depending on previous answers, the system captures users' feedback or in-depth needs for the j-th analysis result, ensuring the semantic relevance between subsequent queries and previous answers. Through the association records from the j-th result to the (j+1)-th query, a continuous semantic chain is provided for analyzing the evolution trend of user needs, avoiding the mixing of irrelevant queries into the same interaction process.

[0049] Subsequent queries by users may be related to or partially extended from the initial domain. Therefore, the second tag needs to be updated to adapt to the new domain, while the first tag representing the user's identity and the third tag anchoring the knowledge source remain unchanged. By reusing the first and third tags, the stability of identity and knowledge context is ensured. By updating the second tag, the response is adapted to new needs, forming a response mode of stable framework + dynamic content, avoiding the disconnect between subsequent answers and the initial identity or knowledge domain.

[0050] The specific conditions for the target client to end its interaction with the query platform can be set by the implementer according to the actual situation. For example, the interaction could end when the target client enters a preset end text, or when the target client does not enter any new question text within a preset time period. The interaction data for each interaction round is then constructed using the target query statement and target analysis result text for each interaction round. The preset end text and preset time period can be set by the implementer according to the actual situation. For example, the preset end text could be texts such as "Understood, thank you" or "I understand, thank you," and the preset time period could be 3 minutes, 5 minutes, 10 minutes, etc.

[0051] As described above, the two-dimensional positioning of user identity and question topic is formed by the first and second tags, and their synergistic effect significantly narrows the scope of knowledge base retrieval. Through a mapping mechanism that maps the first tag in the pre-set tag library to the third tag in the query platform, a precise association is achieved between user identity and the query knowledge base. This narrows the retrieval scope from the entire knowledge base to a subset of identity-related knowledge bases, and integrates relevant knowledge resources scattered across multiple databases into a query knowledge base set oriented towards a specific identity. This significantly improves the retrieval efficiency and analytical accuracy of the pre-set large model. By collaboratively inputting four types of inputs into a pre-defined large model, the model generates target analysis result text that simultaneously satisfies four conditions under the four-dimensional constraints of "user specific needs, user identity, query topic, and knowledge scenario resources": "semantic fit with query needs, perspective matching user identity, content covering the scope of the topic, and arguments sourced from a related knowledge base." This ensures that the analysis results align with the user's cognitive level and professional needs, and avoids generating generalized content that is detached from actual requirements. This improves the quality of information interaction between the target client and the large model. Furthermore, it iterates based on the interaction pattern of subsequent user queries depending on previous answers. During the iteration process, the first and third tags are reused to ensure the stability of identity and knowledge scenario, while the second tag is updated to ensure the response adapts to new needs. This forms a response mode of stable framework + dynamic content, preventing subsequent answers from becoming disconnected from the initial identity or knowledge domain. This ensures high-quality interaction data and provides a high-quality data foundation for generating in-depth, logical, and targeted reports.

[0052] S200: Based on all target query statements corresponding to all interaction rounds, obtain the question depth trend corresponding to the target client, where the question depth trend is either positive or negative.

[0053] The question depth trend refers to the direction of change in the semantic complexity and professional depth of user questions across multiple rounds of interaction, including both positive and negative trends. A positive trend indicates that user questions progress from surface information to deeper needs, meaning the questions become increasingly in-depth, specific, and professional, such as from "What is hypertension?" to "Medication contraindications for hypertension patients." Conversely, a negative trend indicates that user questions degenerate from deeper needs to surface information, meaning the questions become increasingly superficial, broad, and basic, such as from "Specific treatment plans" to "Disease definition." The question depth trend can be used to quantify the changes in the depth value of information during user interaction, providing a quantitative basis for the degree of information adoption corresponding to each round of interaction.

[0054] In one specific embodiment, S200 includes the following steps: S210, obtain the round order of the target query statement corresponding to each interaction round.

[0055] S220: Input all target query statements corresponding to all interaction rounds, the round order of each target query statement, and preset prompt words into the preset large model to obtain the question depth trend corresponding to the target client.

[0056] The round order refers to the sequential numbering of each interaction round within the entire interaction process. This serves as a timeline for semantic analysis, enabling the pre-defined model to determine the direction of the user's question's progression. The form of the round order can be set by the implementer based on the actual situation; for example, the round order could be round one, round two, etc.

[0057] Preset prompts are pre-defined instruction texts used to guide a pre-defined large model to complete a trend judgment task. In this step, the preset prompts can be "Based on the input target query and the corresponding round order, determine the depth trend of the question corresponding to the target client and output a positive or negative trend result."

