A Method and System for Training Scientific and Technological Intelligence Agents Based on Reinforcement Learning

Abstract

This invention provides a training method and system for a science and technology intelligence agent based on reinforcement learning, relating to the field of reinforcement learning technology. First, it determines the multi-stage evolution characteristics of the intelligence task corresponding to the training of the science and technology intelligence agent, including initial exploration, deep analysis, and result output stage attributes. Then, it constructs a reinforcement learning exploration space adapted to this space, containing optional sets of behaviors for each stage and stage transition conditions. Next, it generates staged intelligence sample clusters adapted to this reinforcement learning exploration space, and calls the agent to interact with the sample clusters in the exploration space stage by stage to form a cross-stage behavior sequence. Then, it constructs a multi-stage reward function based on the multi-stage evolution characteristics, generates stage reward signals based on the cross-stage behavior sequence, and updates the agent's decision parameters, completing a single training iteration. This effectively improves the agent's processing capability in a multi-stage science and technology intelligence task environment.

Description

Technical Field

[0001] This invention relates to the field of reinforcement learning technology, and more specifically, to a method and system for training a technology intelligence agent based on reinforcement learning. Background Technology

[0002] In today's rapidly developing technological landscape, the acquisition, analysis, and utilization of scientific and technological intelligence play a crucial role in scientific research decision-making, corporate strategic planning, and the formulation of national science and technology policies. Traditional methods of processing scientific and technological intelligence primarily rely on manual collection and analysis. These methods are not only inefficient but also susceptible to human interference, making it difficult to guarantee the accuracy and comprehensiveness of the intelligence processing.

[0003] With the development of artificial intelligence technology, some scientific and technological intelligence processing systems based on simple machine learning algorithms have gradually emerged. While these systems can automate the processing of scientific and technological intelligence to a certain extent, they typically employ fixed processing patterns and lack the ability to adapt to the dynamic and changing characteristics of scientific and technological intelligence tasks. Scientific and technological intelligence tasks often exhibit multi-stage evolution, with significant differences in the requirements and behavioral choices for intelligence processing at different stages. For example, the initial exploration stage requires the extensive collection of various relevant intelligence, while the in-depth analysis stage focuses on in-depth mining of key information. Existing technologies struggle to flexibly adjust processing strategies according to these multi-stage characteristics, resulting in an inability to efficiently and accurately complete intelligence tasks in complex and ever-changing scientific and technological intelligence environments, thus failing to meet the needs of practical applications. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for training a technology intelligence agent based on reinforcement learning, the method comprising:

[0005] The multi-stage evolution characteristics of intelligence tasks corresponding to the training of science and technology intelligence intelligent agents are determined. The multi-stage evolution characteristics of intelligence tasks include attributes of the initial exploration stage, attributes of the deep analysis stage, and attributes of the result output stage.

[0006] Construct a reinforcement learning exploration space adapted to the multi-stage evolution characteristics of the intelligence mission, wherein the reinforcement learning exploration space includes the set of possible behaviors and stage transition conditions corresponding to each stage;

[0007] Generate a phased intelligence sample cluster adapted to the reinforcement learning exploration space. The phased intelligence sample cluster includes an initial exploration phase intelligence sample subset corresponding to the initial exploration phase attribute, a deep analysis phase intelligence sample subset corresponding to the deep analysis phase attribute, and a result output phase intelligence sample subset corresponding to the result output phase attribute.

[0008] The technology intelligence agent is invoked to interact with the phased intelligence sample clusters in the reinforcement learning exploration space in stages, forming a cross-stage agent behavior sequence;

[0009] A multi-stage reward function is constructed based on the multi-stage evolution characteristics of the intelligence task. Stage reward signals are generated according to the agent's behavior sequence. The decision parameters of the science and technology intelligence agent are updated using the stage reward signals to complete a single training iteration.

[0010] Furthermore, embodiments of the present invention also provide a training system for a technology intelligence agent based on reinforcement learning, characterized in that it includes:

[0011] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described reinforcement learning-based scientific and technological intelligence agent training method by executing the machine-executable instructions.

[0012] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, a processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, and the processor executing the machine-executable instructions, causing the computer device to execute the above-described reinforcement learning-based scientific and technological intelligence agent training method.

[0013] Based on the above, by identifying the multi-stage evolution characteristics of intelligence tasks corresponding to the training of science and technology intelligence agents, including initial exploration, in-depth analysis, and result output stages, the dynamic change patterns of science and technology intelligence tasks at different stages are accurately grasped. A reinforcement learning exploration space adapted to the multi-stage evolution characteristics is constructed, clarifying the behavioral options and stage transition conditions for each stage. This enables the agent to explore and make decisions within a reasonable range, improving the targeting and effectiveness of training. A staged intelligence sample cluster adapted to the reinforcement learning exploration space is generated, providing the agent with training data matching the characteristics of each stage. This helps the agent better understand the intelligence characteristics and processing needs at different stages. The science and technology intelligence agent interacts with the sample clusters in the exploration space stage by stage to form a cross-stage agent behavior sequence, realizing continuous learning and experience accumulation at different stages. A multi-stage reward function is constructed based on the multi-stage evolution characteristics, and the decision parameters of the agent are updated using stage reward signals. This allows for precise feedback and optimization based on the agent's performance at different stages, enabling the agent to continuously adapt to the requirements of multi-stage intelligence tasks and significantly improving the processing capability and adaptability of the science and technology intelligence agent in complex multi-stage task environments. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the execution flow of the reinforcement learning-based scientific and technological intelligence agent training method provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of exemplary hardware and software components of the reinforcement learning-based intelligent agent training system for scientific and technological intelligence provided in an embodiment of the present invention. Detailed Implementation

[0016] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a reinforcement learning-based training method for a scientific and technological intelligence agent, as provided in one embodiment of the present invention. The following is a detailed description of this reinforcement learning-based training method for a scientific and technological intelligence agent.

[0017] Step S110: Determine the multi-stage evolution characteristics of the intelligence task corresponding to the training of the science and technology intelligence agent. The multi-stage evolution characteristics of the intelligence task include the initial exploration stage attributes, the deep analysis stage attributes, and the result output stage attributes.

[0018] In this embodiment, the "Analysis of the Development Trends of Natural Language Processing Technology" in the field of artificial intelligence is used as the unified application scenario throughout the text. When determining the multi-stage evolution characteristics of intelligence tasks, it is first necessary to clarify that the overall goal of the intelligence task in this scenario is to comprehensively and deeply analyze the development history, current status, and future trends of natural language processing technology.

[0019] The initial exploration phase is the starting point of an intelligence mission, and its core lies in the extensive collection and preliminary screening of basic information in the field of natural language processing (NLP) technology. Specifically, this phase requires covering fundamental intelligence features such as basic concepts of NLP, major technical branches (e.g., machine translation, text classification, sentiment analysis), relevant research institutions, and key researchers.

[0020] The attributes of the in-depth analysis stage are those that involve further mining and analysis of the acquired intelligence, building upon the initial exploration. In the scenario of analyzing the development trends of natural language processing technology, this stage needs to focus on advanced intelligence characteristics such as the evolution path of technology, the connections and influences between different technological branches, key technological breakthroughs, and existing problems and challenges in current research. The aim is to reveal the inherent laws and potential trends of technological development.

[0021] The output stage attributes are those that organize, summarize, and present the intelligence obtained from in-depth analysis. For the task of analyzing the development trend of natural language processing technology, this stage needs to output the analysis results in a clear, accurate, and easy-to-understand way, such as forming technology development trend reports, key technology indicator comparison charts, application scenario matching suggestions, and other end-intelligence features, so that users can quickly understand and effectively utilize the analysis results.

[0022] Step S120: Construct a reinforcement learning exploration space adapted to the multi-stage evolution characteristics of the intelligence task. The reinforcement learning exploration space includes the set of possible behaviors and stage transition conditions corresponding to each stage.

[0023] Step S121: Analyze the initial exploration stage attributes in the multi-stage evolution characteristics of the intelligence mission, and extract the core objectives of the initial exploration stage. The core objectives of the initial exploration stage include the initial screening of intelligence sources, the rapid identification of key information, and the elimination of invalid information.

[0024] In the context of analyzing the development trends of natural language processing (NLP) technology, the initial exploration phase is analyzed for its fundamental role in the overall intelligence task. A thorough analysis of this attribute clarifies the core objectives of the initial exploration phase. Specifically, the initial intelligence source screening aims to select high-quality intelligence sources relevant to NLP technology from numerous potential sources, such as authoritative academic databases, well-known technical blogs, and industry research report publishing platforms, ensuring a reliable foundation for subsequent analysis. The rapid identification of key information requires the agent to quickly locate and extract key information related to core concepts, major technical branches, and important events in NLP technology from the selected intelligence sources, improving intelligence acquisition efficiency. The elimination of invalid information aims to remove content irrelevant to the analysis of NLP technology development trends, such as advertisements, irrelevant technical introductions, and redundant descriptions, ensuring the purity and effectiveness of the intelligence.

[0025] Step S122: Based on the core objectives of the initial exploration phase, determine the basic behaviors that the science and technology intelligence agent can perform in the initial exploration phase, and form a set of optional behaviors for the initial exploration phase. The basic behaviors that can be performed in the initial exploration phase include sample unit calling behavior, keyword matching behavior, and information fragment marking behavior.

[0026] Step S1221: For the initial screening of intelligence sources in the initial exploration phase, analyze the operations that the science and technology intelligence agent needs to perform, and define the specific manifestations of sample unit calling behavior. The specific manifestations of sample unit calling behavior include selecting a single intelligence sample unit from the phased intelligence sample cluster, batch retrieving similar intelligence sample units, and retrieving relevant intelligence sample units based on topic tags.

[0027] In the scenario of analyzing the development trend of natural language processing (NLP) technology, the specific manifestations of sample unit retrieval behavior need to be diversified to meet different screening needs for the initial screening of intelligence sources. Selecting a single intelligence sample unit from a phased intelligence sample cluster is suitable for preliminary evaluation and review of specific intelligence sources. For example, selecting an abstract of a paper on the latest NLP technology from a conference proceedings to determine whether the conference proceedings have value as an intelligence source. Batch retrieval of similar intelligence sample units can improve screening efficiency. For instance, simultaneously retrieving abstracts of papers related to NLP technology from multiple different academic journals allows for comparison of overall quality and relevance. Retrieving relevant intelligence sample units based on topic tags utilizes pre-set topic tags (such as "Natural Language Processing - Machine Translation" or "NLP - Text Generation") to quickly locate and retrieve sample units that match the theme of NLP technology development trend analysis, further narrowing the screening scope.

[0028] Step S1222: For the sample unit calling behavior, set the triggering condition for the execution of the sample unit calling behavior. The triggering condition for the execution of the sample unit calling behavior is that the sample unit calling behavior is automatically triggered when the science and technology intelligence agent enters the initial exploration stage and has not loaded any intelligence sample units.

[0029] In the context of analyzing the development trends of natural language processing technology, the triggering conditions for sample unit invocation need to be clear and easily determined. When a science and technology intelligence agent begins its initial exploration phase and no intelligence sample units for analysis have been loaded yet, the system will automatically trigger sample unit invocation to initiate subsequent operations such as key information identification and invalid information elimination. This triggering condition ensures that the agent can proactively acquire intelligence samples during the initial exploration phase, avoiding task stagnation due to a lack of processing objects.

[0030] Step S1223: Quickly identify targets based on key information in the initial exploration phase, analyze the operations that the science and technology intelligence agent needs to perform, and define the specific manifestations of keyword matching behavior. The specific manifestations of keyword matching behavior include comparing preset technical keywords with words in intelligence sample units, extracting terms in intelligence sample units that meet preset standards as candidate keywords, and verifying the relevance of candidate keywords to the task theme.

[0031] In the scenario of analyzing the development trends of natural language processing (NLP) technologies, keyword matching needs to be conducted from multiple dimensions to achieve rapid identification of key information. Comparing preset technical keywords with the vocabulary in the intelligence sample units is a basic and direct approach. Preset technical keywords cover core terms, technology names, and research methods in the field of NLP, such as "BERT model," "attention mechanism," and "pre-trained language model." By comparing the occurrence of these keywords in the sample units, a preliminary judgment can be made as to whether the sample units contain key information. Extracting terms from the intelligence sample units that meet preset criteria as candidate keywords can uncover potentially important terms within the sample units. These terms may not be included in the preset keyword list but are crucial for understanding the content of the sample units and identifying key information. For example, in the abstract of a paper on a new NLP technology, a high frequency of a certain emerging technology term may be key information reflecting the latest developments in that technology. Verifying the relevance of candidate keywords to the task topic ensures that the extracted candidate keywords are indeed relevant to the task of analyzing the development trends of NLP technologies, avoiding the introduction of irrelevant terms and guaranteeing the accuracy of key information identification.

