Research report generation method, electronic device, and computer-readable storage medium

Abstract

The application discloses a research report generation method, an electronic device and a computer readable storage medium. The method comprises: obtaining conference text, which is obtained based on transcribing conference recording data; generating a topic prompt word, the topic prompt word comprising preset research requirements to be met by each research attribute; using a second large language model to extract a topic based on the topic prompt word and the conference text to obtain a topic extraction result, the topic extraction result comprising several research topics; taking at least one research topic as a target research topic, searching at least one Internet data source using keywords corresponding to the target research topic to obtain several first topic texts related to the target research topic; and using a first large language model to generate a research report of the target research topic based on the several first topic texts and second topic texts related to the target research topic in the conference text. In this way, the application can generate a more comprehensive research report.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a research report generation method, electronic device, and computer-readable storage medium. Background Technology

[0002] In practical research scenarios, such as in the consumer electronics sector, formulating new product pricing strategies is crucial for companies. However, numerous pain points exist in actual research practices. Traditional methods of manually compiling recordings of pricing strategy brainstorming sessions are inefficient and prone to omissions and errors. Internal discussions are limited by information barriers, creating blind spots in understanding the actual market situation. It's difficult to comprehensively and accurately grasp key market information such as competitor pricing ranges, core strengths and weaknesses, and there's a lack of crucial information on distribution channels. Even after identifying a research topic, manually searching the internet for information is time-consuming and laborious, and it's difficult to effectively integrate market data with key points from internal discussions. This results in pricing strategy reports that are detached from reality or unimplementable, failing to develop scientifically effective pricing solutions. Summary of the Invention

[0003] The main technical problem addressed by this application is to provide a research report generation method, electronic device, and computer-readable storage medium that can generate more comprehensive research reports.

[0004] To address the aforementioned technical problems, this application provides a research report generation method. The method includes: acquiring meeting text, wherein the meeting text is obtained by transcribing meeting recording data; generating topic prompts, wherein the topic prompts include preset research requirements that each research attribute must meet; using a second language model to extract topics based on the topic prompts and the meeting text, obtaining topic extraction results, wherein the topic extraction results include several research topics; selecting at least one of the research topics as a target research topic, and using keywords corresponding to the target research topic to search in at least one internet data source to obtain several first topic texts related to the target research topic; and using a first language model to generate a research report on the target research topic based on the several first topic texts and second topic texts related to the target research topic in the meeting text.

[0005] To address the aforementioned technical problems, a second aspect of this application provides an electronic device comprising a memory and a processor, wherein the memory stores program instructions and the processor executes the program instructions to implement the aforementioned research report generation method.

[0006] To address the aforementioned technical problems, a third aspect of this application provides a computer-readable storage medium for storing program instructions that can be executed to implement the aforementioned research report generation method.

[0007] The aforementioned technical solution involves extracting several research topics from conference texts that meet corresponding preset research requirements based on their research attributes. Therefore, these extracted or refined research topics are suitable for in-depth research, possess significant research implications, and have substantial research value, avoiding the blind and arbitrary selection of research topics. Furthermore, several primary research topic texts related to the target research topic are obtained by searching internet data sources using keywords corresponding to the target research topic. Therefore, generating a research report on the target research topic based on these primary topic texts and related secondary topic texts from the conference involves a comprehensive analysis and integration of both texts, organically incorporating them into the research report. This results in a more comprehensive and systematic reflection of all aspects of the target research topic, improving the report's quality and value. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating an embodiment of the research report generation method provided in this application; Figure 2 This is a flowchart illustrating an embodiment of the present application that uses the second largest language model to extract topics based on topic prompts and meeting text to obtain topic extraction results; Figure 3 This is a flowchart illustrating an embodiment of the training method for the second largest language model provided in this application; Figure 4 This is a schematic diagram of the structure of an embodiment of the research report generation device provided in this application; Figure 5 This is a schematic diagram of the structure of an embodiment of the electronic device provided in this application; Figure 6 This is a schematic diagram of an embodiment of the computer-readable storage medium provided in this application. Detailed Implementation

[0009] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0010] In the following description, specific details such as particular system architectures, interfaces, and technologies are presented for illustrative purposes rather than for limiting purposes, in order to provide a thorough understanding of this application.

[0011] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this document means two or more. Moreover, the term "at least one" in this document means any combination of at least two of any one or more of a plurality of objects. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0012] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the research report generation method provided in this application. It should be noted that if substantially the same result is obtained, this embodiment is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, this embodiment includes: Step S11: Obtain the meeting text.

[0013] In this embodiment, the meeting text is obtained; wherein, the meeting text is obtained based on the transcription of meeting recording data. The meeting recording data can be audio data directly recorded from online / offline meetings, or audio content extracted from video data recorded from online / offline meetings.

[0014] In one embodiment, obtaining the meeting text can specifically involve: collecting meeting recording data during the meeting; transcribing the meeting recording data into a transcribed text; and using the transcribed text as the meeting text. In other words, the transcribed text obtained from transcribing the meeting recording data is used as the meeting text.

[0015] In other implementations, obtaining the meeting text can specifically involve: collecting meeting recording data during the meeting; transcribing the meeting recording data into a transcribed text; and converting the transcribed text into a structured minutes to serve as the meeting text. A structured minutes is a summary generated by using technical means to extract, categorize, and organize the text content from the transcribed text according to predefined templates or data fields. It can be viewed as a meeting summary about the meeting, or the meeting summary about the meeting can be used as the meeting text.

[0016] In one specific implementation, Automatic Speech Recognition (ASR) technology can be used to transcribe meeting recordings into text. ASR technology can process meeting recordings, converting the speech of all speakers into text, word by word (i.e., transcribed text). Furthermore, during the transcription process, ASR technology can be combined with voiceprint recognition technology to automatically distinguish different speakers and tag each sentence in the transcribed text with the speaker's name, thus significantly improving the readability and usability of the transcribed text.

[0017] Step S12: Generate topic selection prompts and use the second language model to extract topics based on the topic selection prompts and the meeting text to obtain the topic extraction results.

[0018] In this implementation, topic selection prompts are generated, and a second language model is used to extract topics based on these prompts and the conference text. The extracted topics include pre-defined research requirements for each research attribute, and the extracted topics comprise several research topics. These research topics are extracted from the conference text and meet the corresponding pre-defined research requirements. Therefore, the research topics extracted or refined from the conference text are suitable for in-depth research, have significant research implications, and possess in-depth research value, avoiding the blindness and arbitrariness in research topic selection.

[0019] In one embodiment, the research attributes include at least one of the following: content complexity, problem openness, knowledge openness, and knowledge breadth, wherein content complexity characterizes the depth of the topic, problem openness characterizes the diversity of solutions to the topic, knowledge openness characterizes the availability of the knowledge involved in the topic, and knowledge breadth characterizes the breadth of knowledge sources involved in the topic.