[0058] The above-mentioned method, by recording the order of rounds, provides a time-series anchor point for trend analysis. By combining the semantic connotation of the target query statement with the time-series relationship of the round order, the trend direction of the question depth can be determined. This solves the problem of trend misjudgment caused by only looking at the content and not the order in traditional keyword analysis. It also outputs positive and negative results to reflect the evolution of users' real needs, and provides key behavioral feature data for the degree of information adoption corresponding to each subsequent interaction round.

[0059] S300 maps the information adoption rate for each interaction round based on the question depth trend corresponding to the target client.

[0060] Among them, the degree of information adoption is the likelihood that a user will accept the information provided by the query platform in a certain interaction round. The higher the degree of information adoption, the more likely the information in that interaction round is to be adopted by the user and influence their decision. This allows high-adoption-degree interaction content to be prioritized for display or focused analysis in report generation, thus solving the limitations of traditional methods that distribute information weight equally or only look at the effect of a single round of interaction.

[0061] In one specific embodiment, S300 includes the following steps: S310: Based on the problem depth trend of the target client, obtain the third mapping rule corresponding to the target client.

[0062] S320, according to the third mapping rule, maps the turn order of each target query statement to the degree of information adoption.

[0063] In one specific embodiment, S310 includes the following steps: S311, if the problem depth trend corresponding to the target client is positive, then the round order and the degree of information adoption in the third mapping rule corresponding to the target client are positively correlated.

[0064] S312, if the problem depth trend corresponding to the target client is reversed, then the round order and information adoption degree in the third mapping rule corresponding to the target client are negatively correlated.

[0065] The third mapping rule defines the correspondence between the round order and the degree of information adoption. It is determined by the trend of question depth and is used to ensure that the calculation of the degree of information adoption conforms to the evolution logic of user needs.

[0066] Specifically, under a positive trend, user needs gradually focus, and later interactions are closer to the core needs. Therefore, the later the round order, the higher the degree of information adoption. The positive correlation mapping under the positive trend enables the interaction report to focus on the core information that users ultimately care about, avoiding interference from the preliminary content.

[0067] When user needs diverge or deviate under a reverse trend, early interactions are more likely to reflect core needs. Therefore, the earlier the interaction sequence, the higher the degree of information adoption. By using negative correlation mapping, the information value of early interactions can be preserved, and interaction reports can avoid ignoring core needs due to user deviations in later stages.

[0068] The third mapping rule transforms the round order into a quantifiable degree of information adoption. For example, under a positive trend, the degree of information adoption X1 = 1 / (1 + exp(-B)), and under a negative trend, the degree of information adoption X1 = exp(-B), where X1 is the degree of information adoption, B is the round order, and exp() is an exponential function with base e. Correspondingly, the range of the degree of information adoption is (0, 1).

[0069] The above describes a dynamic mechanism that uses trends to determine mapping rules and transforms the order of rounds into quantifiable levels of information adoption. In a positive trend, positive correlation mapping ensures that key information in the process of deepening user needs is highly adopted, avoiding interference from preliminary content. In a negative trend, negative correlation mapping preserves the information value of early core user needs, preventing core information from being ignored due to later deviations. This allows interaction reports to accurately capture the core content that users ultimately focus on, thereby improving the relevance and professionalism of the report's summary of user interactions.

[0070] S400 inputs the interaction data, the information adoption rate for each interaction round, and the preset report template into the preset large model to obtain the interaction report between the target client and the query platform.

[0071] Interaction data serves as the raw material for interaction reports, providing specific content on the platform's answers to questions from target clients. The degree of information adoption acts as a value filter for interaction reports, guiding the pre-set large model to prioritize displaying interaction content with high adoption rates and avoiding interference from irrelevant information. The pre-set report template serves as the structural skeleton of interaction reports, defining the chapter framework and content focus of interaction reports. For example, financial reports focus on investment strategies, while medical reports focus on treatment suggestions, enabling interaction reports to both retain the original details of user interactions and highlight the value of key information.

[0072] Pre-designed report templates are pre-designed framework documents that define the report structure and content direction. They include fixed sections (such as interaction overview, requirements analysis, conclusions and recommendations) and variable fill areas (such as user issue statistics, trend charts, etc.) to ensure that interaction reports from different clients have a unified format, making them easy to read and compare.

[0073] Correspondingly, the interaction report is an analysis document generated by the pre-set large model based on interaction data, information adoption level and pre-set template. It includes user needs interpretation, information value assessment, trend prediction and other content. It provides users with an interaction summary to improve user experience, and provides platform operators with user behavior insights to guide content optimization.

[0074] The above-mentioned content selection based on the degree of information adoption ensures that the interaction report focuses on the core information that users really care about, avoids bias caused by subjective selection, and through the semantic analysis capabilities of the pre-set large model, it can explore potential patterns in the interaction data and provide high-quality interaction reports that are highly targeted and professional for the target client.