[0032] Step S1224: For keyword matching behavior, set the triggering condition for keyword matching behavior execution. The triggering condition for keyword matching behavior execution is that after the science and technology intelligence agent completes the sample unit calling behavior and loads the intelligence sample unit, the keyword matching behavior is automatically triggered.

[0033] In the context of analyzing the development trends of natural language processing (NLP) technology, the triggering of keyword matching depends on the loading of sample units. Once the science and technology intelligence agent completes the sample unit retrieval process—that is, successfully retrieves and loads intelligence sample units related to NLP technology from the intelligence source—the system automatically triggers keyword matching to promptly identify key information within the sample units. This triggering mechanism ensures the continuity and timeliness of the key information identification process, enabling the agent to extract key information immediately after acquiring the sample units.

[0034] Step S1225: Targeting the elimination of invalid information in the initial exploration phase, analyze the operations that the science and technology intelligence agent needs to perform, and define the specific manifestations of information fragment marking behavior. The specific manifestations of information fragment marking behavior include marking paragraphs in intelligence sample units that are irrelevant to the task topic, marking recurring information content, and marking statements whose credibility does not meet the preset requirements.

[0035] In the analysis of natural language processing (NLP) technology development trends, information fragment labeling is a crucial operation for eliminating invalid information. Labeling paragraphs irrelevant to the task's theme within intelligence sample units is essential. For example, in a comprehensive article containing NLP technology and other unrelated technologies (such as computer vision), paragraphs related to computer vision are labeled as irrelevant information for later exclusion. Labeling recurring information, such as identical descriptions of basic NLP concepts appearing in multiple sample units, avoids redundant processing. Labeling statements lacking credibility, such as exaggerated claims about breakthroughs in NLP technology from non-authoritative sources without data support, allows for more cautious handling or exclusion of unreliable information in subsequent analysis, ensuring the quality of the intelligence.

[0036] Step S1226: For the information fragment marking behavior, set the triggering conditions for the execution of the information fragment marking behavior. The triggering conditions for the execution of the information fragment marking behavior are that the information fragment marking behavior is automatically triggered after the science and technology intelligence agent completes the keyword matching behavior and identifies key information.

[0037] In the context of analyzing the development trends of natural language processing technology, the triggering of information fragment labeling should occur after key information identification. Once the science and technology intelligence agent completes keyword matching and successfully identifies key information within the intelligence sample unit, the system automatically triggers information fragment labeling to eliminate invalid information. At this point, the agent has clearly identified the key content within the sample unit, enabling it to more accurately determine which information fragments are invalid and label them accordingly, thereby improving the efficiency and accuracy of invalid information removal.

[0038] Step S1227: Collect the specific manifestations and triggering conditions of sample unit calling behavior, keyword matching behavior, and information fragment marking behavior to form a list of behavior descriptions.

[0039] In the context of analyzing the development trends of natural language processing technology, to comprehensively and clearly define the optional set of behaviors in the initial exploration phase, it is necessary to systematically collect the specific manifestations and triggering conditions of sample unit invocation behavior, keyword matching behavior, and information fragment marking behavior. For sample unit invocation behavior, this involves collecting its specific manifestations, such as selecting a single intelligence sample unit from a phased intelligence sample cluster, batch retrieving similar intelligence sample units, and searching for related intelligence sample units based on topic tags, as well as the conditions for automatic triggering when entering the initial exploration phase without loading any intelligence sample units. Similarly, for keyword matching behavior, this involves collecting its specific manifestations, such as comparing preset technical keywords with words in intelligence sample units, and the conditions for automatic triggering after completing the sample unit invocation behavior and loading the intelligence sample unit. Likewise, for information fragment marking behavior, this involves collecting its specific manifestations and the conditions for automatic triggering after completing the keyword matching behavior and identifying key information.

[0040] Step S1228: Deduplicate and logically sort the behaviors in the behavior description list to ensure that there are no conflicts between the behaviors and that they conform to the execution flow of the initial exploration phase.

[0041] In the context of analyzing the development trends of natural language processing technology, the list of behavior descriptions may contain duplicate descriptions or logically inconsistent orders between behaviors. To address this, the list is deduplicated, removing identical or substantially similar behavior descriptions to ensure the uniqueness of each behavior. Then, a logical ordering process is performed, following the natural flow of intelligence processing in the initial exploration phase: first, sample unit calls are made to obtain intelligence samples; then, keyword matching is performed to identify key information; finally, information fragment labeling is performed to eliminate invalid information and rank the behaviors in the list. This ensures that there are no logical conflicts between behaviors and that the actions follow the normal order in which the agent performs tasks during the initial exploration phase, enabling the agent to work systematically.

[0042] Step S1229: Determine the sorted list of behavior descriptions as the initial exploration phase behavior option set, and clarify that the execution object of each behavior in the initial exploration phase behavior option set is the initial exploration phase intelligence sample subset in the phased intelligence sample cluster.

[0043] In the scenario of analyzing the development trend of natural language processing technology, the list of behavior descriptions, after deduplication and logical sorting, clearly outlines the basic behaviors that a scientific intelligence agent can execute in the initial exploration phase, along with their triggering conditions and specific manifestations. This list is formally designated as the action selection set for the initial exploration phase, providing a clear range of choices for the agent's actions at this stage. Simultaneously, it is clarified that the execution target of each behavior in this action selection set is a subset of the initial exploration phase intelligence samples within the phased intelligence sample cluster. That is, all basic actions of the agent revolve around the intelligence sample units within this subset, ensuring the targeted and effective execution of behaviors and avoiding operations on irrelevant samples.

[0044] Step S123: Analyze the deep analysis stage attributes in the multi-stage evolution characteristics of the intelligence mission, and extract the core objectives of the deep analysis stage. The core objectives of the deep analysis stage include intelligence correlation mining objectives, technology trend prediction objectives, and contradictory information identification objectives.

[0045] In the analysis of natural language processing (NLP) technology development trends, when analyzing the attributes of the deep analysis stage, the focus is on its role in the in-depth processing and value mining of intelligence. By analyzing these attributes, the core objectives of the deep analysis stage are extracted. The intelligence correlation mining objective aims to discover the inherent connections between different intelligence reports, such as the mutual learning between different research methods in NLP technology and the synergistic effects of different technological branches, thereby forming a holistic understanding of technological development. The technology trend prediction objective, based on existing intelligence and the discovered correlations, predicts the future development direction of NLP technology, potential new technologies, and the expansion of technology applications, providing users with forward-looking intelligence support. The contradictory information identification objective identifies and analyzes contradictory information within the intelligence, such as different evaluations of the performance of the same NLP technology in different research reports, or different opinions from different experts on technological development trends. By identifying these contradictions, the underlying reasons and essence of the information are revealed, improving the accuracy and reliability of intelligence analysis.

[0046] Step S124: Based on the core objectives of the deep analysis phase, determine the advanced behaviors that the science and technology intelligence agent can perform in the deep analysis phase, and form a set of optional behaviors for the deep analysis phase. The advanced behaviors that can be performed in the deep analysis phase include sample correlation calculation behavior, time series feature extraction behavior, and information credibility assessment behavior.

[0047] In the context of analyzing the development trends of natural language processing (NLP) technology, based on the core objectives of the deep analysis phase, the science and technology intelligence agent needs to possess more advanced behavioral capabilities. The sample correlation calculation behavior quantifies the degree of correlation between different intelligence samples. For example, by calculating the thematic similarity and citation relationships between two papers on different branches of NLP technology, the closeness of their correlation can be measured to achieve the goal of intelligence correlation mining. The temporal feature extraction behavior targets the time dimension of NLP technology development, extracting the characteristics of changes in technical indicators over time and the temporal distribution characteristics of research hotspots from intelligence samples, providing data support for predicting technology trends. The information credibility assessment behavior evaluates the credibility of intelligence information by assessing the authority of the intelligence source, the completeness and logic of the data, and the rationality of the argumentation process, thereby effectively identifying contradictory information. Integrating these advanced behaviors forms the optional set of behaviors for the deep analysis phase.

[0048] Step S125: Analyze the result output stage attributes in the multi-stage evolution characteristics of the intelligence task, and extract the core objectives of the result output stage. The core objectives of the result output stage include intelligence summary generation objectives, conclusion visualization objectives, and application scenario matching objectives.

[0049] In the scenario of analyzing the development trend of natural language processing technology, when parsing the attributes of the output stage, the focus is on how to transform the complex intelligence obtained from in-depth analysis into a form that is easy for users to understand and use. By analyzing this attribute, the core objectives of the output stage are clarified. The intelligence summary generation objective requires condensing and refining the large amount of intelligence content after in-depth analysis to form a concise, accurate, and comprehensive summary, summarizing the key information and main conclusions of the development of natural language processing technology, so that users can quickly grasp the core content. The conclusion visualization objective is to display the analysis results and data in intuitive ways such as charts and graphs, such as technology development timeline charts and bar charts comparing the performance of different technology branches, making complex data and relationships clearer and easier to understand. The application scenario matching objective is to match potential application scenarios based on the analyzed characteristics and development trends of natural language processing technology, such as intelligent customer service, machine translation, and intelligent recommendation systems, and explain the applicability and advantages of different technology branches in each application scenario, providing guidance for users' practical applications.

[0050] Step S126: Based on the core objectives of the result output stage, determine the terminal behaviors that the science and technology intelligence agent can perform in the result output stage, forming a set of optional behaviors for the result output stage. The terminal behaviors that can be performed in the result output stage include content condensation behavior, chart generation behavior, and scene tag matching behavior.

[0051] In the context of analyzing the development trends of natural language processing technology, based on the core objective of the results output stage, the science and technology intelligence agent needs to perform corresponding terminal behaviors to complete the intelligence output. Content condensation behavior can highly summarize and simplify the intelligence after in-depth analysis, removing redundant information, retaining core viewpoints and key data, and generating a concise summary that meets the intelligence summary generation objective. Chart generation behavior selects appropriate chart types (such as line charts, pie charts, radar charts, etc.) based on the data and relationships in the analysis results, presenting the conclusions in a visual form to meet the conclusion visualization objective. Scene tag matching behavior analyzes and compares the characteristics of natural language processing technology and the characteristics of application scenarios to match the technology with corresponding scene tags, such as matching "sentiment analysis technology" with tags like "social media sentiment monitoring scenario" and "customer feedback analysis scenario," achieving the application scenario matching objective. Integrating the above terminal behaviors forms an optional set of behaviors for the results output stage.

[0052] Step S127: Determine the conditions for transitioning from the initial exploration phase to the deep analysis phase. The conditions for transitioning from the initial exploration phase to the deep analysis phase are that the science and technology intelligence agent completes a preset number of basic behaviors in the selectable behavior set of the initial exploration phase and the key information identification coverage reaches the threshold of the initial exploration phase.

[0053] In the context of analyzing the development trends of natural language processing technology, the transition from the initial exploration phase to the deep analysis phase requires certain conditions to ensure that the initial exploration phase has provided a sufficient and high-quality intelligence foundation for deep analysis. The preset quantity of basic behaviors refers to the total number of sample unit invocation, keyword matching, and information fragment labeling behaviors that the agent needs to complete in the initial exploration phase, reaching a pre-set standard. For example, this might involve completing the invocation, matching, and labeling of a certain number of academic paper abstract sample units. The key information identification coverage rate refers to the proportion of key information identified in the initial exploration phase relative to the total amount of key information that should be identified in that phase. When this proportion reaches a certain threshold in the initial exploration phase, it indicates that the agent has acquired sufficient key information. Only when both conditions are met simultaneously can the agent transition from the initial exploration phase to the deep analysis phase, ensuring the smooth progress of deep analysis and the reliability of the analysis results.

[0054] Step S128: Determine the conditions for transitioning from the deep analysis stage to the result output stage. The conditions for transitioning from the deep analysis stage to the result output stage are that the science and technology intelligence agent completes a preset number of advanced behaviors in the selectable set of behaviors in the deep analysis stage and the intelligence association mining completeness reaches the threshold of the deep analysis stage.

[0055] In the context of analyzing the development trends of natural language processing technology, the transition conditions from the deep analysis stage to the result output stage also need to be strictly set. The preset number of advanced behaviors refers to the total number of actions performed by the agent during the deep analysis stage, including sample correlation calculation, temporal feature extraction, and information credibility assessment, reaching a preset standard. Examples include calculating the correlation of a certain number of intelligence sample pairs and extracting temporal features from multiple time-series intelligence data. The intelligence correlation mining completeness refers to the proportion of intelligence correlations mined during the deep analysis stage to the total number of correlations that should be mined in that stage. When the threshold for the deep analysis stage is set such that this proportion reaches a certain level, it indicates that the agent has comprehensively mined the correlations between intelligences. Only when both conditions are met simultaneously can the agent move from the deep analysis stage to the result output stage, ensuring that the output results are based on sufficient deep analysis and have high quality and value.