[0020] The depth of a chosen topic refers to its rich layers and in-depth details at the theoretical, logical, or practical levels. It emphasizes not simply piling up information, but rather the interaction of multiple dimensions, their numerous variations, and their inherent connections. For example, using Yang's Triangle to solve the incentive mechanism for a sales team requires a deep analysis of the performance and mutual influence of multiple dimensions such as organizational capabilities, employee mindset, and employee governance within the specific environment of a sales team. The relationships between these factors are complex and open to in-depth exploration. In contrast, the Pythagorean theorem, as a clear mathematical theorem, is concise and logically simple, lacking the space for multi-layered in-depth discussion and thus lacking sufficient depth.

[0021] Selecting topics whose content complexity meets the pre-defined research requirements allows for the selection of research topics with a certain level of complexity. Topics with a certain level of complexity possess rich layers and details to be explored at the theoretical, logical, or practical levels. Therefore, topics with a certain level of complexity are suitable for in-depth research and have in-depth research value, thus avoiding the selection of topics that are too simple and lack in-depth research value.

[0022] The diversity of solutions to a research topic means that there is no fixed, unique answer or solution; multiple possible interpretations, approaches, or outcomes exist. It encourages analysis and exploration from different angles, theoretical frameworks, and practical backgrounds. For example, when a company explores the feasibility of fragrance as a second growth curve, it can consider multiple aspects such as market demand, product development, brand positioning, and marketing strategies. Different people, based on different experiences and understandings, may arrive at different conclusions and implementation paths. In contrast, the Fibonacci sequence has a clear mathematical definition and calculation method, resulting in relatively fixed answers and less diversity of solutions.

[0023] Selecting research topics that meet the pre-defined research requirements based on their openness to the problem allows for the selection of topics with a certain degree of openness. Topics with a certain degree of openness involve multiple possible explanations, approaches, or results. Therefore, topics with a certain degree of openness are suitable for in-depth research and have in-depth research value, enabling in-depth research from different angles and levels, and generating rich and diverse research results.

[0024] The openness and breadth of knowledge required for a research topic necessitate that the knowledge involved be readily available from public sources, and that the sources be broad and comprehensive. This facilitates multifaceted comparison, analysis, and verification, making the research more reliable and comprehensive. For example, analyzing the competitiveness of a company's compensation mechanism can be done by collecting a large amount of salary data from similar companies and industry reports online; however, analyzing the growth potential of 100 employees within a company relies on internal private data, which is limited in availability and lacks external comparison, resulting in poor data accessibility and breadth.

[0025] Selecting research topics that meet the pre-defined research requirements in terms of knowledge openness allows for the selection of topics with a certain degree of knowledge openness. Topics with a certain degree of knowledge openness are those where the knowledge involved can be relatively easily obtained from public channels. Therefore, topics with a certain degree of knowledge openness are suitable for in-depth research and have in-depth research value. It is easier to obtain relevant materials, data, and information related to the topic, providing a solid foundation for in-depth research.

[0026] Selecting research topics that meet the pre-defined research requirements in terms of breadth of knowledge allows for the selection of topics with a certain breadth of knowledge. Topics with a certain breadth of knowledge involve a wide range of knowledge sources and have broad coverage. Therefore, topics with a certain breadth of knowledge are suitable for in-depth research and have in-depth research value. They can obtain more relevant resources, data, and information from different sources related to the topic, providing a solid foundation for in-depth research.

[0027] In one implementation, the topic extraction results also include keywords corresponding to each research topic.

[0028] In one implementation, the second language model is used to extract topics based on topic prompts and meeting text to obtain topic extraction results. Specifically, the meeting text can be decomposed into topic-level sub-topics to obtain several parent topics; each parent topic can be further decomposed to obtain several sub-topics corresponding to each parent topic, which are used as the topic extraction results corresponding to the meeting text.

[0029] The conference text undergoes overall topic modeling and semantic aggregation to extract its core research topics. These topics are macro-level, encompassing the core objectives of the conference text, guiding all research directions, and possessing global guidance. They are unique or exhibit some parallelism and serve as the root node of the entire topic tree. For each topic, it is automatically broken down into several sub-topics, resulting in parallel, complementary, and non-overlapping second-level sub-topics. Therefore, the topic extraction results from the conference text form a tree-like structure of topic-sub-topics, ensuring a more complete coverage of the conference text.

[0030] Compared to single-layer, flat topic extraction techniques, the top-down, multi-level, tree-like topic decomposition constructs a structured topic selection system with a master topic and supporting sub-topics, solving the technical shortcomings of traditional techniques such as disorganized, lack of hierarchy, and lack of focus in topic selection. Furthermore, through multi-level, progressively granular decomposition, it comprehensively exhaustively covers all dimensions of conference text research, ensuring both the integrity of the overall framework of the subsequently generated research report and the implementation of detailed micro-level information. This addresses the problems of insufficient research dimensions and inadequate report depth in traditional solutions.

[0031] In one implementation, the second language model is used to extract topics based on topic prompts and meeting text to obtain topic extraction results. Specifically, this can be: extracting several explicit topics from the meeting text, and extracting several implicit topics from the meeting text based on the textual semantics of the meeting text; and using the several explicit topics and the several implicit topics as the topic extraction results corresponding to the meeting text.

[0032] Explicit research topics are directly extracted from the conference text. These are research tasks, analytical directions, and discussion topics that are explicitly and proactively mentioned in the conference text, representing concrete research needs that users clearly want to explore. Implicit research topics, on the other hand, are potential research gaps and assessment needs that are not explicitly mentioned, assigned research tasks, or have no direct research sentences in the conference text, but can be inferred through contextual semantic relationships, logical deduction, judgment of decision-making status, or analysis of points of contention. Implicit research topics do not rely on literal keywords but rather on in-depth reasoning and mining based on implicit features such as sentence semantic relationships, contextual logical chains, expressions of uncertainty, conflicting viewpoints, or decision-making gaps.

[0033] Abandoning the traditional, singular explicit keyword matching and extraction model, we construct a dual-track topic identification architecture combining explicit and implicit keywords, achieving a technological upgrade from literal extraction to deep semantic understanding of conference research needs. Through a complementary and integrated explicit-implicit topic selection mechanism, we ensure that the generated research reports align with the user's explicit research intentions while supplementing core decision-making content such as in-depth analysis, risk assessment, and feasibility verification, significantly enhancing the professionalism and decision-making value of the generated research reports.

[0034] Step S13: Select at least one research topic as the target research topic, and use the keywords corresponding to the target research topic to search in at least one Internet data source to obtain several first topic texts related to the target research topic.

[0035] In this implementation, at least one research topic is selected as the target research topic. Keywords corresponding to the target research topic are used to search at least one internet data source to obtain several first-topic texts related to the target research topic. Internet data sources contain massive amounts of information, covering various fields and data types. By using keywords corresponding to the target research topic for searching, first-topic texts related to the topic can be obtained from multiple different websites, platforms, and databases, greatly broadening the channels for information acquisition. Furthermore, information on the internet updates extremely quickly, reflecting the latest industry developments, research results, and policy and regulatory changes in a timely manner. The first-topic texts obtained through the search allow researchers to understand the current status of the target research topic, making the research report more timely and relevant. Moreover, the internet contains various voices and viewpoints. For the same target research topic, different people may have different opinions, analyses, and interpretations. The first-topic texts obtained through the search allow researchers to access these diverse perspectives, thereby gaining a more comprehensive understanding of all aspects of the topic and avoiding one-sided research.