[0075] In one specific implementation, the report generation method based on large models further includes the following steps: S10, for any first tag, obtain the interaction data of the current first tag for several interaction rounds for several preset users. The interaction data of each interaction round includes several sets of interaction text between a preset user and a preset large model. Each set of interaction text includes the question text of the preset user and the answer text of the preset large model for the question text.

[0076] S20, extract keywords for each set of interactive text to obtain a set of interactive keywords corresponding to each set of interactive text.

[0077] S30: Analyze all groups of interactive keywords corresponding to each interaction round to obtain the question depth trend of the preset user corresponding to each interaction round, wherein the question depth trend is positive or negative.

[0078] S40: Based on the question depth trend corresponding to all interaction rounds, obtain the influence weight of each interaction keyword corresponding to the current first tag.

[0079] S50: Based on the influence weight of each interactive keyword corresponding to the current first tag and the number of times each interactive keyword appears, obtain the target keyword corresponding to the current first tag.

[0080] S60: Map the third tag corresponding to the target keyword in the preset tag library to the third tag corresponding to the current first tag.

[0081] S70, iterate through all the first tags and get the third tag corresponding to each first tag.

[0082] The interaction rounds are a continuous interactive process between the same pre-defined user and a pre-defined large model around a specific topic, encompassing the complete cycle from the initial question to the end of the dialogue. Since different users displaying the same identity label to the pre-defined large model will exhibit commonalities in their question types and focus areas when interacting with it, these commonalities can be extracted from the interaction data. Therefore, analyzing the dynamic changes in user needs through continuous dialogue provides time-series data for analyzing the depth and trends of questions.

[0083] The preset users are typical users selected to train the label mapping relationship, covering the identity domain corresponding to the first label. They are used to extract common patterns from the interaction data of typical users, ensuring the universality of the label mapping rules and avoiding interference with analysis results due to abnormal interactions of individual users. In this embodiment, representative interaction samples are selected to ensure that the data covers common interaction scenarios of users with this identity, thus avoiding bias in subsequent analysis due to a single sample.

[0084] By utilizing keyword extraction techniques from natural language processing, non-core information is extracted from question and answer texts, retaining only core vocabulary that reflects the interaction theme. This ensures that the extracted keywords simultaneously cover both user needs and model responses, avoiding one-sided information extraction from only the question or answer side. Those skilled in the art will recognize that any existing keyword extraction technique falls within the scope of this invention, such as TF-IDF, TextRank, and the BERT keyword model, and will not be elaborated upon here.

[0085] In this embodiment, the question depth trend can characterize the direction of change in the semantic complexity and professional depth of user questions in an interaction round, including positive and negative trends. This is used to quantify the changes in the depth value of information during user interaction and to provide a basis for keyword weight calculation.

[0086] Influence weights are used to characterize the contribution of interactive keywords to the mapping relationship between the first and third tags. Their function is to filter redundant keywords, ensuring that target keywords accurately reflect the core characteristics of the identity domain.

[0087] Target keywords are the most representative keywords within a given primary tag, considering both their influence weight and frequency of occurrence. They serve as a bridge between the primary and secondary tags, improving tag matching accuracy. For example, target keywords extracted from the interaction data of a predefined user group with the primary tag "teacher" could be "lesson plan design" or "classroom management."

[0088] The preset tag library stores the mapping relationship between keywords and third tags. By matching the target keyword with the preset tag library, the third tag corresponding to the first tag is obtained, which provides a basis for limiting the scope of professional fields.

[0089] The above-mentioned data collection through multiple preset users and multiple interaction rounds ensures the representativeness of the analysis samples, providing a reliable foundation for subsequent pattern extraction and avoiding mapping bias caused by partial data. Improving the extraction of interaction keywords simplifies the text analysis dimension, transforming complex natural language into quantifiable core vocabulary, providing structured input for subsequent calculations. Increasing the introduction of question depth trends captures the dynamic changes in user needs, making subsequent analysis more closely aligned with the actual scenarios where user needs change with communication in real interactions, thus improving the accuracy of keyword importance judgment. Increasing the quantification of keyword influence weights avoids filtering keywords solely based on frequency of occurrence, ensuring the retention of core words that substantially impact user needs. Combining influence weights and frequency of occurrence to select target keywords balances universality and the importance of driving deeper understanding of needs, enabling target keywords to accurately represent the core concerns of users with different identities. These keywords also serve as an intermediate bridge between the first and third tags, improving tag matching accuracy and providing a basis for further defining the scope of professional fields. This ultimately improves the domain adaptability and response efficiency of the target analysis results text for users with different identities during information interaction.