[0056] Step S129: Structure and integrate the optional sets of behaviors in the initial exploration stage, the optional sets of behaviors in the deep analysis stage, the optional sets of behaviors in the result output stage, the conditions for transitioning from the initial exploration stage to the deep analysis stage, and the conditions for transitioning from the deep analysis stage to the result output stage, to form a reinforcement learning exploration space containing stage identifiers, behavior lists, and transition rules.

[0057] In the context of analyzing the development trends of natural language processing technology, the final step in forming a reinforcement learning exploration space is to structurally integrate the previously constructed sets of action options and stage transition conditions for each stage. Each stage is clearly labeled, such as "Initial Exploration Stage," "Deep Analysis Stage," and "Result Output Stage," so that the agent can clearly identify its current stage. The action options for each stage are organized into a action list, detailing the name, specific manifestation, and triggering conditions of each action. Simultaneously, the stage transition conditions are defined as transition rules, specifying under what circumstances the agent can transition from one stage to the next. Through this structural integration, a complete reinforcement learning exploration space is formed.

[0058] Step S130: Generate a phased intelligence sample cluster adapted to the reinforcement learning exploration space. The phased intelligence sample cluster includes an initial exploration phase intelligence sample subset corresponding to the initial exploration phase attribute, a deep analysis phase intelligence sample subset corresponding to the deep analysis phase attribute, and a result output phase intelligence sample subset corresponding to the result output phase attribute.

[0059] Step S131: Extract the initial exploration stage identifier and the corresponding initial exploration stage behavior selection set from the reinforcement learning exploration space, and determine the basic feature requirements of the intelligence sample for the initial exploration stage. The basic feature requirements of the intelligence sample for the initial exploration stage include the theme clarity feature requirements, keyword density feature requirements, and information fragment integrity feature requirements.

[0060] In the context of analyzing the development trends of natural language processing (NLP) technology, this study extracts the initial exploration stage identifiers and corresponding behavioral options from the reinforcement learning exploration space, clarifying the task focus and behavioral requirements of this stage. Based on this information, the basic feature requirements for intelligence samples in the initial exploration stage are determined. The theme clarity feature requires that the core theme of the intelligence sample must be clear and prominent, directly reflecting content related to NLP technology, avoiding samples with vague themes or irrelevant to the task. The keyword density feature requires that the frequency of keywords related to NLP technology in the sample reaches a certain standard to ensure that the sample contains sufficient key information, facilitating keyword matching and key information identification by the agent. The information fragment completeness feature requires that the basic information fragments in the sample (such as explanations of technical concepts, introductions to research background, etc.) must be complete, without missing important information.

[0061] Step S132: Based on the basic characteristic requirements of intelligence samples in the initial exploration stage, select intelligence samples that meet the requirements from the science and technology intelligence database to form an intelligence sample subset for the initial exploration stage. The intelligence sample subset for the initial exploration stage includes technical field overview samples, research direction introduction samples, and core terminology explanation samples.

[0062] Step S1321: Analyze the requirements for the topic clarity of intelligence samples in the initial exploration stage, and define the criteria for measuring topic clarity. The criteria for measuring topic clarity include the consistency between the title and core content of the intelligence sample reaching the preset topic clarity threshold, the frequency of occurrence of topic keywords in the abstract meeting the preset frequency requirements, and the degree of concentration of topic discussion in the main text meeting the preset concentration standard.

[0063] In the analysis of natural language processing (NLP) technology development trends, specific criteria need to be established when defining the thematic clarity requirements for intelligence samples in the initial exploration phase. The consistency between the sample title and its core content is a crucial criterion; the theme reflected in the sample title should be highly consistent with the core content of the sample text. For example, a sample titled "Research on Transformer Models in Natural Language Processing" must have core content focusing on research related to Transformer models in NLP. This consistency must meet a preset thematic clarity threshold. The frequency of keywords in the abstract must also meet preset frequency requirements. As a concise summary of the sample content, the abstract should contain a sufficient number of keywords related to NLP technology (such as "natural language processing," "Transformer model," and "pre-training") to indicate the sample's core theme. Finally, the concentration of thematic discussion in the main text must meet preset concentration standards. The main text should dedicate a significant portion to discussions of NLP technology, avoiding extensive discussions of irrelevant content to ensure thematic focus.

[0064] Step S1322: Based on the criteria for determining the clarity of the subject matter, evaluate each intelligence sample in the science and technology intelligence database, retain intelligence samples that meet the requirements for the clarity of the subject matter, and form a subject-selected subset.

[0065] In the context of analyzing the development trends of natural language processing technology, each intelligence sample in the science and technology intelligence database is carefully evaluated based on the aforementioned criteria for topic clarity. For each sample, the consistency between its title and core content is first checked to determine if it meets the preset topic clarity threshold; then, the frequency of keywords in the abstract is counted to see if it meets the preset frequency requirements; finally, the concentration of topic discussion in the main text is analyzed to confirm whether it meets the preset concentration standard. Only intelligence samples that simultaneously meet these three requirements are retained to form the topic selection subset.

[0066] Step S1323: Analyze the keyword density feature requirements of the intelligence sample in the initial exploration stage, define the calculation method of keyword density, which is the ratio of the total number of occurrences of the core keywords in the intelligence sample to the total number of characters in the sample, and require that this ratio reach the preset density threshold.

[0067] In the context of analyzing the development trends of natural language processing (NLP) technology, when interpreting the keyword density requirements for intelligence samples in the initial exploration phase, the specific calculation method for keyword density needs to be clarified. Core keywords refer to terms and technical names closely related to NLP technology, such as "word embedding," "recurrent neural network," and "attention mechanism." Keyword density is calculated by dividing the total number of occurrences of all core keywords in the intelligence sample by the total number of characters in the sample; the resulting ratio is the keyword density. This ratio must reach a preset density threshold to ensure that the sample contains a sufficient number of core keywords, thereby guaranteeing the relevance and information richness of the sample to the initial exploration phase task.

[0068] Step S1324: Calculate the intelligence samples in the topic selection subset according to the keyword density calculation method, retain the intelligence samples whose keyword density meets the requirements, and form the keyword density selection subset.

[0069] In the context of analyzing the development trends of natural language processing technology, keyword density is calculated for each intelligence sample in the topic selection subset according to a defined keyword density calculation method. The total frequency of core keywords in each sample is counted, and the ratio of this frequency to the total number of characters in the sample is calculated. The calculated keyword density is compared with a preset density threshold, and intelligence samples whose ratio reaches or exceeds the threshold are retained, forming a keyword density selection subset. This step further filters out samples containing sufficient core keywords, ensuring that the key information content of the samples meets the needs of the initial exploration phase.

[0070] Step S1325: Analyze the information fragment integrity feature requirements of the intelligence sample in the initial exploration stage, define the judgment criteria for information fragment integrity, including the integrity of the technical background introduction reaching the preset background integrity threshold, the integrity of the core concept definition meeting the preset concept integrity standard, and the integrity of the research scope definition meeting the preset scope definition requirements.

[0071] In the scenario of analyzing the development trend of natural language processing (NLP) technology, when analyzing the information fragment integrity requirements of intelligence samples in the initial exploration stage, judgment criteria for information fragment integrity are established. The integrity requirement for the technical background introduction is that the sample must comprehensively and detailedly introduce the research background, development history, and current status of NLP technology, reaching a preset background integrity threshold, enabling the agent to understand the origin and development of the technology. The integrity requirement for the core concept definition is that the sample's definition of the relevant core NLP concepts (such as "natural language understanding" and "natural language generation") must be accurate, clear, and complete, meeting the preset concept integrity standard, ensuring the agent can correctly understand the key concepts. The integrity requirement for the research scope definition is that the sample clearly states the specific scope of the NLP technology being studied, the content it covers, and the content it does not include, meeting the preset scope definition requirements and avoiding misunderstandings of the sample's research content by the agent.

[0072] Step S1326: Based on the judgment criteria for information fragment integrity, check the intelligence samples in the keyword density screening subset, retain the intelligence samples whose information fragment integrity meets the requirements, and form the information integrity screening subset.

[0073] In the context of analyzing the development trends of natural language processing technology, the intelligence samples in the keyword density filtering subset are examined one by one according to the criteria for judging the integrity of information fragments. The checks include: whether the technical background introduction of the sample is complete and meets the preset background integrity threshold; whether the core concept definition is complete and meets the preset concept integrity standard; and whether the research scope definition is complete and meets the preset scope definition requirements. Only samples that meet all the criteria for judging the integrity of information fragments are retained, forming the information integrity filtering subset.

[0074] Step S1327: Select intelligence samples from the information integrity screening subset that describe the overall situation of the technical field to form a technical field overview sample group. The technical field overview sample group includes field overview samples of different technical branches.

[0075] In the scenario of analyzing the development trend of natural language processing (NLP) technology, intelligence samples that describe the overall situation of the technology field are further filtered from the information integrity screening subset. These samples typically provide a comprehensive overview of the overall development status of NLP technology, the composition of major technical branches, the research focus of each branch, and their interrelationships. These samples are then grouped into a technology field overview sample group, which includes field overview samples for different technical branches, such as "Overview of the Overall Development of Natural Language Processing Technology," "Overview of Machine Translation Technology," and "Overview of Text Classification Technology," enabling the intelligent agent to understand the overall picture of the NLP technology field from a macro perspective.

[0076] Step S1328: Select intelligence samples with specific research directions as the description objects from the information integrity screening subset to form a research direction introduction sample group. The research direction introduction sample group includes the core content and development status samples of each research direction.

[0077] In the context of analyzing the development trends of natural language processing (NLP) technology, intelligence samples are selected from a subset of information integrity screening, focusing on specific research directions. These samples concentrate on a particular research direction within NLP technology, such as "Research on Sentiment Analysis Based on Deep Learning" or "Research on the Application of Pre-trained Language Models in Text Generation." The samples detail the core content, main methods used, current research progress, achievements, and challenges of each research direction. These samples are grouped into a research direction overview sample group, providing the agent with a deeper understanding of each specific research direction and supplementing the macro-level perspective of the technology overview sample group.

[0078] Step S1329: Select intelligence samples with technical terminology explanations as the description object from the information integrity screening subset to form a core terminology explanation sample group. The core terminology explanation sample group includes definition samples of basic terms, professional terms and emerging terms.

[0079] In the scenario of analyzing the development trend of natural language processing technology, intelligence samples focusing on the explanation of technical terms are selected from a subset of information integrity screening. These samples specifically explain and describe various technical terms in the field of natural language processing, including basic terms (such as "corpus" and "word vectors"), specialized terms (such as "bidirectional long short-term memory network" and "self-attention mechanism"), and emerging terms (such as "large language model" and "cue learning"). Each term definition sample includes the precise meaning of the term, relevant background, and application scenarios. These samples are combined into a core term explanation sample group to help the agent accurately understand and master the key terms in natural language processing technology.

[0080] Step S13210: Merge the sample groups of technical field overview, research direction introduction, and core terminology explanation, remove duplicate samples, and form an intelligence sample subset for the initial exploration stage.

[0081] In the context of analyzing the development trends of natural language processing technology, three sample groups—an overview of the technical field, a brief introduction to research directions, and an explanation of core terms—are merged. During the merging process, all samples undergo deduplication to remove those with identical or substantially repetitive content, thus avoiding redundancy. After merging and deduplication, a subset of intelligence samples for the initial exploration phase is formed. This subset integrates a macro-level overview of the technical field, detailed introductions to specific research directions, and accurate explanations of core terms, providing comprehensive, diverse, and non-duplicative intelligence sample support for the scientific and technological intelligence agent to perform basic actions such as sample unit invocation, keyword matching, and information fragment labeling during the initial exploration phase.

[0082] Step S13211: Add a stage identifier and topic tag to each sample in the intelligence sample subset of the initial exploration stage to facilitate the rapid identification and retrieval of the science and technology intelligence agent during the interaction process.

[0083] In the scenario of analyzing the development trend of natural language processing technology, each sample in the intelligence sample subset of the initial exploration stage is given a clear stage identifier, namely "initial exploration stage," enabling the agent to clearly identify which stage the intelligence sample belongs to. Simultaneously, based on the specific content of each sample, corresponding topic tags are added, such as "Overview of Natural Language Processing Technology," "Research Directions in Machine Translation," and "Explanation of Transformer Model Terminology." These topic tags accurately reflect the core topic content of the sample, facilitating the intelligent agent to quickly locate and retrieve the required intelligence samples based on its own behavioral needs (such as topic tag retrieval in sample unit invocation behavior) during interaction with the samples, thereby improving interaction efficiency.