[0036] In one embodiment, the keywords corresponding to the target research topic include original topic words and extended topic words. The original topic words include subject and theme-related words in the target research topic. The extended topic words include at least one of the following: a first extended word related to the subject based on subject fission, and a second extended word related to theme-related words based on theme fission.

[0037] The original keywords for the research topic clearly define the core subject and related terms, serving as a precise entry point for topic searching. Expanded keywords further broaden the search scope, uncovering information relevant to the target research topic from different dimensions. In other words, the original keywords provide a clear direction for the search, ensuring that search results are closely related to the core content of the target research topic; expanded keywords, derived from the subject and related terms, can precisely pinpoint specific areas and details related to the target research topic.

[0038] By generating first-level extended terms based on the subject, information related to the subject can be explored more deeply from different perspectives. These first-level extended terms are not limited to the direct attributes of the subject itself, but can also involve related industries, markets, users, and other aspects. Generating second-level extended terms based on the theme allows for further refinement of the theme, clarifying specific research directions and priorities. These second-level extended terms can be generated based on the latest development trends and hot topics within the theme.

[0039] For example, taking the competitive analysis of a company's compensation mechanism as the research topic: the subject of the research topic could be a new energy vehicle company, and the first extended terms related to the subject based on subject fission could be AI startups, etc.; the theme-related terms in the research topic could be compensation mechanisms, and the second extended terms related to theme-related terms based on theme fission could be long-term incentive mechanisms, compensation and benefits, etc.

[0040] In one implementation, before searching at least one internet data source using keywords corresponding to the target research topic to obtain several first-topic texts related to the target research topic, at least one keyword corresponding to the research topic is also obtained and presented to the user. In response to user adjustments to the keywords and / or additions to the keywords, the keywords are updated. Different users have different understandings and focuses regarding research topics. By presenting users with the initially obtained keywords, they can adjust the keywords according to their research intentions, professional background, and specific needs, making the search more accurately match their actual needs. Furthermore, user adjustments can remove keywords that may lead to irrelevant search results, thus avoiding obtaining a large amount of irrelevant content during the search process and improving search efficiency and accuracy. Users, with their professional knowledge and in-depth understanding of the research field, may discover omissions in the initially obtained keywords. Through keyword addition, users can add important keywords closely related to the research topic, expanding the search scope and making the search results more comprehensive.

[0041] In one embodiment, at least one internet data source includes API interfaces provided by search engines and third-party knowledge bases. These third-party knowledge bases include, but are not limited to, Baidu Wenku and CNKI (China National Knowledge Infrastructure), etc., and are not limited thereto.

[0042] In one embodiment, at least one internet data source includes an API interface provided by a search engine and a third-party knowledge base. Searching for keywords corresponding to the target research topic within the at least one internet data source to obtain several first-topic texts related to the target research topic can specifically involve: searching for several webpage links matching the keywords corresponding to the target research topic through the API interface provided by the search engine, accessing the corresponding webpages through each webpage link, and extracting the first-topic texts related to the target research topic from each accessed webpage; and searching for several documents matching the keywords corresponding to the target research topic from the third-party knowledge base, and extracting the first-topic texts related to the target research topic from each matched document.

[0043] Based on keywords corresponding to the target research topic, multi-source internet content search can be achieved. Through the API interface provided by the search engine, webpage links matching the keywords of the target research topic can be searched and obtained. Then, the content of each webpage linked to can be extracted to obtain the primary research topic text. For simple webpages, content extraction can be performed using webpage content crawling, while for complex webpages, web crawlers can be used to extract the content; no limitation is made here. From third-party knowledge bases, several documents matching the keywords of the target research topic can be searched and obtained. Then, the content of each document can be extracted to obtain the primary research topic text. MCP tools can be used to search and obtain several documents matching the keywords of the target research topic from third-party knowledge bases.

[0044] In one specific implementation, extracting first topic text related to the target research topic from each visited webpage, or extracting first topic text related to the target research topic from each matched document, can specifically involve: using each webpage or document as the target text; segmenting the target text and converting each segment of the target text into a first embedding vector; obtaining second embedding vectors of keywords corresponding to the target research topic; obtaining the vector similarity between the second embedding vector and the first embedding vector of each segment, and selecting segments with vector similarity greater than a similarity threshold as the first topic text.

[0045] For web pages or documents related to the target research topic, the relevance to the research topic is often only a part of the content. Therefore, the target text (web page or document) is segmented and embedded, and the keywords corresponding to the target research topic are also embedded. Then, through vector retrieval, text segments with high similarity to the keywords corresponding to the target research topic (vector similarity greater than the similarity threshold) are extracted as the primary topic text related to the target research topic. This can improve the accuracy of the primary topic text related to the target research topic and avoid the problem of the context content being too long and exceeding the processing limit of the first language model when generating the research report.

[0046] In other specific implementations, extracting first topic text related to the target research topic from each visited webpage, or extracting first topic text related to the target research topic from each matched document, can specifically involve: using each webpage or document as the target text; performing synonym expansion on the keywords corresponding to the target research topic to obtain at least one synonym; forming a search term set by combining the keywords corresponding to the target research topic with at least one synonym; performing word segmentation on the target text to obtain several word segments of the target text; determining the position of the word segment matching each search term in the search term set in the target text as the hit position; and using paragraphs in the target text where the number of hit positions is greater than a preset threshold as the first topic text.

[0047] Specifically, the keywords corresponding to the target research topic are expanded with synonyms, and the keywords corresponding to the target research topic are combined with at least one synonym to form a search term set; multiple search terms are retrieved in several word segments of the target text, and the position of the word segment matching the search term in the target text is recorded as the hit position; hotspot aggregation is performed on the hit positions of the word segment matching the search term in the target text. Hotspot aggregation can be performed using a segmented histogram method, such as cutting the target text into paragraphs, accumulating the hit positions of each paragraph, and extracting the paragraphs with hit positions exceeding a threshold as the first topic text related to the target research topic.

[0048] Expanding keywords corresponding to the target research topic with synonyms to form a search term set can cover different expressions of the same concept. Different terms may be used to describe the same research content or topic in web pages or documents. For example, for the target research topic "Application of Artificial Intelligence in Medical Diagnosis," synonyms for the keyword "artificial intelligence" might include "AI" and "machine intelligence," while synonyms for "medical diagnosis" might include "medical diagnosis" and "clinical diagnosis." By including these synonyms in the search term set, we can more accurately match content related to the target research topic in the target text, avoiding the omission of important information due to vocabulary differences, thereby improving the accuracy of information extraction. Determining the position of the word segment matching each search term in the target text as the hit position, and selecting paragraphs with more than a preset threshold of hit positions as the first topic text, ensures that the extracted paragraphs are highly relevant to the target research topic. Only when a paragraph contains multiple occurrences of a word segment matching the search term set is it considered closely related to the target research topic, thus reducing the possibility of misjudgment and improving the accuracy of information extraction.