[0090] In one specific embodiment, S10 includes the following steps: S101, for any interaction round corresponding to any preset user corresponding to the current first label, obtain the first answer text corresponding to the preset large language model based on the first question text input by the current preset user into the preset large model.

[0091] S102, based on the current preset user's input of the i-th answer text into the preset large model for the (i+1)-th question text, obtain the (i+1)-th answer text corresponding to the preset large language model, where i > 0.

[0092] S103, update i=i+1, repeat step S102 until the current preset user ends the interaction with the current interaction round of the preset large language model, and determine all the question text and answer text of the current first tag for the current preset user's current interaction round as the interaction data of the current first tag for the current preset user's current interaction round.

[0093] S104, iterate through all the preset users and all the interaction rounds corresponding to the current first label, and obtain the interaction data of the current first label for several preset users and several interaction rounds.

[0094] The first question from the user is used as the starting point of the round. The initial dialogue context is established by the first answer of the pre-set big model. Subsequent questions from the user depend on the previous answers of the pre-set big model, forming multiple rounds of question-and-answer dialogues. This constitutes a semantic progression chain within the round, providing temporal dimension interaction data for capturing the dynamic evolution of user needs.

[0095] The specific conditions for the user to end the current interaction round with the preset large language model can be set by the implementer according to the actual situation. For example, it can be that the user enters a preset end text, or that the user does not enter a new question text within a preset time period, in which case the interaction round is determined to end, and all question-and-answer pairs in the interaction round constitute complete interaction data.

[0096] In one specific embodiment, S30 includes the following steps: S301, for any interaction round, obtain the interaction order of the interaction text corresponding to each group of interaction keywords in the current interaction round.

[0097] S302, input all groups of interaction keywords corresponding to the current interaction round, the interaction order corresponding to each group of interaction keywords, and the preset prompt words into the preset large model to obtain the question depth trend of the preset user corresponding to the current interaction round.

[0098] Specifically, the interaction sequence is used to accurately mark the position of each group of interaction keywords within a round, ensuring that the temporal relationship is not confused. The form of the interaction sequence can be set by the implementer according to the actual situation; for example, the interaction sequence can be group one, group two, etc.

[0099] The pre-built large model possesses temporal semantic reasoning capabilities. By combining the semantic connotations of interactive keywords with the temporal relationship of the interaction sequence, it can determine the trend direction of question depth. Specifically, the pre-trained large model uses pre-trained domain knowledge to understand the depth levels of keywords, and then, based on the changes in depth levels with the interaction sequence, outputs positive or negative conclusions. Specifically, the question depth for keywords related to "definition" can be set to shallow, for keywords related to "application" to medium, and for keywords related to "principle derivation" to deep. Correspondingly, when the pre-built large model analyzes the question depth from shallow to deep, the corresponding question depth trend is positive; conversely, the corresponding question depth trend is negative.

[0100] Preset prompts are pre-defined instruction texts used to guide a pre-defined large model to complete a trend judgment task. For example, a preset prompt might be "Based on the input keywords and the corresponding interaction sequence, determine the depth trend of the questions asked by the pre-defined user in that interaction round, and output a positive or negative trend result."

[0101] The above-mentioned method provides a time-series anchor point for trend analysis by recording the interaction sequence. By combining the semantic connotation of the interaction keywords with the temporal relationship of the interaction sequence, the trend direction of the question depth can be determined. This solves the problem of misjudgment of trends caused by only looking at the content and not the order in traditional keyword analysis. It also outputs positive and negative results to reflect the evolution of the real needs of the preset users in a single round of interaction, providing key behavioral feature data for the accurate mapping of subsequent tags.

[0102] In one specific embodiment, S40 includes the following steps: S401, for any interaction round, obtain the first mapping rule corresponding to the current interaction round based on the question depth trend corresponding to the current interaction round.

[0103] S402, according to the first mapping rule, the interaction order corresponding to each group of interaction keywords in the current interaction round is mapped to the first priority corresponding to each group of interaction keywords in the current interaction round, wherein the first priority is used to characterize the relative importance between different interaction keywords in an interaction round.

[0104] S403, according to the second mapping rule, the question depth trend corresponding to the current interaction round is mapped to the second priority corresponding to each group of interaction keywords in the current interaction round, wherein the second priority is used to characterize the relative importance between different interaction rounds.

[0105] S404. Based on the first and second priorities of each group of interactive keywords in the current interaction round, obtain the influence weight corresponding to each group of interactive keywords in the current interaction round.

[0106] In one specific embodiment, S401 includes the following steps: S4011, if the problem depth trend corresponding to the current interaction round is positive, then the interaction order and the first priority in the first mapping rule corresponding to the current interaction round are positively correlated.