[0084] Step S133: Extract the deep analysis stage identifiers and corresponding deep analysis stage behavior options in the reinforcement learning exploration space, and determine the advanced feature requirements of the deep analysis stage for intelligence samples. The advanced feature requirements of the deep analysis stage for intelligence samples include data correlation feature requirements, time series continuity feature requirements, and opinion conflict feature requirements.

[0085] In the context of analyzing the development trends of natural language processing technology, this paper extracts the identifiers of the deep analysis stage and the corresponding set of possible behaviors in the reinforcement learning exploration space, clarifying the advanced processing requirements of intelligence samples at this stage. Based on this, the advanced feature requirements of intelligence samples in the deep analysis stage are determined. Data correlation features require that the data contained in the intelligence sample can be correlated with the data of other samples, such as the comparative relationship between different experimental data, the correlation between different technical indicators, etc., so that the agent can calculate the sample correlation degree. Time series continuity features require that the data or information in the sample can form a continuous sequence in chronological order, such as the change data of a certain natural language processing technology performance indicator over time, the distribution data of research hotspots in different years, etc., supporting the agent in extracting time-series features. Conflicting viewpoint features require that the sample contains conflicting information such as different academic viewpoints, research conclusions, or technical evaluations, providing material for the agent to distinguish contradictory information.

[0086] Step S134: Based on the advanced feature requirements of intelligence samples in the in-depth analysis stage, select intelligence samples that meet the requirements from the science and technology intelligence database to form a subset of intelligence samples for the in-depth analysis stage. The subset of intelligence samples for the in-depth analysis stage includes experimental data comparison samples, technology evolution process samples, and academic opinion debate samples.

[0087] In the scenario of analyzing the development trend of natural language processing technology, suitable intelligence samples are selected from the science and technology intelligence database based on the advanced feature requirements of the deep analysis stage. For the data correlation feature requirement, samples containing experimental data that can be compared and analyzed are selected; for the time series continuity feature requirement, samples with time series data or records of the technological development process in chronological order are selected; for the viewpoint conflict feature requirement, samples containing different academic viewpoints or research conclusions are selected. These selected samples are categorized into experimental data comparison samples, technological evolution process samples, and academic viewpoint debate samples, collectively forming the intelligence sample subset for the deep analysis stage. This subset meets the needs of the intelligent agent to perform advanced behaviors such as sample correlation calculation, time series feature extraction, and information credibility assessment at this stage.

[0088] Step S135: Extract the result output stage identifier and the corresponding result output stage behavior selection set in the reinforcement learning exploration space, and determine the final feature requirements of the result output stage for the intelligence sample. The final feature requirements of the result output stage for the intelligence sample include the conclusion conciseness feature requirements, the application scenario adaptability feature requirements, and the visualization compatibility feature requirements.

[0089] In the scenario of analyzing the development trend of natural language processing technology, the identifiers of the output stage and the corresponding set of possible behaviors in the reinforcement learning exploration space are extracted to clarify the final presentation requirements of the intelligence sample at this stage. Based on these, the final feature requirements of the intelligence sample at the output stage are determined. The conclusion conciseness feature requires that the conclusion part of the sample must be concise, accurate, and highly generalized, clearly expressing the core viewpoints and key findings, facilitating content condensation by the agent. The application scenario adaptability feature requires that the technology or conclusion in the sample be relevant to actual application scenarios, including application cases and applicable conditions in different scenarios, supporting scenario label matching by the agent. The visualization compatibility feature requires that the data or information in the sample be suitable for display in visualization forms such as charts, such as data having clear classifications and numerical relationships, facilitating chart generation by the agent.

[0090] Step S136: Based on the final feature requirements of the intelligence samples in the result output stage, select the intelligence samples that meet the requirements from the science and technology intelligence database to form a subset of intelligence samples for the result output stage. The subset of intelligence samples for the result output stage includes research conclusion summary samples, technology application case samples, and data visualization template samples.

[0091] In the scenario of analyzing the development trend of natural language processing technology, intelligence samples are selected from the overall science and technology intelligence database based on the terminal characteristic requirements of the output stage. The selection includes: research conclusion summary samples with concise conclusion characteristics (containing highly summarized conclusions related to natural language processing technology research); technology application case samples with application scenario adaptability characteristics (detailing the application of the technology in real-world scenarios); and data visualization template samples with visualization compatibility characteristics (providing chart templates and examples suitable for displaying data related to natural language processing technology). These samples form the intelligence sample subset for the output stage, providing strong sample support for the agent to perform terminal actions such as content refinement, chart generation, and scene label matching at this stage.

[0092] Step S137: Analyze the content association between the intelligence sample subset in the initial exploration phase and the intelligence sample subset in the in-depth analysis phase, and calculate the thematic overlap between the intelligence sample subset in the initial exploration phase and the intelligence sample subset in the in-depth analysis phase.

[0093] In the context of analyzing the development trends of natural language processing (NLP) technologies, analyzing the content correlation between the intelligence sample subsets in the initial exploration phase and the deep analysis phase is crucial to ensuring that the intelligence samples from the two phases can be interconnected and supportive of each other. By comparing the thematic content of the two subsets, common NLP themes are identified, such as "pre-trained language models" and "sentiment analysis." Then, the proportion of these common themes in the two subsets is calculated, i.e., the theme overlap. The theme overlap can be calculated by statistically analyzing the proportion of samples with common themes to the total number of samples in both subsets.

[0094] Step S138: Analyze the content relationship between the intelligence sample subset in the in-depth analysis stage and the intelligence sample subset in the result output stage, and calculate the logical coherence between the intelligence sample subset in the in-depth analysis stage and the intelligence sample subset in the result output stage.

[0095] In the context of analyzing the development trends of natural language processing technology, this study focuses on the content correlation between the intelligence sample subsets in the in-depth analysis phase and the intelligence sample subsets in the output phase, with a particular emphasis on their logical coherence. Logical coherence refers to the ability of the intelligence analysis results in the in-depth analysis phase to reasonably derive and support the conclusions, recommendations, or visualizations in the output phase. This is determined by examining the logical relationships between the analysis process, data support, and conclusions in the in-depth analysis phase samples and the final conclusions, application recommendations, and visualizations in the output phase samples, assessing their consistency and reasonableness. Logical coherence can be calculated by evaluating the degree to which the intelligence in the in-depth analysis phase supports the intelligence in the output phase, such as the proportion of supported conclusions to the total number of conclusions in the output phase, ensuring that the results of the in-depth analysis are effectively transformed into output results.

[0096] Step S139: Based on topic overlap and logical coherence, adjust the sample quantity ratio of the intelligence sample subset in the initial exploration stage, the intelligence sample subset in the in-depth analysis stage, and the intelligence sample subset in the result output stage, so that the correlation strength of the sample subsets in adjacent stages adapts to the stage transition requirements of the reinforcement learning exploration space.

[0097] In the scenario of analyzing the development trend of natural language processing technology, the sample size ratio of the intelligence sample subsets in the three stages is adjusted based on the calculated topic overlap and logical coherence. If the topic overlap between the initial exploration stage and the deep analysis stage is low, it indicates that the samples in the initial exploration stage do not provide a sufficiently relevant foundation for the deep analysis stage. It may be necessary to increase the number of samples related to the deep analysis topic in the initial exploration stage, or decrease the number of samples with low relevance to the initial exploration topic in the deep analysis stage. If the logical coherence between the deep analysis stage and the result output stage is insufficient, it may be necessary to adjust the analysis depth of the samples in the deep analysis stage or the conclusion basis of the samples in the result output stage, and adjust the sample size ratio of the two subsets accordingly. Through the above adjustments, the correlation strength of the sample subsets in adjacent stages can meet the requirements of stage transition in the reinforcement learning exploration space. That is, after the initial exploration stage is completed, the deep analysis stage has enough relevant samples for analysis; after the deep analysis stage is completed, the result output stage has sufficient and logically coherent intelligence to support the output results.

[0098] Step S1310: The adjusted initial exploration phase intelligence sample subset, in-depth analysis phase intelligence sample subset, and result output phase intelligence sample subset are ordered to form a phased intelligence sample cluster containing phase labels and sample association maps.

[0099] In the context of analyzing the development trends of natural language processing technology, the adjusted three-stage intelligence sample subsets are sequentially combined according to the initial exploration stage, the in-depth analysis stage, and the result output stage. Each sample subset is labeled with a specific stage, such as "Initial Exploration Stage Sample Subset," "In-Depth Analysis Stage Sample Subset," and "Result Output Stage Sample Subset." Simultaneously, a sample association graph is constructed, graphically displaying the relationships between sample subsets at different stages, as well as the relationships between samples within the same subset. For example, a sample outlining a specific technological field in the initial exploration stage may be associated with multiple related technological evolution samples in the in-depth analysis stage, and these technological evolution samples may in turn be associated with corresponding technological application case samples in the result output stage. Through this sequential combination and the construction of the sample association graph, a complete cluster of staged intelligence samples is formed.

[0100] Step S140: Invoke the technology intelligence agent to interact with the phased intelligence sample cluster in the reinforcement learning exploration space in stages to form a cross-stage agent behavior sequence.

[0101] Step S141: Configure the initial state of the science and technology intelligence agent to correspond to the initial exploration stage of the reinforcement learning exploration space, and load the initial exploration stage intelligence sample subset from the staged intelligence sample cluster.

[0102] In the context of analyzing the development trends of natural language processing technology, the first step in invoking a scientific and technological intelligence agent is to configure its initial state. The agent's initial state is set to correspond to the initial exploration phase in the reinforcement learning exploration space, clarifying its current stage and task objectives. Simultaneously, a subset of intelligence samples for the initial exploration phase is loaded from the phased intelligence sample cluster. All sample units within this subset (such as technical field overview samples, research direction introduction samples, and core terminology explanation samples) are provided to the agent as the objects of its basic operations during the initial exploration phase, including sample unit invocation, keyword matching, and information fragment labeling. This ensures the agent can successfully initiate the initial exploration phase task.

[0103] Step S142: Call the science and technology intelligence agent to select basic behaviors from the initial exploration phase behavior selection set, execute the selected basic behaviors on the initial exploration phase intelligence sample subset, and record the basic behavior execution process and the corresponding sample response information.

[0104] In the scenario of analyzing the development trend of natural language processing technology, after completing the initial state configuration and sample loading, the technology intelligence agent selects basic behaviors from the initial exploration phase's set of available behaviors. Based on the current task requirements and sample conditions, the agent autonomously selects appropriate basic behaviors. For example, it first executes the sample unit invocation behavior to retrieve a sample outlining a specific technical field, then performs keyword matching on that sample to find keywords related to "the development history of natural language processing technology," and finally performs information fragment labeling on irrelevant advertising information in the sample. During the execution of each basic behavior, detailed records are kept of the execution steps, the sample units involved, the behavior parameters, and the sample's response information under the influence of the behavior, such as keyword matching results and labeled invalid information fragments.

[0105] Step S143: Monitor the basic behavior execution progress of the science and technology intelligence agent. When the conditions for transitioning from the initial exploration phase to the deep analysis phase are met, switch the state of the science and technology intelligence agent to the deep analysis phase and load the deep analysis phase intelligence sample subset from the phased intelligence sample cluster.

[0106] In the context of analyzing the development trends of natural language processing technology, the execution progress of the basic behaviors of the science and technology intelligence agent in the initial exploration phase is monitored in real time. Specifically, it monitors whether the number of basic behaviors completed by the agent reaches a preset number, and whether the key information identification coverage reaches the threshold of the initial exploration phase. For example, whether the agent has completed the calling, matching, and labeling of a preset number of sample units, and whether the proportion of key information identified in these samples to the total amount of key information to be identified reaches a threshold. When both conditions are met, it indicates that the initial exploration phase has been completed, and the state of the science and technology intelligence agent is switched from the initial exploration phase to the deep analysis phase. At the same time, a subset of intelligence samples for the deep analysis phase is loaded from the phased intelligence sample cluster.

[0107] Step S144: Call the science and technology intelligence agent to select advanced behaviors from the set of behaviors available in the deep analysis stage, execute the selected advanced behaviors on the subset of intelligence samples in the deep analysis stage, record the execution process of the advanced behaviors and the corresponding sample response information, and associate them with the basic behavior records of the initial exploration stage.

[0108] Step S1441: Extract the advanced behavior list of the optional set of behaviors in the deep analysis stage in the reinforcement learning exploration space, and show the behavior descriptions and execution objects of the sample correlation calculation behavior, temporal feature extraction behavior, and information credibility assessment behavior to the science and technology intelligence agent.