[0049] The preset threshold number is not limited and can be set according to actual usage needs. Furthermore, the word segmentation of the target text can be dictionary-based, statistical-based, or deep learning-based; no specific limitation is made here. Dictionary-based segmentation relies on a pre-built dictionary, matching the target text against words in the dictionary according to a certain strategy. If a matching word is found in the dictionary, it is segmented. Dictionary-based segmentation is intuitive in principle, relatively easy to implement, and has low development and maintenance costs. Because it only requires dictionary matching, the segmentation speed is fast, meeting the needs of applications with high real-time requirements. Statistical-based segmentation learns from a large amount of annotated segmented text, statistically analyzing the combination frequency and mutual information between adjacent characters to determine which characters are more likely to form words. Because it's based on statistical rules and doesn't rely on a fixed dictionary, it can identify and segment new words and proper nouns that aren't in the dictionary to a certain extent, based on contextual information and statistical patterns. For ambiguous text, statistical segmentation can analyze contextual information, calculate the probability of different segmentation methods, and select the one with the highest probability, thus better handling ambiguity. Deep learning-based segmentation uses deep learning models to extract and model text features, automatically learning segmentation patterns. It can capture contextual information and semantic features, learning how to accurately segment text through extensive training data. Deep learning models can automatically learn complex features and patterns in text, and have a stronger ability to capture contextual information, thus achieving higher segmentation accuracy when processing various complex texts.

[0050] For example, if the keyword corresponding to the target research topic is "mobile phone", synonyms for "mobile phone" could be "mobile phone", "smartphone", "cell phone", etc. Similarly, if the target text is "I went to the library to read books", the word segments for "I went to the library to read books" would be "I", "went", "library", and "read books".

[0051] In one implementation, at least one research topic is designated as the target research topic. Specifically, this can be achieved by: displaying several research topics in response to a user's request to view them; and designating the user's selected research topic as the target research topic in response to the user's selection of at least one research topic. Users can select target research topics based on their needs and interests, ensuring that the selected topics closely align with the user's research intentions. This avoids biases in topic selection due to inaccurate understanding of user needs, allowing research to more precisely focus on the user's concerns and improving the effectiveness and relevance of the research.

[0052] In one specific implementation, when displaying research topics, relevant information for each research topic can be provided simultaneously, such as the research background and objectives of the research topic.

[0053] In one embodiment, a search is performed in at least one Internet data source using keywords corresponding to the target research topic to obtain several first topic texts related to the target research topic. Specifically, this can be done by: searching in at least one Internet data source using keywords corresponding to the target research topic to obtain several initial topic texts related to the target research topic; selecting at least one initial topic text that meets the search requirements from the several initial topic texts as the first topic text, wherein the initial topic text that meets the search requirements is: an initial topic text whose publication time is greater than a preset time.

[0054] By using a preset time as a filtering threshold, only initial topic selection texts published later than the preset time are retained as the first topic selection texts, thus filtering out outdated and obsolete initial topic selection texts. This achieves an initial screening of the timeliness of internet search data, effectively preventing a large amount of outdated, delayed, and invalid information from being mixed into the research argument system. On the one hand, it can quickly remove outdated data such as market trends, competitor dynamics, and industry hot topics that iterate rapidly, avoiding the problem of distorted research conclusions and biased judgments due to the use of outdated information, ensuring the timeliness and practical reference value of research content from the source of the data. On the other hand, through time threshold filtering rules, the standardization and automation of the timeliness screening of initial topic selection texts are achieved, eliminating the need for manual verification of the newness of information one by one, greatly reducing the manual intervention cost of initial topic selection text screening, and improving the overall efficiency of initial topic text acquisition and cleaning.

[0055] In one specific implementation, the preset timeframes for different types of initial topic selection texts can be different; that is, different validity periods are set for different types of initial topic selection texts. Configuring differentiated preset time thresholds for different types of initial topic selection texts breaks away from a uniform, one-size-fits-all timeliness filtering logic, balancing information timeliness with content retention integrity. For initial topic selection texts related to policies and regulations, industry basic standards, and underlying technical principles, a more lenient and longer-term preset timeframe is configured to prevent the accidental deletion of authoritative existing data with long-term reference value due to a uniform short-term window, thus preserving core arguments such as industry benchmarks and compliance evidence. For iterative initial topic selection texts related to market price trends, competitor new product dynamics, and short-term consumption trends, a tighter, more recent preset timeframe is configured to accurately filter outdated and invalid data, avoiding distortion of pricing and competitive analysis conclusions due to outdated market information.

[0056] For example, initial topic texts on policies and regulations are valid for a long period or have a relatively long validity period; therefore, for initial topic texts on policies and regulations, the time window can be broadened, and the preset time period can be set shorter. On the other hand, initial topic texts on market trends are valid for a short period and need to be strictly limited to a recent time frame to ensure timeliness while avoiding the omission of high-value information; therefore, for initial topic texts on market trends, the preset time period can be set longer.

[0057] Of course, in other specific implementations, the preset time corresponding to different types of initial topic texts can also be the same, and no specific limitation is made here.

[0058] Step S14: Using the first language model, generate a research report on the target research topic based on several first-topic texts and second-topic texts related to the target research topic in the conference text.

[0059] In this implementation, a research report on the target research topic is generated using a first large language model based on several first-topic texts and second-topic texts related to the target research topic from the conference proceedings. The first large language model possesses powerful text generation capabilities, enabling it to process large amounts of text information quickly and generate a research report according to requirements. Furthermore, the large language model can comprehensively analyze and integrate the first and second-topic texts, organically incorporating them into the research report. This makes the research report more comprehensive and systematic in reflecting all aspects of the target research topic, improving the report's quality and value.

[0060] In one embodiment, the second largest language model is either the first largest language model or another largest language model other than the first largest language model, and no limitation is made here.

[0061] In one implementation, the research report format includes an abstract, overview, background description, industry information related to the target research topic, macroeconomic data, at least one report argument and corresponding supporting evidence, and a conclusion. The conclusion can be understood as the core viewpoints and conclusions of the research report. Furthermore, there is no limit to the number of report arguments; they can be set according to actual usage needs.

[0062] In one embodiment, at least one paragraph of the research report is identified with the reference data cited in that paragraph.

[0063] In one embodiment, the research report also includes an internet content source access interface corresponding to the reference data cited in the research report. This access interface can access the internet content source to obtain the reference data, which includes at least one of the following: web pages and documents in third-party knowledge bases. The research report includes an internet content source access interface corresponding to the reference data cited in the research report, allowing users to directly open and access the internet content source through this interface.