[0107] S4012, if the problem depth trend corresponding to the current interaction round is reversed, then the interaction order and the first priority in the first mapping rule corresponding to the current interaction round are negatively correlated.

[0108] The first mapping rule is the correspondence between the interaction order and the first priority, which is determined by the question depth trend of the current interaction round. It dynamically adapts to the keyword importance distribution pattern under different trends, avoiding deviations caused by fixed order weights (such as defaulting to later keywords being more important) (such as mistakenly increasing the weight of later irrelevant keywords in a reverse trend).

[0109] Correspondingly, in a positive trend, user questions become more focused on core needs as interaction deepens. Therefore, in the first mapping rule corresponding to a positive question depth trend, the interaction order and the first priority are positively correlated; that is, the later the interaction order, the higher the first priority. Conversely, in a negative trend, user questions deviate from the core as interaction deepens. Therefore, in the first mapping rule corresponding to a negative question depth trend, the interaction order and the first priority are negatively correlated; that is, the earlier the interaction order, the higher the first priority.

[0110] The first priority is a score indicating the relative importance of keywords across different interaction sequences within the same interaction round. Through a first mapping rule, the interaction sequence is transformed into a quantifiable first priority. For example, in a positive trend round, the first priority Y1 = 1 / (1 + exp(-C)), and in a negative trend round, the first priority Y1 = exp(-C), where Y1 is the first priority, C is the interaction sequence, and exp() is an exponential function with base e. Correspondingly, the range of the first priority is (0, 1).

[0111] The second mapping rule is the correspondence between the question depth trend and the second priority. It is used to quantify the overall value of different rounds, ensuring that keywords in positive rounds that reflect the core needs of users receive higher weight, thus solving the problem of invalid interaction keywords interfering with all rounds due to the traditional approach of treating all rounds equally. Correspondingly, the second priority represents the overall value of different interaction rounds; the higher the second priority, the greater the reference significance of that round in reflecting the core needs of users.

[0112] The second mapping rule can be a fixed score mapping, with the specific score set by the implementer according to the actual situation. For example, a positive trend corresponds to a priority of 5, and a negative trend corresponds to a priority of 3.

[0113] The influence weight of each set of interactive keywords in the current interaction round is obtained by multiplying the first priority and the second priority of each set of interactive keywords.

[0114] The above-described dynamic adaptation mechanism, which determines the mapping rules based on trends, allows keywords with the same interaction order to receive differentiated priorities under different trends. This solves the problem of misjudgment of weight caused by traditional fixed rules, such as assuming that later keywords are more important, in reverse trends. By quantifying the importance differences within the same round through the first priority, keywords that drive demand deepening and those that lay the foundation are clearly distinguished, avoiding insufficient weight of core information due to treating all keywords equally. By quantifying the value differences between different rounds through the second priority, keywords in positive trend rounds receive higher base scores, while the weight of keywords in reverse trend rounds is reasonably suppressed. This solves the problem of invalid interaction interference caused by traditional fixed rules that only consider the importance of a single round and ignore the value of the round itself. Finally, through the comprehensive calculation of the first and second priorities, the output influence weight more closely matches the true importance ranking of users' core needs, providing a precise quantitative basis for subsequent target keyword selection.

[0115] In one specific embodiment, S50 includes the following steps: S501, the product of the influence weight and the number of times each interactive keyword appears is used to determine the importance of each interactive keyword.

[0116] S502, determine the largest interactive keyword corresponding to the current first tag as the target keyword corresponding to the current first tag.

[0117] By comprehensively quantifying the influence weight and frequency of occurrence, the most representative target keywords are selected from all interactive keywords of the current first tag. This ensures that the target keywords not only reflect the core needs of the user with this identity, but also have general representativeness, providing a precise anchor point for the subsequent mapping from the first tag to the third tag. This ensures that the interactions of users with different identities can automatically match the most relevant knowledge scenarios, improving the accuracy and efficiency of the entire information interaction process.

[0118] As described above, by collecting user query and platform response data from all interaction rounds, the interaction data retains the consistency of user identity while dynamically adapting to changes in the problem domain. This provides high-quality interaction data for analyzing the dynamic changes in user needs. Based on the target query statements of all interaction rounds, the temporal semantic reasoning capability of the pre-set large model is used to determine the direction of the problem's depth evolution. This allows the interaction report to reveal from a dynamic perspective whether user needs are deepening or generalizing. Furthermore, the round order is converted into information adoption degree based on the problem depth trend, ensuring that the information value assessment conforms to the evolution logic of users' real needs. This avoids the core information being buried due to traditional average weighting. Finally, by deeply integrating the interaction data, information adoption degree, and pre-set report template, the pre-set large model is used for content generation and structural arrangement. This allows the interaction report to retain the original details of user interactions while highlighting the value of key information. At the same time, it achieves format standardization and insight depth, providing highly targeted, highly professional, and high-quality interaction reports for the target client.