[0109] In the scenario of analyzing the development trend of natural language processing technology, when the science and technology intelligence agent enters the deep analysis stage, the system first extracts a list of advanced behaviors from the optional set of behaviors in the deep analysis stage within the reinforcement learning exploration space. This list details three advanced behaviors: sample correlation calculation, temporal feature extraction, and information credibility assessment. Then, the system displays the behavior description and execution target for each behavior to the agent. For example, the description of sample correlation calculation is "to calculate the degree of correlation between samples by analyzing common technical terms, research methods, and experimental conclusions," and its execution target is the experimental data comparison samples in the intelligence sample subset of the deep analysis stage; the description of temporal feature extraction is "to extract the trend characteristics, periodic characteristics, and mutation characteristics of data or information in samples in chronological order," and its execution target is samples of technological evolution; the description of information credibility assessment is "to analyze the authority of the sample source, the strength of data support, and consistency with other samples to assess the credibility level of the information," and its execution target is samples of academic debate. Through this display, the agent clearly understands the content and applicable targets of each advanced behavior, enabling it to make appropriate choices.

[0110] Step S1442: Based on the content characteristics of the currently loaded deep analysis stage intelligence sample subset, the science and technology intelligence agent selects the first advanced behavior that matches the sample content characteristics from the advanced behavior list. If the sample content contains multiple related data, the sample correlation calculation behavior is selected first.

[0111] In the scenario of analyzing the development trend of natural language processing technology, after receiving a list of advanced actions along with their descriptions and execution objects, the science and technology intelligence agent analyzes the content characteristics of the currently loaded subset of intelligence samples for the deep analysis phase. For example, if the currently loaded sample subset contains a large number of experimental data comparison samples, and these samples involve multiple related experimental data (such as performance test data of the same natural language processing model on different datasets), that is, the sample content contains multiple related data, then according to the set priority rules, the agent prioritizes the sample correlation calculation action as the first advanced action, because this action can effectively mine the inherent connections between these related data, which is in line with the core objective of intelligence correlation mining in the deep analysis phase.

[0112] Step S1443: Call the science and technology intelligence agent to perform pairwise comparisons on multiple samples in the intelligence sample subset of the in-depth analysis stage, extract the common technical terms, research methods and experimental conclusions among the samples, calculate the proportion of these common elements as the correlation parameter, and require the correlation parameter to reach the preset correlation threshold.

[0113] In the scenario of analyzing the development trend of natural language processing technology, a science and technology intelligence agent is invoked to perform sample correlation calculation. The agent performs pairwise comparisons on multiple samples from a subset of intelligence samples in the deep analysis phase, for example, selecting two experimental data comparison samples on different natural language processing models in text classification tasks. During the pairwise comparison process, the agent extracts common technical terms (such as "convolutional neural network," "accuracy," "recall," etc.), research methods (such as "supervised learning," "cross-validation," etc.), and experimental conclusions (such as "model A outperforms model B on a small dataset," etc.) between the two samples. Then, the number of these common elements is counted, and their proportion to the total number of elements in both samples is calculated. This proportion is used as a correlation parameter. This correlation parameter is required to reach a preset correlation threshold to determine whether the two samples have a strong correlation. If the threshold is reached, the two samples are considered closely related, and further correlation analysis can be performed.

[0114] Step S1444: Record the execution process of the sample correlation calculation behavior. The execution process of the sample correlation calculation behavior includes the compared sample pairs, the extracted common elements and the calculated correlation parameters. At the same time, record the response changes of the samples in the intelligence sample subset in the deep analysis stage. The response changes of the samples in the intelligence sample subset in the deep analysis stage are the states of the samples after the correlation parameters are marked.

[0115] In the scenario of analyzing the development trend of natural language processing technology, the execution process of the agent performing the sample correlation calculation is recorded in detail. Specifically, the identification information of the compared sample pairs (such as sample numbers), the list of common elements extracted from the sample pairs (including common technical terms, research methods, and experimental conclusions), and the calculated correlation parameter values ​​are recorded. Simultaneously, close attention is paid to the response changes of samples in the intelligence sample subset during the deep analysis phase under the action of this behavior, i.e., the state of the samples after being labeled with correlation parameters. For example, correlation parameter labels are added to the samples participating in the comparison, such as "the correlation parameter with sample X is Y," and these labels are stored in the sample's attribute information so that subsequent actions (such as temporal feature extraction) can refer to this correlation information, or reflect the changes in the sample's correlation state when forming the agent's action sequence.

[0116] Step S1445: Call the science and technology intelligence agent to process the samples with time series attributes in the intelligence sample subset of the in-depth analysis stage, arrange the data points in the samples in chronological order, and extract the trend characteristics, periodic characteristics and mutation characteristics of the data points over time.

[0117] In the scenario of analyzing the development trend of natural language processing technology, a science and technology intelligence agent is invoked to perform time-series feature extraction. The agent first filters samples with time-series attributes from the intelligence sample subset of the deep analysis phase, such as samples of technological evolution. These samples contain development data of natural language processing technology at different time points (such as the number of core papers published each year, the annual average of major technical indicators, etc.). The agent arranges the data points in the samples in chronological order (such as year order from earliest to latest). Then, it analyzes the arranged data points to extract their trend characteristics over time (such as the overall upward or downward trend of technical indicators), periodic characteristics (such as regular fluctuation cycles in technological development), and abrupt change characteristics (such as a sudden and significant change in technical indicators within a certain period, which may correspond to a key technological breakthrough).

[0118] Step S1446: Record the execution process of time series feature extraction. The execution process of time series feature extraction includes the division method of time series, the extracted trend feature parameters, periodic feature parameters and mutation feature parameters. At the same time, record the response changes of samples in the intelligence sample subset in the deep analysis stage. The response changes of samples in the intelligence sample subset in the deep analysis stage are the state of the samples after they are labeled with time series features.

[0119] In the scenario of analyzing the development trend of natural language processing technology, when recording the execution process of time-series feature extraction, the method of dividing the time series is recorded in detail, such as dividing the data points by year, quarter, or month. Simultaneously, the extracted trend feature parameters (such as the slope and intercept of the trend line, etc.), periodic feature parameters (such as the period length and amplitude, etc.), and abrupt change feature parameters (such as the time point of the abrupt change and the amount of change in data before and after the abrupt change, etc.) are recorded. Furthermore, the response changes of samples in the intelligence sample subset during the deep analysis phase under this action are recorded, i.e., the state of the samples after being labeled with time-series features. For example, time-series feature labels are added to the samples, such as "Trend feature: rising; Periodic feature: no obvious period; Abrupt change feature: significant increase in 2020," and these labels are associated with the samples for storage, so that the time-series feature information of the samples is clearly reflected.

[0120] Step S1447: Call the science and technology intelligence agent to process the samples containing viewpoints or conclusions in the intelligence sample subset of the in-depth analysis stage, analyze the source authority of the samples, the strength of data support and the degree of consistency with other samples, and generate a credibility level description. The credibility level description is divided into multiple levels according to preset standards.

[0121] In the context of analyzing the development trends of natural language processing technology, a science and technology intelligence agent is invoked to perform information credibility assessment. The agent processes samples containing viewpoints or conclusions (primarily samples of academic debates) within the intelligence sample subset from the deep analysis phase. First, it analyzes the source authority of the sample, such as determining whether the journal in which the sample was published is an authoritative academic journal and whether the authors are well-known researchers in the field. Then, it assesses the data support strength for the viewpoints or conclusions in the sample, checking whether there is sufficient experimental data, case studies, etc., to support the viewpoints or conclusions. Next, it analyzes the degree of consistency between the viewpoints or conclusions of the sample and other relevant samples from the deep analysis phase, determining whether there are any contradictions or conflicts. Based on the combined analysis results of source authority, data support strength, and consistency with other samples, a credibility level description for the sample is generated according to preset standards (such as excellent, good, average, poor, etc.).

[0122] Step S1448: Record the execution process of the information credibility assessment behavior. The execution process of the information credibility assessment behavior includes the source analysis basis, the data support strength judgment standard and the consistency comparison result. At the same time, record the response changes of the samples in the intelligence sample subset in the deep analysis stage. The response changes of the samples in the intelligence sample subset in the deep analysis stage are the state after the samples are marked with credibility level.

[0123] In the scenario of analyzing the development trend of natural language processing technology, when recording the execution process of information credibility assessment, detailed records should be kept of the source analysis basis (such as the journal's impact factor, the author's H-index, etc.), the criteria for judging the strength of data support (such as the size of the data, the rationality of the experimental design, etc.), and the consistency comparison results with other samples (such as how many samples are consistent with each other, how many samples conflict with each other, etc.). Simultaneously, the response changes of samples in the intelligence sample subset during the in-depth analysis stage under the action of this behavior should be recorded, that is, the state of the sample after being labeled with a credibility level. For example, a credibility level label, such as "credibility level: good," can be added to the sample, and the source analysis basis and data support strength information collected during the assessment process can be stored as supplementary explanations for the label in the sample's attributes, so that subsequent behavior sequences can reflect the results of the sample credibility assessment.

[0124] Step S1449: Retrieve the basic behavior records of the initial exploration phase, identify the basic behaviors related to the samples in the current deep analysis phase during the initial exploration phase, including sample unit calling behavior or keyword matching behavior for the same topic, and establish the association identifier between the basic behaviors related to the samples in the initial exploration phase and the advanced behaviors in the deep analysis phase.

[0125] In the context of analyzing the development trends of natural language processing technology, to link the advanced behaviors of the deep analysis phase with the basic behaviors of the initial exploration phase, all basic behavior records from the initial exploration phase are retrieved. These records are analyzed to identify basic behaviors related to the samples being processed in the current deep analysis phase. For example, if the current deep analysis phase is processing experimental data comparison samples related to "the application of Transformer models in machine translation," by reviewing the basic behavior records from the initial exploration phase, sample unit invocation behaviors and keyword matching behaviors previously performed on the same topic, "Introduction to Research Directions of Transformer Models," can be identified. Then, association markers are established between these related basic behaviors and the current advanced deep analysis behaviors (such as sample correlation calculation behaviors). For instance, an identifier such as "Associated with Initial Exploration Behavior ID: XXX" is added to the advanced behavior records, and an identifier such as "Associated with Deep Analysis Behavior ID: YYY" is added to the basic behavior records, thus forming a cross-phase behavioral association chain in the behavior sequence.

[0126] Step S14410: Integrate the advanced behavior execution process, sample response information, and association identifiers with the basic behavior records of the initial exploration phase in the deep analysis phase to form a complete interaction record of the deep analysis phase.

[0127] In the scenario of analyzing the development trend of natural language processing technology, the execution process of each advanced behavior in the deep analysis stage (such as sample pairs, common elements, and correlation parameters for calculating sample correlation; time series division method and various feature parameters for time series feature extraction, etc.), sample response information (such as the status of the labeled correlation parameters, time series feature status, and credibility level status of the sample, etc.), and the association identifiers with the basic behavior records of the initial exploration stage (such as the associated behavior ID) are systematically integrated. The above information is organized according to the chronological order of behavior execution to form a complete interaction record for the deep analysis stage. This record comprehensively reflects the interaction process between the agent and the intelligence samples in the deep analysis stage, the correlation between behaviors, and the changes in sample status, and is an important component for subsequently forming cross-stage agent behavior sequences.

[0128] Step S145: Monitor the progress of the advanced behavior of the science and technology intelligence agent. When the conditions for transitioning from the deep analysis stage to the result output stage are met, switch the state of the science and technology intelligence agent to the result output stage and load the result output stage intelligence sample subset from the staged intelligence sample cluster.

[0129] In the context of analyzing the development trends of natural language processing technology, the progress of advanced behaviors of the science and technology intelligence agent in the deep analysis phase is continuously monitored. This includes monitoring whether the number of advanced behaviors completed by the agent reaches a preset limit, and whether the completeness of intelligence association mining reaches the threshold for the deep analysis phase. For example, whether the agent has completed the association calculation, temporal feature extraction, and information credibility assessment for a preset number of sample pairs, and whether the proportion of the number of mined intelligence associations to the total number of associations to be mined in this phase reaches a threshold. When both conditions are met, it indicates that the deep analysis phase is complete, and the agent's state is switched to the result output phase. Simultaneously, a subset of intelligence samples from the phased intelligence sample cluster is loaded from the result output phase.

[0130] Step S146: Call the science and technology intelligence agent to select the terminal behavior from the set of optional behaviors in the result output stage, execute the selected terminal behavior on the subset of intelligence samples in the result output stage, record the execution process of the terminal behavior and the corresponding sample response information, and at the same time associate it with the advanced behavior records in the deep analysis stage.