[0064] In one embodiment, a research report on the target research topic is generated using a first language model based on several first topic texts and second topic texts related to the target research topic in the conference text. Specifically, this can involve: comparing whether the arguments between the first topic texts and the second topic texts are consistent; in response to the consistency of arguments between the first topic texts and the second topic texts, using the consistent arguments between the first topic texts and the second topic texts as the report arguments of the research report; in response to the second topic texts containing supplementary arguments other than those in the first topic texts, combining the arguments of the first topic texts and the supplementary arguments contained in the second topic texts as the report arguments of the research report, or using the arguments of the first topic texts and the supplementary arguments contained in the second topic texts as different report arguments of the research report; in response to the conflicting arguments between the first topic texts and the second topic texts, listing the arguments of the first topic texts and the arguments of the second topic texts respectively as multiple report arguments of the research report.

[0065] When the arguments in the first and second topic texts contradict each other, listing the different arguments separately helps to comprehensively present the controversies and differing viewpoints surrounding the target research topic, making the research report more rigorous. In other words, the research report does not provide a definitive conclusion on the correctness of the arguments, but only provides clear information about the arguments and presents the different arguments in a structured manner, facilitating efficient reading for users.

[0066] If the second topic text contains supplementary arguments beyond those in the first topic text, combining these supplementary arguments with the arguments in the first topic text or using them separately as the main arguments of the report can enrich the content of the research report. For example, in researching the topic of "the development trend of online education," the first topic text mainly discusses "technological progress driving the personalized development of online education," while the second topic text supplements this with the argument "policy support promoting the equitable development of online education." Combining both as the main arguments of the research report will make the report's explanation of the development trend of online education more comprehensive and in-depth.

[0067] In one implementation, a triangulation method will be introduced to cross-validate and integrate multi-source data based on the target research topic, several first topic texts, and second topic texts related to the target research topic in the conference proceedings.

[0068] In one specific implementation, the triangulation method is used to decompose and cross-validate multi-source data in layers, specifically divided into three levels: internal insight layer, external evidence layer, and cross-validation layer.

[0069] The internal insights layer performs meticulous cleaning, noise reduction, and thematic coding on the transcribed meeting texts, eliminating redundant spoken language, ineffective small talk, and repetitive discussions. It also precisely extracts core business viewpoints, trend judgments, preliminary decision conclusions, and key arguments formed during the meeting discussions. The information from the internal insights layer originates from real business discussion scenarios, possessing strong relevance, alignment with business needs, and high decision-making reference value, thus providing a customized and precise core analytical framework for the overall research report.

[0070] The external evidence layer provides publicly available, objective, timely, and compliant external empirical support for the generation of research reports. It aggregates multi-source internet data, including web pages, industry documents, journal articles, market data, and policy documents, after timeliness filtering, vector retrieval, and paragraph extraction, forming an objective external empirical evidence library. The external evidence layer performs a comprehensive and refined purification of massive publicly available internet data sources, sequentially completing precise keyword retrieval, differentiated timeliness threshold filtering, semantic vector similarity matching, redundant content deduplication, and low-quality source filtering. It selectively integrates high-quality information from various types, including web pages, industry standard documents, academic journal articles, authoritative market statistics, policy and regulatory documents, and industry research reports, constructing a standardized and highly reliable external empirical evidence library. The information in the external evidence layer has the advantages of broad coverage, strong objectivity, excellent timeliness, and public traceability, effectively compensating for the problems of insufficient data volume, limited industry perspective, and lack of theoretical basis in internal meeting information.

[0071] The cross-validation layer is the core collaborative verification link of the triangular information fusion mechanism, realizing two-way linkage comparison and intelligent fusion of internal subjective judgment and external objective data. The cross-validation layer conducts two-way mutual verification analysis between the experience judgments and subjective opinions formed in internal meetings and external internet empirical data. It objectively categorizes three situations: consistent viewpoints, external supplementation and extension, and differing viewpoints, presenting only a structured presentation without subjective judgment of right or wrong, ensuring the neutrality and completeness of the subsequent research report. The cross-validation layer performs comprehensive semantic alignment and dimensional matching between the business experience, subjective judgments, and preliminary conclusions extracted from the internal insight layer and the standardized industry data, market rules, policy requirements, and academic viewpoints accumulated from the external evidence layer. It then structures and presents three types of information relationships: consistent internal and external viewpoints, external information supplementing and extending internal understanding, and internal and external viewpoints exhibiting dimensional differences or disagreements in judgment.

[0072] Please see Figure 2 , Figure 2This is a flowchart illustrating an embodiment of the method provided in this application for extracting topics based on topic prompts and meeting text using a second-largest language model. It should be noted that if substantially the same result is obtained, this embodiment is not necessarily identical. Figure 2 The illustrated process sequence is limited. For example... Figure 2 As shown, this embodiment includes: Step S21: Use the second language model to identify multiple topics from the conference text.

[0073] In this embodiment, multiple topics are identified from the meeting text.

[0074] Step S22: Analyze the selected topic to obtain its various research attributes.

[0075] In this implementation method, the selected topic is analyzed to obtain its various research attributes.

[0076] In one implementation, the research attributes include content complexity, which can be obtained by evaluating the semantic depth of the text related to the topic in the conference text and / or the degree of overlap in knowledge domains.

[0077] In one specific implementation, the topic-related text in the meeting text can be split into several sentences; from these sentences, sentences with complex sentence structures (e.g., long and difficult sentences) can be identified; and the proportion of sentences with complex sentence structures can be counted as the content complexity of the topic.

[0078] In one specific implementation, the topic-related text in the conference text can be split into several words, and the semantic vector of each word can be obtained. Based on the semantic vector of each word, the semantic similarity between all pairs of words in the conference text can be calculated. Cosine similarity can be used as a measure. The closer the similarity value is to 1, the more similar the two words are semantically; the closer it is to 0, the greater the semantic difference. According to the semantic similarity calculation results, a word semantic relationship network is constructed, and each word is regarded as a node in the network. The semantic similarity between words is used as the weight of the edge. Only when the semantic similarity between two words exceeds a preset threshold will a connection be established between their corresponding nodes.

[0079] The complexity of a topic can be assessed by counting the number of polysemous words in the selected text (the more polysemous words in the selected text, the more semantic ambiguity there is, and the higher the complexity of the topic); or by measuring the density of the semantic relationship network of words (indicating the tightness of the semantic connection between words) (the higher the density of the semantic relationship network of words, the tighter the semantic connection and the higher the degree of interweaving between words in the selected text, and the higher the complexity of the topic).

[0080] Specifically, the density of the semantic relationship network of words can be calculated using the following formula:

[0081] Where D represents the density of the semantic relation network of words; E represents the actual number of edges; This represents the maximum possible number of edges. In a word semantic relation network, there are n nodes, and at most one edge connects any two distinct nodes. Therefore, the maximum possible number of edges is... , Only when the semantic similarity between two words exceeds a preset threshold will a connection be established between their corresponding nodes, and only then will an edge exist between their corresponding nodes.

[0082] For example, to determine the total number of network nodes n: after segmenting and deduplicating the topic text, count the number of unique words (e.g., if there are 50 words after segmentation and 30 words after deduplication, n=30); count the actual number of edges E: iterate through the results of pairwise similarity calculations of all words and count the number of word pairs with similarity exceeding a preset threshold (undirected, AB and BA count as 1 pair, only 1 edge is recorded); substitute into the formula to calculate density: substitute n and E into the above formula to obtain the final word semantic relationship network density; association complexity: use the density value as the content complexity of the topic, the higher the density, the higher the content complexity of the topic.