[0119] Example 2 like Figure 2 As shown, this embodiment two provides a report generation device based on a large model, which includes: The data interaction module 21 is used to acquire the interaction data between the target client and the query platform for several interaction rounds. The interaction data for each interaction round includes the target query statement input by the target client into the pre-trained preset large model in the query platform, and the target analysis result text of the preset large model for the target query statement.

[0120] The trend analysis module 222 is used to obtain the question depth trend corresponding to the target client based on all target query statements corresponding to all interaction rounds, wherein the question depth trend is either positive or negative.

[0121] The degree mapping module 23 is used to map the degree of information adoption for each interaction round based on the question depth trend of the target client.

[0122] The report generation module 24 is used to input the interaction data, the information adoption level corresponding to each interaction round, and the preset report template into the preset large model to obtain the interaction report between the target client and the query platform.

[0123] In one specific embodiment, the data interaction module 21 includes: The tag acquisition submodule is used to input the first target query statement from the target client into the pre-trained large model in the query platform to obtain the first tag and the second tag corresponding to the first target query statement. The first tag is used to represent the identity domain corresponding to the first target query statement, and the second tag is used to represent the question domain corresponding to the first target query statement.

[0124] The third tag matching submodule is used to match the first tag with the preset tag library in the query platform to obtain the third tag corresponding to the first target query statement. The preset tag library includes the mapping relationship between the first tag and the third tag, and each third tag corresponds to several query knowledge bases in the query platform.

[0125] The first result text acquisition submodule is used to input the first target query statement and the first label, second label and third label corresponding to the first target query statement into the preset large model, and obtain the first target analysis result text corresponding to the first target query statement.

[0126] The target query statement acquisition submodule is used to obtain the (j+1)th target query statement from the target client for the analysis result text of the j-th target, where j > 0.

[0127] The second tag acquisition submodule is used to input the (j+1)th target query statement into the preset large model and obtain the second tag corresponding to the (j+1)th target query statement.

[0128] The second result text acquisition submodule is used to input the (j+1)th target query statement, the second label corresponding to the (j+1)th target query statement, and the first label and third label corresponding to the first target query statement into the preset large model to obtain the (j+1)th target analysis result text corresponding to the (j+1)th target query statement.

[0129] The interactive data acquisition submodule is used to update j=j+1 and repeatedly execute the target query statement acquisition submodule until the target client ends its interaction with the query platform. The target query statement and target analysis result text between the target client and the query platform for each interaction round are identified as the interactive data between the target client and the query platform for each interaction round.

[0130] In one specific embodiment, the large model-based report generation device further includes: The interaction data acquisition module is used to acquire the interaction data of the current first tag for several preset users in several interaction rounds for any first tag. The interaction data of each interaction round includes several sets of interaction text between a preset user and a preset large model. Each set of interaction text includes the question text of the preset user and the answer text of the preset large model to the question text.

[0131] The interactive keyword acquisition module is used to extract keywords from each set of interactive text, obtaining a set of interactive keywords corresponding to each set of interactive text.

[0132] The question depth trend acquisition module is used to analyze all groups of interaction keywords corresponding to each interaction round to obtain the question depth trend of the preset user for each interaction round. The question depth trend can be positive or negative.

[0133] The influence weight acquisition module is used to obtain the influence weight of each interactive keyword corresponding to the current first tag based on the question depth trend corresponding to all interaction rounds.

[0134] The target keyword acquisition module is used to obtain the target keywords corresponding to the current first tag based on the influence weight of each interactive keyword and the number of times each interactive keyword appears.

[0135] The third tag mapping module is used to map the third tag corresponding to the target keyword in the preset tag library to the third tag corresponding to the current first tag.

[0136] The first tag traversal module is used to traverse all first tags and obtain the third tag corresponding to each first tag.

[0137] In one specific implementation, the influence weight acquisition module includes: The first mapping rule acquisition submodule is used to obtain the first mapping rule corresponding to the current interaction round based on the question depth trend corresponding to the current interaction round.

[0138] The first priority acquisition submodule is used to map the interaction order corresponding to each group of interaction keywords in the current interaction round to the first priority corresponding to each group of interaction keywords in the current interaction round according to the first mapping rule. The first priority is used to characterize the relative importance between different interaction keywords in an interaction round.

[0139] The second priority acquisition submodule is used to map the question depth trend corresponding to the current interaction round to the second priority corresponding to each group of interaction keywords in the current interaction round according to the second mapping rule. The second priority is used to represent the relative importance between different interaction rounds.