[0131] In the scenario of analyzing the development trends of natural language processing technology, after the agent enters the result output stage and loads a subset of samples, it selects a terminal action from the set of available actions for the result output stage. Based on the core objectives and sample characteristics of the result output stage, the agent selects an appropriate terminal action. For example, for research conclusion summary samples, it performs content condensation to generate a concise summary of technology development trends; for technology application case samples, it performs scenario tag matching to match appropriate application scenario tags; and for data visualization template samples, it performs chart generation to transform the analyzed data into visual charts. During the execution of the terminal action, the execution details, processed samples, and sample response information (such as the generated summary text, matched scenario tags, and generated chart types) are recorded. The terminal action records are also associated with advanced action records related to the deep analysis stage. For example, the content condensation action is associated with previous actions such as temporal feature extraction and information credibility assessment of relevant samples to ensure that the output results are supported by in-depth analysis.

[0132] Step S147: Collect basic behavior records from the initial exploration phase, advanced behavior records from the deep analysis phase, and final behavior records from the result output phase. Together with the sample response information corresponding to each phase, organize them in chronological order to form an agent behavior sequence containing phase identifiers, behavior types, and response associations.

[0133] In the scenario of analyzing the development trend of natural language processing technology, all basic behavioral records of the science and technology intelligence agent in the initial exploration stage (including the execution process, sample response information, etc.), all advanced behavioral records in the deep analysis stage (including the execution process, sample response information, and association identifiers with the initial exploration stage behaviors, etc.), and all terminal behavioral records in the result output stage (including the execution process, sample response information, and association identifiers with the deep analysis stage behaviors, etc.) are collected. Then, these records, along with the corresponding sample response information for each stage, are uniformly organized according to the chronological order of behavior execution. During the organization process, stage identifiers (such as "initial exploration," "deep analysis," "result output") are added to each behavioral record, and the behavior type is clearly defined (such as "sample unit call," "sample correlation calculation," "content refinement," etc.). Association identifiers are used to reflect the response correlations between behaviors in different stages and between behaviors and sample responses. Finally, a complete agent behavior sequence is formed, which comprehensively and systematically records all behaviors and interactions of the agent throughout the entire intelligence task.

[0134] Step S148: Mark the connection points between the initial exploration stage behavior and the deep analysis stage behavior, and between the deep analysis stage behavior and the result output stage behavior in the agent behavior sequence, clarify the causal relationship between different stage behaviors, so that the agent behavior sequence fully reflects the cross-stage interaction process of the science and technology intelligence agent in the reinforcement learning exploration space.

[0135] In the scenario of analyzing the development trend of natural language processing technology, the connection points between different stages of behavior are marked in the formed agent behavior sequence. Specifically, the connection points between the last basic behavior of the initial exploration stage and the first advanced behavior of the deep analysis stage, and between the last advanced behavior of the deep analysis stage and the first terminal behavior of the result output stage, are marked. These connection points are marked by adding explicit labels such as "Stage transition: Initial exploration → Deep analysis" and "Stage transition: Deep analysis → Result output" to the behavior sequence. At the same time, based on the previously established behavior association labels, the causal relationships between different stages of behavior are clarified in the sequence. For example, it is explained that a certain advanced behavior in the deep analysis stage is based on the results of which basic behaviors in the initial exploration stage, and a certain terminal behavior in the result output stage is based on the analysis results of which advanced behaviors in the deep analysis stage. Through these labels and association descriptions, the agent behavior sequence can completely and clearly reflect the cross-stage interaction process of the scientific and technological intelligence agent in the reinforcement learning exploration space from initial exploration to deep analysis to result output, as well as the inherent logical connections between the behaviors of each stage.

[0136] Step S150: Construct a multi-stage reward function based on the multi-stage evolution characteristics of the intelligence task, generate stage reward signals according to the agent's behavior sequence, update the decision parameters of the science and technology intelligence agent using the stage reward signals, and complete a single training iteration.

[0137] Step S151: For the initial exploration stage attribute in the multi-stage evolution characteristics of intelligence missions, construct an initial exploration stage reward sub-function. The input of the initial exploration stage reward sub-function is the basic behavior of the science and technology intelligence agent in the initial exploration stage and the corresponding sample response information. The output of the initial exploration stage reward sub-function is the reward value corresponding to the effect of the basic behavior. The reward value corresponding to the effect of the basic behavior is positively correlated with the increase of the key information recognition coverage.

[0138] In the context of analyzing the development trends of natural language processing technology, a reward sub-function for the initial exploration phase is constructed based on the attributes of the initial exploration phase. The purpose of this sub-function is to evaluate the performance of the agent's basic behaviors during the initial exploration phase. Its input parameters include the various basic behaviors performed by the agent in the initial exploration phase (such as sample unit invocation, keyword matching, and information fragment labeling) and the corresponding sample response information (such as keyword matching results and labeled invalid information fragments). The output of the sub-function is the reward value corresponding to the performance of the basic behaviors. Among these, the key information recognition coverage rate is an important indicator for measuring the performance of the basic behaviors; that is, the proportion of key information recognized by the agent in the initial exploration phase out of the total amount of key information that should be recognized in this phase. The reward value is positively correlated with the key information recognition coverage rate; that is, the higher the key information recognition coverage rate, the larger the reward value, and vice versa. Through this design, the agent is incentivized to recognize as much key information as possible during the initial exploration phase.

[0139] Step S152: For the deep analysis stage attribute in the multi-stage evolution characteristics of intelligence tasks, construct a deep analysis stage reward sub-function. The input of the deep analysis stage reward sub-function is the advanced behavior of the science and technology intelligence agent in the deep analysis stage and the corresponding sample response information. The output of the deep analysis stage reward sub-function is the reward value corresponding to the effect of the advanced behavior. The reward value corresponding to the effect of the advanced behavior is positively correlated with the improvement of the completeness of intelligence association mining.

[0140] In the scenario of analyzing the development trend of natural language processing technology, a reward sub-function for the deep analysis stage is constructed based on the attributes of the deep analysis stage. This sub-function is used to evaluate the performance of the agent's advanced behaviors in the deep analysis stage. Its inputs are the advanced behaviors performed by the agent (such as sample correlation calculation, temporal feature extraction, and information credibility assessment) and the corresponding sample response information (such as correlation parameters, temporal feature parameters, credibility level, etc.). The output is the reward value corresponding to the effect of the advanced behavior. The completeness of intelligence correlation mining is a key indicator for measuring the effect of advanced behaviors, referring to the proportion of intelligence correlations mined by the agent in the deep analysis stage to the total number of correlations that should be mined in this stage. The reward value is positively correlated with the completeness of intelligence correlation mining, that is, the more complete the intelligence correlations mined, the higher the reward value, thereby incentivizing the agent to delve deeper into the correlations between intelligences.

[0141] Step S153: For the result output stage attribute in the multi-stage evolution characteristics of intelligence tasks, construct a result output stage reward sub-function. The input of the result output stage reward sub-function is the terminal behavior of the science and technology intelligence agent in the result output stage and the corresponding sample response information. The output of the result output stage reward sub-function is the reward value corresponding to the effect of the terminal behavior. The reward value corresponding to the effect of the terminal behavior is positively correlated with the improvement of the matching degree between the conclusion and the application scenario.

[0142] In the context of analyzing the development trends of natural language processing technology, a reward sub-function for the result output stage is constructed based on the attributes of the result output stage. This sub-function is used to evaluate the performance of the agent's terminal behavior in the result output stage. The input consists of the terminal behavior performed by the agent (such as content condensation, chart generation, and scene tag matching) and the corresponding sample response information (such as generated summary text, chart data, and matched scene tags). The output is the reward value corresponding to the effect of the terminal behavior. The matching degree between the conclusion and the application scenario is the core indicator for measuring the effect of the terminal behavior, that is, the degree to which the technical conclusions and suggestions output by the agent fit the actual application scenario. The reward value is positively correlated with this matching degree; the higher the matching degree, the larger the reward value, thus incentivizing the agent to generate output results that are highly adapted to the application scenario.

[0143] Step S154: Set the inter-stage reward correlation coefficient. The inter-stage reward correlation coefficient is used to quantify the degree of influence of the reward value of the previous stage on the reward value of the next stage, so that the reward value of the subsequent stage can reflect the cumulative effect of the previous behavior.

[0144] In the context of analyzing the development trends of natural language processing technology, setting an inter-stage reward correlation coefficient is crucial to account for the mutual influence between different stages. Since the stages of an intelligence task are interconnected and progressively advance, the effectiveness of actions in previous stages will affect the execution of actions in subsequent stages. For example, a high coverage rate of key information identification in the initial exploration stage provides a better intelligence foundation for the deep analysis stage, potentially making it easier to achieve a higher level of intelligence correlation and mining completeness in the deep analysis stage. The inter-stage reward correlation coefficient is a parameter used to quantify the degree of this influence; its range and specific value are set according to the closeness of the correlation between different stages. By introducing this coefficient, the reward value of the subsequent stage depends not only on the effectiveness of actions in this stage but also on the reward value of the previous stage, thus allowing the reward value to reflect the cumulative effect of previous actions and more comprehensively evaluate the overall performance of the agent.

[0145] Step S155: Integrate the reward sub-function of the initial exploration stage, the reward sub-function of the deep analysis stage, the reward sub-function of the result output stage, and the reward correlation coefficient between stages to form a multi-stage reward function.

[0146] In the context of analyzing the development trends of natural language processing technology, the constructed reward sub-functions for the initial exploration stage, deep analysis stage, and result output stage, along with the established inter-stage reward correlation coefficients, are integrated to form a complete multi-stage reward function. During this integration, it is necessary to clarify the connection methods between the sub-functions and the functioning of the inter-stage reward correlation coefficients. For example, the overall reward value of the multi-stage reward function can be expressed as the weighted sum or other combinations of the output value of the initial exploration stage reward sub-function, the output value of the deep analysis stage reward sub-function adjusted by the inter-stage reward correlation coefficients, and the output value of the result output stage reward sub-function adjusted again by the inter-stage reward correlation coefficients. Through this integration, the multi-stage reward function can comprehensively evaluate the agent's performance throughout the intelligence task, considering both the independent effects of each stage and the cumulative impact between stages.

[0147] Step S156: Divide the agent's behavior sequence into the initial exploration stage, the deep analysis stage, and the result output stage. Extract the behavior and sample response information of each stage and input them into the corresponding stage reward sub-functions to obtain the initial reward value of the initial exploration stage, the initial reward value of the deep analysis stage, and the initial reward value of the result output stage.

[0148] In the scenario of analyzing the development trend of natural language processing technology, the complete agent behavior sequence is first broken down into three stages: initial exploration, deep analysis, and result output, resulting in independent behavior record segments for each stage. Then, the actions performed by the agent (basic, advanced, and terminal behaviors) and corresponding sample response information are extracted from each stage's behavior record segment. The behavior and sample response information of the initial exploration stage are input into the initial exploration stage reward function to calculate the initial reward value; the behavior and sample response information of the deep analysis stage are input into the deep analysis stage reward function to obtain the initial reward value; and the behavior and sample response information of the result output stage are input into the result output stage reward function to obtain the initial reward value. These initial reward values ​​only reflect a preliminary assessment of the effectiveness of each stage's behavior and do not yet consider the inter-stage correlation.

[0149] Step S157: Based on the inter-stage reward correlation coefficient, adjust the initial reward value of the initial exploration stage, the initial reward value of the in-depth analysis stage, and the initial reward value of the result output stage to obtain a stage reward signal that includes the inter-stage cumulative effect.

[0150] In the context of analyzing the development trends of natural language processing technology, the initial reward values ​​for each stage are adjusted based on a predefined inter-stage reward correlation coefficient. For example, assuming the correlation coefficient from the initial exploration stage to the deep analysis stage is 'a', and the correlation coefficient from the deep analysis stage to the result output stage is 'b', then the adjusted reward value for the deep analysis stage can be the initial reward value for the deep analysis stage plus the initial reward value for the initial exploration stage multiplied by the correlation coefficient 'a'; the adjusted reward value for the result output stage can be the initial reward value for the result output stage plus the adjusted reward value for the deep analysis stage multiplied by the correlation coefficient 'b'. Through these adjustments, the reward value for the deep analysis stage incorporates the impact of the behavioral effects of the initial exploration stage, and the reward value for the result output stage incorporates the cumulative impact of the behavioral effects of both the initial exploration and deep analysis stages. The adjusted reward values ​​for each stage collectively constitute a stage reward signal that includes the cumulative effect between stages, more comprehensively reflecting the agent's behavioral performance throughout the entire task process.

[0151] Step S158: Transmit the stage reward signal in reverse chronological order to the decision model of the science and technology intelligence agent, and calculate the gradient change for each action.