[0083] In one implementation, the research attributes include problem openness, which can be obtained by evaluating the solution path and / or potential influencing factors of the selected topic.

[0084] In one specific implementation, the number of solution paths for a given topic can be used as a measure of its openness. The more solution paths a topic has, the higher its openness.

[0085] In one specific implementation, causal analysis can be used to identify various potential factors influencing the topic selection. For example, for a topic about "factors affecting student academic performance," the analysis can be conducted from multiple perspectives, including student-related factors (such as learning attitude, learning methods, and intelligence level), family factors (such as family economic status and parenting styles), school factors (such as teaching quality and school atmosphere), and social factors (such as social culture and employment situation). The number of complexly interrelated potential factors is determined as the openness of the topic. A large number of complexly interrelated potential factors indicates a high degree of openness to the topic, suggesting multiple possible solutions.

[0086] In one implementation, the research attribute includes knowledge openness, which can be obtained by assessing the availability of literature resources for the selected topic.

[0087] In one specific implementation, academic search engines (such as Google Scholar and Baidu Scholar) and professional databases (such as CNKI, Wanfang, and Web of Science) can be used to search for academic literature, research reports, industry information, and other materials related to the research topic. The number of high-quality materials found is used as a measure of the knowledge openness of the research topic. If the proportion of high-quality resources is large, it indicates that a large amount of high-quality related materials can be easily obtained, indicating that the knowledge openness of the research topic is high; conversely, it indicates that the knowledge openness is low.

[0088] In one implementation, the research attributes include breadth of knowledge, which can be obtained by assessing the scope of the research topic's domain and / or the sources of knowledge.

[0089] In one specific implementation, the subject areas involved in the selected topic can be determined by classifying and labeling the relevant texts by discipline. The number of subject areas involved in the topic can be counted as the breadth of knowledge of the topic; if a large number of subject areas are involved, it indicates a greater breadth of knowledge. For example, a topic on "the impact of climate change on ecosystems and socio-economics" involves multiple subject areas such as meteorology, ecology, economics, and sociology.

[0090] In one specific implementation, the sources of knowledge related to the chosen topic can be identified, including academic literature, industry reports, news media, and business practices. The number of different sources is counted to represent the breadth of knowledge for the chosen topic; a higher number of sources indicates a broader knowledge base. For example, when researching a company's management issues, analysis can be conducted by combining knowledge from multiple sources, such as academic theories, industry cases, and internal company data.

[0091] Step S23: Select a research topic from multiple options that meets the corresponding preset research requirements for each research attribute.

[0092] In this implementation, a research topic is selected from multiple options if all research attributes meet the corresponding preset research requirements. A topic is selected only if it meets all research attributes, which means that the selected research topic has good conditions in all key aspects, ensuring that the research has high feasibility and quality.

[0093] In one implementation, the research attribute meets the corresponding preset research requirements, which may be that the research attribute is greater than or equal to the corresponding preset research threshold.

[0094] In one implementation, the various research attributes of the selected topic can be weighted to obtain a total researchable score. Topics whose total researchable score meets the corresponding preset research requirements are selected as research topics. By assigning weights to different research attributes, the screening criteria can be adjusted according to actual research needs and the importance of each research attribute. This allows for a more reasonable assessment of the comprehensive value of the selected topic, highlighting the role of key research attributes while taking into account the influence of other research attributes.

[0095] In one specific implementation, the total researchable score satisfies the corresponding preset research requirements, which can be that the total researchable score is greater than or equal to the corresponding preset research score.

[0096] Please see Figure 3 , Figure 3 This is a flowchart illustrating an embodiment of the training method for the second largest language model provided in this application. It should be noted that if substantially the same result is obtained, this embodiment does not necessarily reflect that outcome. Figure 3 The illustrated process sequence is limited. For example... Figure 3 As shown, this embodiment includes: Step S31: Construct the dataset.

[0097] In this embodiment, a dataset is constructed; wherein, the dataset includes multiple sample texts, and each sample text is labeled with the target topic selection result.

[0098] There is no limit to the number of sample texts in the dataset; this can be set according to actual usage needs. For example, the number of sample texts in the dataset could be 10,000.

[0099] In one implementation, the target topic selection results of the sample text annotation include the research topic corresponding to the sample text.

[0100] Of course, in other implementations, the target topic selection results of sample text annotation may also include the research topic corresponding to the sample text and the keywords corresponding to each research topic, which is not specifically limited here.

[0101] Step S32: Divide the sample texts in the dataset into categories, and select a specific proportion of sample texts from each category to form a training set.

[0102] In this implementation, the sample texts in the dataset are divided into categories, and a specific proportion of sample texts are selected from each category to form a training set. After dividing the dataset into categories, sample texts are extracted to form the training set. Subsequently, the training set is used to train the base language model, enabling the base language model to learn to recognize and generate topic selection results corresponding to sample texts of each category.

[0103] Furthermore, after selecting a specific proportion of sample texts from each category to form the training set, the remaining sample texts will automatically become the test set or validation set. The distribution of the test set or validation set in each category is consistent with that of the training set, making the evaluation results of the second largest language model more credible and statistically significant.

[0104] There is no specific limit to the ratio; it can be set according to actual usage needs. For example, a specific ratio of 70% or 75% is acceptable. Furthermore, the specific ratios for different categories can be the same or different; this is not limited here.

[0105] For example, the sample texts in each category are divided into training set, validation set and test set according to a ratio of 7:1.5:1.5; then 70% of the sample texts in each category are randomly selected to form the training set.

[0106] In one implementation, the sample texts in the dataset can be divided according to the category of the target topic selection results labeled on them.

[0107] Step S33: Train the base large language model using the training set to obtain the second large language model.

[0108] In this embodiment, the base language model is trained using a training set to obtain a second large language model. Training the base language model with the training set ensures that the predicted topic selection results generated by the base language model based on the training set closely approximate the target topic selection results of the training set. Since the target topic selection results of the training set are the actual topic selection results, this process ensures that the predicted topic selection results generated by the base language model based on the training set closely approximate the actual topic selection results of the training set. In other words, it drives the predicted topic selection results generated by the base language model based on the training set to be as accurate as possible, enabling the base language model to accurately generate topic selection results.

[0109] In one implementation, the base large language model can be an open-source model such as DeepSeek, Large Language Model Meta AI (LLaMA), or QWen, and is not limited here.

[0110] In one implementation, the base large language model is trained using a training set and then trained using supervised fine-tuning (SFT).