[0140] The influence weight acquisition submodule is used to obtain the influence weight of each group of interactive keywords in the current interaction round based on the first priority and second priority of each group of interactive keywords in the current interaction round.

[0141] In one specific implementation, the trend analysis module 22 includes: The round order acquisition submodule is used to obtain the round order of the target query statement corresponding to each interaction round.

[0142] The Problem Depth Trend Acquisition Submodule is used to input all target query statements corresponding to all interaction rounds, the round order of each target query statement, and preset prompt words into a preset large model to obtain the problem depth trend corresponding to the target client.

[0143] In one specific embodiment, the degree mapping module 23 includes: The third mapping rule acquisition submodule is used to obtain the third mapping rule corresponding to the target client based on the question depth trend of the target client.

[0144] The information adoption level acquisition submodule is used to map the round order of each target query statement to the information adoption level according to the third mapping rule.

[0145] In one specific implementation, the third mapping rule acquisition submodule includes: The first rule-determining unit is used to determine that if the problem depth trend corresponding to the target client is positive, then the round order and the degree of information adoption in the third mapping rule corresponding to the target client are positively correlated.

[0146] The second rule-determining unit is used to determine that if the problem depth trend corresponding to the target client is reversed, then the round order and information adoption degree in the third mapping rule corresponding to the target client are negatively correlated.

[0147] Embodiments of the present invention also provide a non-transitory computer-readable storage medium, which can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a method in the method embodiment. The at least one instruction or the at least one program is loaded and executed by the processor to implement the large model-based report generation method provided in Embodiment 1 above.

[0148] Embodiments of the present invention also provide an electronic device, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0149] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0150] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A report generation method based on a large model, characterized in that, The report generation method based on large models also includes the following steps: S100: Obtain interaction data between the target client and the query platform for several interaction rounds, wherein the interaction data for each interaction round includes the target query statement input by the target client to the pre-trained preset large model in the query platform, and the target analysis result text of the preset large model for the target query statement. S100 includes the following steps: S110, input the first target query statement of the target client into the pre-trained large model in the query platform, and obtain the first label and the second label corresponding to the first target query statement, wherein the first label is used to characterize the identity domain corresponding to the first target query statement, and the second label is used to characterize the question domain corresponding to the first target query statement; S120, the first tag is matched with the preset tag library in the query platform to obtain the third tag corresponding to the first target query statement, wherein the preset tag library includes the mapping relationship between the first tag and the third tag, and each third tag corresponds to several query knowledge bases in the query platform; S130, input the first target query statement and the first label, second label and third label corresponding to the first target query statement into the preset large model to obtain the first target analysis result text corresponding to the first target query statement; S140, obtain the (j+1)th target query statement of the target client for the jth target analysis result text, where j>0; S150, input the (j+1)th target query statement into the preset large model to obtain the second tag corresponding to the (j+1)th target query statement; S160, input the (j+1)th target query statement, the second label corresponding to the (j+1)th target query statement, and the first label and third label corresponding to the first target query statement into the preset large model to obtain the (j+1)th target analysis result text corresponding to the (j+1)th target query statement; S170, update j=j+1, repeat step S140 until the target client ends its interaction with the query platform, and determine the target query statement and target analysis result text between the target client and the query platform for each interaction round as the interaction data between the target client and the query platform for each interaction round; S200, based on all target query statements corresponding to all interaction rounds, obtain the question depth trend corresponding to the target client, wherein the question depth trend is either positive or negative; S300, based on the question depth trend corresponding to the target client, the information adoption degree corresponding to each interaction round is mapped to the result; S400, input the interaction data, the information adoption degree corresponding to each interaction round, and the preset report template into the preset large model to obtain the interaction report between the target client and the query platform.

2. The report generation method based on a large model according to claim 1, characterized in that, The report generation method based on large models also includes the following steps: S10, for any first tag, obtain the interaction data of the current first tag for several interaction rounds for several preset users, wherein the interaction data of each interaction round includes several sets of interaction text between a preset user and a preset large model, and each set of interaction text includes the question text of the preset user and the answer text of the preset large model for the question text. S20, Extract keywords from each set of interactive texts to obtain a set of interactive keywords corresponding to each set of interactive texts; S30, Analyze all groups of interactive keywords corresponding to each interaction round to obtain the question depth trend of the preset user corresponding to each interaction round, wherein the question depth trend is positive or negative; S40: Based on the question depth trend corresponding to all interaction rounds, obtain the influence weight of each interaction keyword corresponding to the current first tag; S50: Based on the influence weight of each interactive keyword corresponding to the current first tag and the number of times each interactive keyword appears, obtain the target keyword corresponding to the current first tag; S60, map the third tag corresponding to the target keyword in the preset tag library to the third tag corresponding to the current first tag; S70, iterate through all the first tags and get the third tag corresponding to each first tag.