[0152] In the scenario of analyzing the development trend of natural language processing technology, the stage reward signal, which includes the cumulative effect between stages, is backpropagated to the decision model of the science and technology intelligence agent according to the chronological order of the agent's actions. The decision model is the core part of the agent used to select and execute actions, and it contains parameters related to action selection. The backpropagation process starts from the last action in the result output stage and proceeds sequentially to the first action in the deep analysis stage and the initial exploration stage. During the propagation process, the gradient change of the parameters corresponding to each action in the decision model is calculated based on the magnitude and direction of the stage reward signal. The gradient change reflects the direction and degree to which the model parameters need to be adjusted to improve the overall reward value. For example, for actions with high reward values, the corresponding parameter gradient change will indicate that the model increases its preference for that action; for actions with low reward values, the parameter gradient change will indicate that the model decreases its preference for that action.

[0153] Step S159: Adjust the parameter weights in the decision model related to the behavior in the initial exploration stage, the behavior in the deep analysis stage, and the behavior in the result output stage based on the gradient change. For behaviors where the stage reward signal is greater than the preset reward threshold, the corresponding parameter weight increase value is calculated according to the preset ratio.

[0154] For example, step S1591: Analyze the structure of the decision-making model of the science and technology intelligence agent, determine the set of parameter weights related to the behavior in the initial exploration stage, and the set of parameter weights related to the behavior in the initial exploration stage includes the selection weight of sample unit calling behavior, the threshold weight of keyword matching behavior, and the judgment weight of information fragment marking behavior.

[0155] In the context of analyzing the development trends of natural language processing technology, understanding the structure of the decision-making model of a science and technology intelligence agent is a prerequisite for adjusting parameter weights. Decision-making models typically contain multiple levels and modules for processing input information, making behavioral decisions, and outputting behavioral selection results. Through in-depth analysis of the model structure, the modules and specific locations of parameter weights related to the behavior in the initial exploration phase are identified. Specifically, the selection weight of the sample unit invocation behavior determines which sample unit the agent chooses to invoke under what circumstances; its value affects the probability of invoking different sample units. The threshold weight of the keyword matching behavior controls the strictness of keyword matching, determining the degree to which a word needs to match preset keywords to be identified as key information. The judgment weight of the information fragment labeling behavior influences the agent's judgment of whether an information fragment is invalid; different weight values ​​lead to different criteria for judging invalid information. Integrating these parameter weights forms a set of parameter weights related to the behavior in the initial exploration phase.

[0156] Step S1592: Analyze the structure of the decision-making model of the science and technology intelligence agent, and determine the set of parameter weights related to the behavior in the deep analysis stage. The set of parameter weights related to the behavior in the deep analysis stage includes the algorithm parameter weights of the sample correlation calculation behavior, the window parameter weights of the time series feature extraction behavior, and the index weights of the information credibility assessment behavior.

[0157] In the context of analyzing the development trends of natural language processing technology, this paper further analyzes the structure of the decision-making model of the science and technology intelligence intelligent agent to identify the parameter weights related to the behavior in the deep analysis stage. The algorithm parameter weights for the sample correlation calculation behavior are the weights of the internal parameters of the algorithm used in the decision model to calculate the sample correlation. These parameters affect the result of the correlation calculation, and thus influence the agent's judgment on the degree of sample correlation. The window parameter weights for the temporal feature extraction behavior are the parameter weights that control the size of the time window during temporal feature extraction; different window sizes will extract temporal features at different scales. The indicator weights for the information credibility assessment behavior are the weights assigned to the importance of assessment indicators such as source authority, data support strength, and consistency in the decision model. These weights determine the contribution of each indicator to the final credibility assessment result. Combining the above parameter weights constitutes a set of parameter weights related to the behavior in the deep analysis stage.

[0158] Step S1593: Analyze the structure of the science and technology intelligence intelligent agent decision-making model, determine the set of parameter weights related to the behavior in the result output stage, and the set of parameter weights related to the behavior in the result output stage includes the compression ratio weight of the content condensation behavior, the dimension weight of the chart generation behavior, and the relevance weight of the scene tag matching behavior.

[0159] In the scenario of analyzing the development trend of natural language processing technology, the decision model structure is further analyzed to determine the parameter weights related to the actions in the result output stage. The compression ratio weight of the content condensation action controls the length of the generated summary text, determining the degree to which the original content is compressed. The dimensionality weight of the chart generation action influences the selection of chart type and the display of data dimensions within the chart; different dimensionality weights lead the agent to prioritize different types of charts to display data. The relevance weight of the scene label matching action measures the closeness of the correlation between the technical conclusion and the characteristics of the application scenario; a higher weight value indicates greater importance of the associated feature in scene matching. These parameter weights are integrated to form a set of parameter weights related to the actions in the result output stage.

[0160] Step S1594: The gradient change obtained by backpropagating the stage reward signal is divided into the initial exploration stage, the depth analysis stage, and the result output stage to obtain the gradient change of the initial exploration stage, the gradient change of the depth analysis stage, and the gradient change of the result output stage.

[0161] In the context of analyzing the development trends of natural language processing technology, the overall gradient change obtained through backpropagation of stage reward signals is broken down into three stages: the initial exploration stage, the deep analysis stage, and the result output stage. This is because the behaviors at different stages correspond to different sets of parameter weights in the decision model, requiring separate calculation of the gradient changes of the parameter weights for each stage. The breakdown process separates the portion of the overall gradient change related to the behaviors in the initial exploration stage, based on the stage identifier of each behavior in the behavior sequence. Similarly, it separates the gradient changes related to the behaviors in the deep analysis stage and the gradient changes related to the behaviors in the result output stage. The gradient change for each stage is a vector, with its dimension corresponding to the dimension of the parameter weight set for that stage. Each element in the vector corresponds to the gradient change value of a parameter weight.

[0162] Step S1595: Assign the gradient change amount in the initial exploration phase to the set of parameter weights related to the behavior in the initial exploration phase, and adjust the value of each parameter weight in the set of parameter weights related to the behavior in the initial exploration phase so that the parameter weight increases as the gradient change amount increases positively and decreases as the gradient change amount increases negatively.

[0163] In the scenario of analyzing the development trend of natural language processing technology, a one-to-one correspondence is established between the gradient change amount in the initial exploration phase and the weight of each parameter in the set of parameter weights related to the behavior in the initial exploration phase. For example, the selection weight of the sample unit calling behavior corresponds to the first element in the gradient change amount vector, the threshold weight of the keyword matching behavior corresponds to the second element, and so on. Then, the value of each parameter weight is adjusted according to the correspondence. The adjustment rule is: if the gradient change amount is positive and the value increases, it means that increasing the parameter weight can improve the reward value, so the value of the parameter weight is increased accordingly; if the gradient change amount is negative and the value increases, it means that decreasing the parameter weight can improve the reward value, so the value of the parameter weight is decreased accordingly. Through the above adjustments, the parameter weights in the initial exploration phase are adjusted in a direction that is conducive to improving the overall reward value.

[0164] Step S1596: Match the gradient change amount in the deep analysis stage with the set of parameter weights related to the behavior in the deep analysis stage, and adjust the value of each parameter weight in the set of parameter weights related to the behavior in the deep analysis stage. The adjustment method is the same as the adjustment method of the parameter weights in the initial exploration stage. For advanced behaviors where the stage reward signal is greater than the preset reward threshold, the corresponding parameter weights are enhanced according to the gradient change amount.

[0165] In the scenario of analyzing the development trend of natural language processing technology, a method similar to that used in the initial exploration phase is adopted to map the gradient change in the deep analysis phase to the set of parameter weights related to the behavior in the deep analysis phase. The value of each parameter weight is adjusted according to the direction and magnitude of the gradient change, in the same way as in the initial exploration phase: the parameter weight increases with a positive gradient change and decreases with a negative gradient change. Furthermore, for advanced behaviors where the stage reward signal is greater than a preset reward threshold (such as the high reward value of the sample correlation calculation behavior), the corresponding parameter weights (such as the algorithm parameter weights of the sample correlation calculation behavior) undergo additional enhancement processing on top of the adjustment based on the gradient change. This enhancement processing can be achieved by multiplying the original adjustment by an enhancement coefficient greater than 1, further increasing the parameter weights corresponding to these efficient advanced behaviors, thereby increasing the probability that the agent will select and execute these behaviors in the future.

[0166] Step S1597: Match the gradient change amount of the result output stage with the set of parameter weights related to the behavior of the result output stage, adjust the value of each parameter weight in the set of parameter weights related to the behavior of the result output stage, and enhance the corresponding parameter weights according to the gradient change amount for the final behavior when the stage reward signal is greater than the preset reward threshold, so as to increase the execution probability of this type of behavior.

[0167] In the scenario of analyzing the development trend of natural language processing technology, the gradient change in the output stage is mapped to the set of parameter weights related to the behavior in the output stage, and the value of each parameter weight is adjusted according to the same adjustment rules as the previous stages. Simultaneously, for terminal behaviors where the stage reward signal is greater than a preset reward threshold (such as the high reward value of scene label matching behavior), the corresponding parameter weights (such as the relevance weight of scene label matching behavior) are enhanced based on the gradient change adjustment. Through these enhancements, the parameter weights corresponding to these efficient terminal behaviors occupy a larger proportion in the decision model, thereby increasing the probability of the agent selecting and executing these behaviors in the output stage to generate higher quality output results.

[0168] Step S1598: Calculate the sum of the parameter weight sets related to the behavior in the initial exploration stage, the parameter weight sets related to the behavior in the deep analysis stage, and the parameter weight sets related to the behavior in the result output stage, so that the sum of parameter weights is maintained within the preset weight range, and the model is avoided from becoming unbalanced due to the weight of a single parameter being too large or too small.

[0169] In the scenario of analyzing the development trend of natural language processing technology, the sum of the weights of the adjusted parameter weight sets for each of the three stages is calculated. For example, the sum of the weights of all parameters in the initial exploration stage is obtained; similarly, the sum of the weights for the deep analysis stage and the result output stage is calculated. Then, it is checked whether these weight sums are within the preset weight range. The preset weight range is set to ensure the stability of the decision model and the balance between the weights of each parameter. If the sum of the weights of a certain stage exceeds the preset range, the parameter weights of that stage need to be appropriately scaled, such as proportionally reducing or increasing all parameter weights to bring the sum back to the preset range. This avoids the model overemphasizing a certain behavior due to an excessively large single parameter weight, or ignoring some important behaviors due to an excessively small single parameter weight, thereby preventing model imbalance and ensuring the diversity and rationality of the agent's behavior choices.

[0170] Step S1599: Reintegrate the adjusted set of parameter weights related to the behavior in the initial exploration phase, the set of parameter weights related to the behavior in the deep analysis phase, and the set of parameter weights related to the behavior in the result output phase into the decision model, replacing the original parameter weights in the decision model.

[0171] In the scenario of analyzing the development trend of natural language processing technology, the parameter weight sets related to the initial exploration stage, the deep analysis stage, and the result output stage, after adjustment and weight summation checks, are reintegrated into the decision model according to their original positions and structures. The new parameter weight sets completely replace the original corresponding parameter weights in the decision model, updating it to a version with the latest parameter adjustments. This step completes the physical update of the decision model parameters, enabling the agent to make decisions based on the new parameter weights in subsequent action selection and execution, demonstrating the effectiveness of training iteration.

[0172] Step S15910: Simulate the process of the science and technology intelligence agent performing basic, advanced and terminal behaviors using the adjusted parameter weights, and count the percentage of parameter weights corresponding to behaviors in which the reward signal of the science and technology intelligence agent is greater than the preset reward threshold during the execution phase, ensuring that the percentage increases according to the preset ratio.

[0173] In the scenario of analyzing the development trend of natural language processing technology, to verify the effect of parameter weight adjustment, the process of a scientific and technological intelligence agent executing the entire intelligence task using the adjusted parameter weights was simulated. During the simulation, the agent sequentially executed basic behaviors in the initial exploration phase, advanced behaviors in the deep analysis phase, and terminal behaviors in the result output phase. Each behavior executed by the agent and its corresponding parameter weights were recorded, and the sum of parameter weights corresponding to behaviors whose stage reward signals exceeded a preset reward threshold, as well as the proportion of these parameter weights to the total sum of all behavior parameter weights, were calculated. This proportion was compared with the proportion before adjustment to ensure that the adjusted proportion increased according to the preset ratio. The preset ratio is set according to the training objective, such as wanting to increase the proportion of parameter weights for efficient behaviors by a certain percentage. If the preset increase ratio is not achieved, it may be necessary to re-examine the parameter weight adjustment process or gradient calculation for problems and make corresponding corrections to ensure that the training iterations achieve the expected results.

[0174] Step S1510: Integrate the adjusted parameter weights to form updated decision parameters, thus completing the update process of the decision parameters of the science and technology intelligence agent.