[0111] Specifically, the hyperparameters are determined as follows: learning rate, batch size, and epochs. The learning rate controls the step size of parameter updates during training of the large language model. If the learning rate is set too large, the parameter update amplitude will be too large, potentially causing the large language model to fail to converge stably during training, or even oscillate or diverge. Conversely, if the learning rate is set too small, the parameter updates will be slow, significantly increasing training time and potentially leading to getting stuck in local optima, failing to find the globally optimal parameter configuration. Batch size refers to the number of data samples input to the large language model for each training iteration. A larger batch size can utilize the parallelism of matrix operations to improve computational efficiency and reduce training time, but it also increases memory consumption. Furthermore, a larger batch size may cause the large language model to become more prone to errors during training. It is easy to get stuck in local optima because the sample information used in each parameter update is relatively similar. Conversely, although the batch size is relatively less computationally efficient, it can provide richer gradient information, which helps the base language model to escape local optima and find better parameter configurations. The training epochs represent the number of times the base language model has completely traversed the entire dataset. Too few training epochs may prevent the base language model from fully learning the features and patterns in the data, resulting in poor performance. Too many training epochs may lead to overfitting, that is, the base language model performs well on the training data but poorly on the test data, because the base language model has overlearned the noise and features of specific samples in the training data and cannot generalize well to new data.

[0112] Training begins and the process is monitored. After determining the initial hyperparameters, the Base Language Model begins training based on the training set. The Base Language Model calculates the predicted topic selection result through forward propagation based on the input sample text. Then, it compares the predicted topic selection result with the labeled target topic selection result in the dataset, calculating the value of the loss function. The loss function measures the difference between the Base Language Model's predicted topic selection result and the target topic selection result. The goal of the Base Language Model is to minimize the value of the loss function by continuously adjusting the parameters. Next, the Base Language Model calculates the gradient of the loss function with respect to each parameter using the backpropagation algorithm. Then, it updates the parameters of the Base Language Model based on the learning rate and gradients, making the predicted topic selection result of the Base Language Model closer to the true target topic selection result. This process is repeated until the Base Language Model completes all preset training epochs.

[0113] Monitoring the training process is crucial. During training, the performance of the base language model needs close monitoring, primarily by observing changes in the loss function to determine its training status. Normally, as the number of training epochs increases, the loss function should gradually decrease, indicating that the base language model is continuously learning features and patterns in the data, and its performance is gradually improving.

[0114] Addressing issues encountered during training: When the loss function of the base language model stops decreasing or even increases during training, it indicates that the model may be failing to converge. This could be due to an excessively high learning rate, problematic data, or an unreasonable model structure. To address this, try reducing the learning rate to make the model's parameters update more slowly and stably; check for noise or mislabeled data and clean and correct it; or adjust the model's structure by increasing or decreasing the number of layers or neurons to improve its expressive power. If the model performs well on the training set with a low loss function, but performs poorly on the validation or test sets with a high loss function, it indicates overfitting. To address overfitting, regularization methods such as L1 and L2 regularization can be employed. These methods, by adding a regularization term to the loss function, limit the size of the base language model's parameters, preventing excessive complexity. Early stopping can also be used, stopping training prematurely when performance on the validation set no longer improves, preventing the base language model from continuing to learn noise and specific sample features from the training data. Furthermore, increasing the amount of data and performing data augmentation can effectively alleviate overfitting. Gradient explosion refers to the phenomenon where gradient values ​​become extremely large during training, leading to excessively large parameter updates in the base language model, even resulting in numerical overflow. This is usually caused by a deep base language model structure, an excessively large learning rate, or the presence of outliers in the data. Gradient clipping can be used, which involves setting a gradient threshold and clipping gradient values ​​to within that threshold range to prevent excessively large gradients from causing abnormal parameter updates. Alternatively, reducing the learning rate can also be attempted to decrease the magnitude of parameter updates. When a large-scale language model encounters problems such as convergence failure, overfitting, or gradient explosion, a detailed analysis of the bad cases encountered during training is necessary. By examining these bad cases, the reasons for the model's prediction errors can be identified, such as special cases in the data, labeling errors, or insufficient understanding of certain features by the model. These bad cases are then relabeled with more accurate labels to enable the model to better learn the features of these samples. Based on the analysis of bad cases and the problems encountered during training, the initial hyperparameter settings should be adjusted. For example, if overfitting is found, the learning rate can be appropriately reduced, the regularization coefficient increased, or early stopping can be used; if the model fails to converge, the learning rate can be reduced or the data itself can be examined for problems.After adjusting the hyperparameters, restart the training process and monitor the training status of the base large language model again until the base large language model can converge stably and achieve good performance on the validation and test sets.

[0115] In one implementation, after extracting topics using a second language model based on topic prompts and conference text, and obtaining the topic extraction results, reinforcement learning is applied to the second language model based on user adjustments to the topic extraction results and / or user selection of target research topics. In other words, user adjustments to the topic extraction results and / or user selection of target research topics serve as a reward mechanism fed back to the second language model for reinforcement learning, continuously iterating the parameter weights of the second language model.

[0116] User adjustments to the topic extraction results provide direct feedback to the second language model. If a user corrects the topic extraction results, the second language model can learn from these corrections, identifying which topic extraction results were incorrect and avoiding similar errors in subsequent topic extraction processes. The user's choice of target research topics helps the second language model better understand user needs and preferences. Based on user choices, the second language model can learn which types of topics are more popular with users, thus focusing more on these aspects when extracting topics and improving the match between topics and user needs.

[0117] In one embodiment, the steps of training the first language model to enable it to generate research reports based on topic texts include: Step 1: Constructing a dataset, wherein the dataset includes multiple sample topic texts, each of which is labeled with a target research report; it should be noted that the remaining related content of Step 1 can be found in Step S31, and will not be repeated here. Step Two: Divide the sample texts in the dataset into categories, and select a specific proportion of sample texts from each category to form the training set. See step S32 for details on step two, which will not be repeated here. Step Three: Train the first language model using the training set. This training aims to make the predicted research reports generated by the first language model based on the training set approximate the target research reports in the training set. Since the target research reports in the training set are the actual research reports in the training set, this ensures that the predicted research reports generated by the first language model based on the training set approximate the actual research reports in the training set, thus driving the first language model to generate research reports as accurately as possible. It should be noted that the remaining details of step three can be found in step S33, and will not be repeated here.

[0118] Please see Figure 4 , Figure 4 This is a schematic diagram of an embodiment of the research report generation device provided in this application. The research report generation device 40 includes an acquisition module 41, an extraction module 42, a search module 43, and a report generation module 44. The acquisition module 41 is used to acquire meeting text, wherein the meeting text is obtained by transcribing meeting recording data. The extraction module 42 is used to generate topic prompts and use a second language model to extract topics based on the topic prompts and the meeting text, obtaining topic extraction results. The topic prompts include preset research requirements that each research attribute must meet, and the topic extraction results include several research topics. The search module 43 is used to select at least one research topic as a target research topic, and use the keywords corresponding to the target research topic to search in at least one Internet data source to obtain several first topic texts related to the target research topic. The report generation module 44 is used to use a first language model to generate a research report on the target research topic based on the several first topic texts and second topic texts related to the target research topic in the meeting text.

[0119] Please see Figure 5 , Figure 5 This is a schematic diagram of an embodiment of the electronic device provided in this application. The electronic device 50 includes a memory 51 and a processor 52 coupled to each other. The processor 52 is used to execute program instructions stored in the memory 51 to implement the steps of any of the above-described research report generation method embodiments. In a specific implementation scenario, the electronic device 50 may include, but is not limited to, a microcomputer or a server. In addition, the electronic device 50 may also include mobile devices such as laptops and tablets, which are not limited here.