3. The report generation method based on a large model according to claim 1, characterized in that, S40 includes the following steps: S401, For any interaction round, obtain the first mapping rule corresponding to the current interaction round based on the question depth trend corresponding to the current interaction round; S402, according to the first mapping rule, the interaction order corresponding to each group of interaction keywords in the current interaction round is mapped to the first priority corresponding to each group of interaction keywords in the current interaction round, wherein the first priority is used to characterize the relative importance between different interaction keywords in an interaction round; S403, According to the second mapping rule, the question depth trend corresponding to the current interaction round is mapped to the second priority corresponding to each group of interaction keywords in the current interaction round, wherein the second priority is used to characterize the relative importance between different interaction rounds; S404. Based on the first and second priorities of each group of interactive keywords in the current interaction round, obtain the influence weight corresponding to each group of interactive keywords in the current interaction round.

4. The report generation method based on a large model according to claim 1, characterized in that, S200 includes the following steps: S210, obtain the round order of the target query statement corresponding to each interaction round; S220, input all target query statements corresponding to all interaction rounds, the round order of each target query statement, and preset prompt words into the preset large model to obtain the question depth trend corresponding to the target client.

5. The report generation method based on a large model according to claim 4, characterized in that, S300 includes the following steps: S310, Based on the question depth trend corresponding to the target client, obtain the third mapping rule corresponding to the target client; S320, according to the third mapping rule, the round order of each target query statement is mapped to the degree of information adoption.

6. The report generation method based on a large model according to claim 5, characterized in that, S310 includes the following steps: S311, if the problem depth trend corresponding to the target client is positive, then the round order and the degree of information adoption in the third mapping rule corresponding to the target client are positively correlated; S312, if the problem depth trend corresponding to the target client is reversed, then the round order and information adoption degree in the third mapping rule corresponding to the target client are negatively correlated.

7. A report generation device based on a large model, characterized in that, The report generation device based on the large model includes: A data interaction module is used to acquire interaction data between the target client and the query platform for several interaction rounds. The interaction data for each round includes the target query statement input by the target client to a pre-trained preset large model in the query platform, and the target analysis result text of the preset large model for the target query statement. The data interaction module includes: The tag acquisition submodule is used to input the first target query statement of the target client into the pre-trained preset large model in the query platform to obtain the first tag and the second tag corresponding to the first target query statement. The first tag is used to represent the identity domain corresponding to the first target query statement, and the second tag is used to represent the question domain corresponding to the first target query statement. The third tag matching submodule is used to match the first tag with the preset tag library in the query platform to obtain the third tag corresponding to the first target query statement. The preset tag library includes the mapping relationship between the first tag and the third tag, and each third tag corresponds to several query knowledge bases in the query platform. The first result text acquisition submodule is used to input the first target query statement and the first label, second label and third label corresponding to the first target query statement into the preset large model to obtain the first target analysis result text corresponding to the first target query statement; The target query statement acquisition submodule is used to acquire the (j+1)th target query statement of the target client for the jth target analysis result text, where j>0; The second tag acquisition submodule is used to input the (j+1)th target query statement into the preset large model and obtain the second tag corresponding to the (j+1)th target query statement; The second result text acquisition submodule is used to input the (j+1)th target query statement, the second label corresponding to the (j+1)th target query statement, and the first label and third label corresponding to the first target query statement into the preset large model to obtain the (j+1)th target analysis result text corresponding to the (j+1)th target query statement. The interactive data acquisition submodule is used to update j=j+1 and repeat step S140 until the target client ends its interaction with the query platform. The target query statement and target analysis result text between the target client and the query platform for each interaction round are determined as the interactive data between the target client and the query platform for each interaction round. The trend analysis module is used to obtain the question depth trend corresponding to the target client based on all target query statements corresponding to all interaction rounds, wherein the question depth trend is either positive or negative; The degree mapping module is used to map the degree of information adoption for each interaction round according to the question depth trend corresponding to the target client. The report generation module is used to input the interaction data, the information adoption level corresponding to each interaction round, and the preset report template into the preset large model to obtain the interaction report between the target client and the query platform.

8. A non-transitory computer-readable storage medium storing at least one instruction or at least one program segment, characterized in that, The at least one instruction or the at least one program segment is loaded and executed by the processor to implement the large model-based report generation method as described in any one of claims 1-6.

9. An electronic device, characterized in that, Includes a processor and the non-transitory computer-readable storage medium as described in claim 8.

Images