[0120] Specifically, processor 52 controls itself and memory 51 to implement the steps of any of the above-described research report generation method embodiments. Processor 52 may also be referred to as a CPU (Central Processing Unit). Processor 52 may be an integrated circuit chip with signal processing capabilities. Processor 52 may also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor. Furthermore, processor 52 may be implemented using integrated circuit chips.

[0121] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of an embodiment of the computer-readable storage medium provided in this application. The computer-readable storage medium 60 of this application embodiment stores program instructions 61. When executed, these program instructions 61 implement the methods provided by any embodiment and any non-conflicting combination of the research report generation method of this application. The program instructions 61 can form a program file and be stored in the aforementioned computer-readable storage medium 60 in the form of a software product, so that a computer device (which may be a personal computer, server, or network device, etc.) can execute all or part of the steps of the methods of various embodiments of this application. The aforementioned computer-readable storage medium 60 includes various media capable of storing program code, such as a USB flash drive, mobile hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or terminal devices such as computers, servers, mobile phones, and tablets.

[0122] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0123] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for generating research reports, characterized in that, include: Obtain the meeting transcript, wherein the meeting transcript is obtained based on the transcription of meeting recording data; Generate topic selection prompts, which include preset research requirements that each research attribute must meet; The second language model is used to extract research topics based on the topic selection prompts and the conference text, resulting in a number of research topics. Take at least one of the research topics as the target research topic, and use the keywords corresponding to the target research topic to search in at least one Internet data source to obtain several first topic texts related to the target research topic; Using the first major language model, a research report on the target research topic is generated based on several first-topic texts and second-topic texts related to the target research topic in the conference text.

2. The method according to claim 1, characterized in that, The research attributes include at least one of the following: content complexity, problem openness, knowledge openness, and knowledge breadth, wherein the content complexity characterizes the depth of the topic, the problem openness characterizes the diversity of solutions to the topic, the knowledge openness characterizes the accessibility of the knowledge involved in the topic, and the knowledge breadth characterizes the breadth of knowledge sources involved in the topic.

3. The method according to claim 1, characterized in that, The topic extraction results also include keywords corresponding to each research topic; And / or, the second largest language model is the first largest language model or another largest language model other than the first largest language model.

4. The method according to claim 1, characterized in that, The method further includes: Construct a dataset, which includes multiple sample texts, each of which is labeled with the target topic selection result; The sample texts in the dataset are divided into categories, and a specific proportion of the sample texts from each category are selected to form a training set. The base large language model is trained using the training set to obtain the second large language model; After obtaining the topic extraction result by using the second language model based on the topic selection prompts and the meeting text, the method further includes: The second language model is subjected to reinforcement learning based on the user's adjustment of the topic extraction results and / or the user's selection of the target research topic.

5. The method according to claim 1, characterized in that, The keywords corresponding to the target research topic include original topic words and extended topic words. The original topic words include subject and theme-related words in the target research topic. The extended topic words include at least one of the following: a first extended word related to the subject based on the subject fission, and a second extended word related to the theme-related words based on the theme fission. And / or, before searching in at least one internet data source using keywords corresponding to the target research topic to obtain several first topic texts related to the target research topic, the method further includes: Obtain at least one keyword corresponding to the research topic, and present at least one keyword corresponding to the research topic to the user; In response to the user's adjustment of keywords corresponding to the research topic and / or the user's addition of keywords to the research topic, the keywords corresponding to the research topic are updated.

6. The method according to claim 1, characterized in that, The at least one internet data source includes API interfaces provided by search engines and third-party knowledge bases; The method involves searching at least one internet data source using keywords corresponding to the target research topic to obtain several first topic texts related to the target research topic, including: Using the API interface provided by the search engine, search for several webpage links that match keywords corresponding to the target research topic, and access the corresponding webpages through each of the webpage links, extracting the first topic text related to the target research topic from each accessed webpage; and, Search the third-party knowledge base for several documents that match the keywords corresponding to the target research topic, and extract the first topic text related to the target research topic from each of the matched documents.

7. The method according to claim 6, characterized in that, The step of extracting first topic text related to the target research topic from each of the visited web pages, or the step of extracting first topic text related to the target research topic from each of the matched documents, includes: Each of the aforementioned web pages or documents is used as the target text; The target text is segmented, and each segment of the target text is converted into a first embedding vector; Obtain the second embedding vector of the keywords corresponding to the target research topic; Obtain the vector similarity between the second embedding vector and the first embedding vector of each segment, and select the segment whose vector similarity is greater than the similarity threshold as the first topic text; Alternatively, the step of extracting first topic text related to the target research topic from each of the accessed web pages, or the step of extracting first topic text related to the target research topic from each of the matched documents, includes: Each of the aforementioned web pages or documents is used as the target text; The keywords corresponding to the target research topic are expanded with synonyms to obtain at least one synonym. The keywords corresponding to the target research topic and the at least one synonym are combined to form a search term set. The target text is segmented into words to obtain several words of the target text; The position of the word segment that matches each search term in the search term set in the target text is determined as the hit position; The paragraphs in the target text whose number of hit locations exceeds a preset threshold are selected as the first topic text.

8. The method according to claim 1, characterized in that, The research report also includes an internet content source access interface corresponding to the reference data cited in the research report. The access interface can access the internet content source to obtain the reference data, which includes at least one of the following: web pages and documents in third-party knowledge bases. And / or, the step of generating a research report on the target research topic using a first language model based on several first topic texts and second topic texts related to the target research topic in the conference text includes: Compare whether the arguments in the first topic text and the second topic text are consistent; In response to the consistency of arguments between the first topic text and the second topic text, the arguments that are consistent between the first topic text and the second topic text shall be taken as the report arguments of the research report; In response to the second topic text containing supplementary arguments other than the first topic text, the arguments of the first topic text and the supplementary arguments contained in the second topic text are combined into the report arguments of the research report, or the arguments of the first topic text and the supplementary arguments contained in the second topic text are respectively used as different report arguments of the research report; In response to a conflict between the arguments in the first topic text and the second topic text, the arguments in the first topic text and the second topic text are listed respectively as multiple report arguments of the research report.

9. The method according to claim 1, characterized in that, The acquisition of meeting text includes: Collect meeting recording data during the meeting; The meeting recording data is transcribed into a transcribed text; The transcribed text may be used as the meeting transcript, or the transcribed text may be converted into a structured minutes to serve as the meeting transcript. And / or, the step of selecting at least one of the research topics as the target research topic includes: In response to a user's request to view research topics, the aforementioned research topics are displayed; In response to a user's selection of at least one of the research topics, the research topic selected by the user is taken as the target research topic.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing program instructions, and the processor executing the program instructions to implement the research report generation method as described in any one of claims 1-9.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program instructions that can be executed to implement the research report generation method as described in any one of claims 1-9.

Images