[0175] In the scenario of analyzing the development trend of natural language processing technology, all parameter weights (including parameter weights related to behavior in the initial exploration stage, deep analysis stage, and result output stage) that have been adjusted, integrated, and verified through the above steps are finally integrated. These parameter weights together constitute the core decision parameters of the science and technology intelligence agent decision-making model. The integrated parameter weights are then determined as the updated decision parameters, formally replacing the agent's original decision parameters. This completes the update process of the science and technology intelligence agent's decision parameters, ending a single training iteration. The updated decision parameters will guide the agent in making behavioral choices and executions in the next training iteration or in the execution of actual intelligence tasks, enabling the agent to continuously learn and optimize its behavioral strategies, gradually improving its ability and effectiveness in completing intelligence tasks.

[0176] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a reinforcement learning-based science and technology intelligence agent training system 100 provided in this application embodiment for executing the above-described reinforcement learning-based science and technology intelligence agent training method. The reinforcement learning-based science and technology intelligence agent training system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0177] Alternatively, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 stores machine-executable instructions for executing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the reinforcement learning-based scientific and technological intelligence agent training method provided in the aforementioned method embodiments.

[0178] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for training a scientific and technological intelligence agent based on reinforcement learning, characterized in that, The method includes: The multi-stage evolution characteristics of intelligence tasks corresponding to the training of science and technology intelligence intelligent agents are determined. The multi-stage evolution characteristics of intelligence tasks include attributes of the initial exploration stage, attributes of the deep analysis stage, and attributes of the result output stage. Construct a reinforcement learning exploration space adapted to the multi-stage evolution characteristics of the intelligence mission, wherein the reinforcement learning exploration space includes the set of possible behaviors and stage transition conditions corresponding to each stage; The basic behaviors that can be performed in the initial exploration phase include sample unit calling, keyword matching, and information fragment marking. The advanced behaviors that can be performed in the deep analysis phase include sample correlation calculation, time series feature extraction, and information credibility assessment. The final behaviors that can be performed in the results output phase include content condensation, chart generation, and scene tag matching. The conditions for transitioning from the initial exploration phase to the deep analysis phase are determined as follows: the science and technology intelligence agent completes a preset number of basic behaviors in the selectable behavior set of the initial exploration phase and the key information identification coverage reaches the threshold of the initial exploration phase. The conditions for transitioning from the deep analysis stage to the result output stage are determined as follows: the science and technology intelligence agent completes a preset number of advanced behaviors in the selectable set of behaviors in the deep analysis stage, and the completeness of intelligence association mining reaches the threshold of the deep analysis stage. Generate a phased intelligence sample cluster adapted to the reinforcement learning exploration space, wherein the phased intelligence sample cluster contains a sample subset corresponding to each phase; The initial exploration phase intelligence sample subset includes samples of technical field overview, research direction introduction, and core terminology explanation; the in-depth analysis phase intelligence sample subset includes samples of experimental data comparison, technology evolution process, and academic debate; and the results output phase intelligence sample subset includes samples of research conclusion summary, technology application case, and data visualization template. The technology intelligence agent is invoked to interact with the phased intelligence sample clusters in the reinforcement learning exploration space in stages, forming a cross-stage agent behavior sequence; The agent behavior sequence is formed by collecting basic behavior records from the initial exploration phase, advanced behavior records from the deep analysis phase, and final behavior records from the result output phase, along with sample response information corresponding to each phase, and is arranged in chronological order, and includes phase identifiers, behavior types, and response associations. A multi-stage reward function is constructed based on the multi-stage evolution characteristics of the intelligence task. Stage reward signals are generated according to the agent's behavior sequence. The decision parameters of the science and technology intelligence agent are updated using the stage reward signals to complete a single training iteration.

2. The method for training a science and technology intelligence agent based on reinforcement learning according to claim 1, characterized in that, The construction of a reinforcement learning exploration space adapted to the multi-stage evolution characteristics of the intelligence mission includes: The initial exploration stage attributes in the multi-stage evolution characteristics of the intelligence mission are analyzed, and the core objectives of the initial exploration stage are extracted. The core objectives of the initial exploration stage include the initial screening of intelligence sources, the rapid identification of key information, and the elimination of invalid information. Based on the core objectives of the initial exploration phase, the basic behaviors that the science and technology intelligence agent can perform in the initial exploration phase are determined, forming a set of optional behaviors for the initial exploration phase. The attributes of the deep analysis stage in the multi-stage evolution characteristics of the intelligence mission are analyzed, and the core objectives of the deep analysis stage are extracted. The core objectives of the deep analysis stage include intelligence correlation mining objectives, technology trend prediction objectives, and contradictory information identification objectives. Based on the core objectives of the deep analysis phase, we determine the advanced behaviors that the science and technology intelligence agent can perform in the deep analysis phase, forming a set of optional behaviors for the deep analysis phase. The results output stage attributes in the multi-stage evolution characteristics of the intelligence mission are analyzed, and the core objectives of the results output stage are extracted. The core objectives of the results output stage include intelligence summary generation objectives, conclusion visualization objectives, and application scenario matching objectives. Based on the core objectives of the result output stage, the executable terminal behaviors of the science and technology intelligence agent in the result output stage are determined, forming a set of optional behaviors for the result output stage. The optional sets of behaviors in the initial exploration phase, the deep analysis phase, and the result output phase, as well as the conditions for transitioning from the initial exploration phase to the deep analysis phase and from the deep analysis phase to the result output phase, are structurally integrated to form a reinforcement learning exploration space that includes phase identifiers, behavior lists, and transition rules.

3. The method for training a science and technology intelligence agent based on reinforcement learning according to claim 2, characterized in that, Based on the core objectives of the initial exploration phase, the basic behaviors that a science and technology intelligence agent can perform in the initial exploration phase are determined, forming a set of optional behaviors for the initial exploration phase, including: For the initial screening of intelligence sources in the initial exploration phase, the analysis of the operations that the science and technology intelligence agent needs to perform is analyzed, and the specific manifestations of sample unit calling behavior are defined. The specific manifestations of sample unit calling behavior include selecting a single intelligence sample unit from a phased intelligence sample cluster, batch retrieving similar intelligence sample units, and retrieving relevant intelligence sample units based on topic tags. For sample unit invocation behavior, the triggering condition for the execution of sample unit invocation behavior is set. The triggering condition for the execution of sample unit invocation behavior is that when the science and technology intelligence agent enters the initial exploration stage and has not loaded any intelligence sample units, the sample unit invocation behavior is automatically triggered. To quickly identify targets based on key information in the initial exploration phase, the analysis of the operations that the science and technology intelligence agent needs to perform is conducted. The specific manifestations of keyword matching behavior are defined, including comparing preset technical keywords with words in intelligence sample units, extracting terms from intelligence sample units whose frequency of occurrence meets preset standards as candidate keywords, and verifying the relevance of candidate keywords to the task theme. For keyword matching behavior, the triggering condition for keyword matching behavior is set. The triggering condition for keyword matching behavior is that after the science and technology intelligence agent completes the sample unit calling behavior and loads the intelligence sample unit, the keyword matching behavior is automatically triggered. To eliminate invalid information in the initial exploration phase, the analysis focuses on the operations that the science and technology intelligence agent needs to perform, and defines the specific manifestations of information fragment marking behavior. These specific manifestations include marking paragraphs in intelligence sample units that are irrelevant to the task topic, marking recurring information content, and marking statements whose credibility does not meet the preset requirements. For the information fragment marking behavior, the triggering conditions for the execution of the information fragment marking behavior are set. The triggering condition for the execution of the information fragment marking behavior is that the information fragment marking behavior is automatically triggered after the science and technology intelligence intelligent agent completes the keyword matching behavior and identifies key information. Collect the specific manifestations and triggering conditions of sample unit calling behavior, keyword matching behavior, and information fragment marking behavior to form a list of behavior descriptions; The behaviors in the behavior description list are deduplicated and logically ordered to ensure that there are no conflicts between the behaviors and that they conform to the execution flow of the initial exploration phase. The sorted list of behavior descriptions is determined as the initial exploration phase behavior option set, and it is clarified that the execution object of each behavior in the initial exploration phase behavior option set is the initial exploration phase intelligence sample subset in the phased intelligence sample cluster.

4. The method for training a science and technology intelligence agent based on reinforcement learning according to claim 1, characterized in that, The generation of a phased intelligence sample cluster adapted to the reinforcement learning exploration space includes: Extract the initial exploration stage identifier and the corresponding initial exploration stage behavior selection set from the reinforcement learning exploration space, and determine the basic feature requirements of the intelligence sample for the initial exploration stage. The basic feature requirements of the intelligence sample for the initial exploration stage include the theme clarity feature requirements, keyword density feature requirements, and information fragment integrity feature requirements. Based on the basic characteristics requirements for intelligence samples in the initial exploration phase, intelligence samples that meet the requirements are selected from the overall science and technology intelligence database to form a subset of intelligence samples for the initial exploration phase. Extract the deep analysis stage identifiers and corresponding deep analysis stage behavior options from the reinforcement learning exploration space, and determine the advanced feature requirements of the deep analysis stage for intelligence samples. The advanced feature requirements of the deep analysis stage for intelligence samples include data correlation feature requirements, time series continuity feature requirements, and opinion conflict feature requirements. Based on the advanced feature requirements of intelligence samples in the in-depth analysis stage, intelligence samples that meet the requirements are selected from the general science and technology intelligence database to form a subset of intelligence samples for the in-depth analysis stage. Extract the result output stage identifiers and corresponding result output stage behavior options from the reinforcement learning exploration space, and determine the final feature requirements of the result output stage for the intelligence sample. The final feature requirements of the result output stage for the intelligence sample include the conclusion conciseness feature requirements, the application scenario adaptability feature requirements, and the visualization compatibility feature requirements. Based on the final characteristic requirements of intelligence samples in the result output stage, intelligence samples that meet the requirements are selected from the overall science and technology intelligence database to form a subset of intelligence samples for the result output stage. Analyze the content association between the intelligence sample subsets in the initial exploration phase and the intelligence sample subsets in the in-depth analysis phase, and calculate the thematic overlap between the intelligence sample subsets in the initial exploration phase and the intelligence sample subsets in the in-depth analysis phase. Analyze the content relationship between the intelligence sample subset in the in-depth analysis stage and the intelligence sample subset in the result output stage, and calculate the logical coherence between the intelligence sample subset in the in-depth analysis stage and the intelligence sample subset in the result output stage. Based on the degree of thematic overlap and logical coherence, the sample quantity ratios of the intelligence sample subsets in the initial exploration stage, the intelligence sample subsets in the in-depth analysis stage, and the intelligence sample subsets in the result output stage are adjusted so that the correlation strength between sample subsets in adjacent stages is adapted to the stage transition requirements of the reinforcement learning exploration space. The adjusted subsets of intelligence samples from the initial exploration phase, the in-depth analysis phase, and the results output phase are combined in an orderly manner to form a phased intelligence sample cluster containing phase labels and sample association maps.

5. The method for training a science and technology intelligence agent based on reinforcement learning according to claim 4, characterized in that, Based on the basic characteristic requirements for intelligence samples in the initial exploration phase, intelligence samples that meet the requirements are selected from the overall science and technology intelligence database to form a subset of intelligence samples for the initial exploration phase, including: The analysis of the requirements for thematic clarity of intelligence samples in the initial exploration stage is used to define the criteria for measuring thematic clarity. The criteria for measuring thematic clarity include the consistency between the title and core content of the intelligence sample reaching the preset thematic clarity threshold, the frequency of occurrence of thematic keywords in the abstract meeting the preset frequency requirements, and the degree of concentration of thematic discussion in the main text meeting the preset concentration standard. Based on the criteria for determining the clarity of the subject matter, each intelligence sample in the science and technology intelligence database is evaluated, and intelligence samples that meet the requirements for the clarity of the subject matter are retained to form a subject-selected subset; The keyword density feature requirements for intelligence samples in the initial exploration phase are analyzed, and the calculation method of keyword density is defined. The keyword density is calculated as the ratio of the total number of occurrences of core keywords in the intelligence sample to the total number of characters in the sample. This ratio is required to reach a preset density threshold. Based on the keyword density calculation method, the intelligence samples in the topic selection subset are calculated, and the intelligence samples that meet the keyword density requirements are retained to form the keyword density selection subset; The analysis of the information fragment integrity characteristics requirements for intelligence samples in the initial exploration stage defines the judgment criteria for information fragment integrity. The judgment criteria for information fragment integrity include the completeness of the technical background introduction reaching the preset background completeness threshold, the completeness of the core concept definition meeting the preset concept completeness standard, and the completeness of the research scope definition meeting the preset scope definition requirements. Based on the criteria for judging the integrity of information fragments, the intelligence samples in the keyword density screening subset are checked, and the intelligence samples that meet the requirements for information fragment integrity are retained to form the information integrity screening subset. Intelligence samples describing the overall situation of a technical field are selected from the information integrity screening subset to form a technical field overview sample group, which includes field overview samples of different technical branches. Intelligence samples with specific research directions as the description objects are selected from the information integrity screening subset to form a research direction introduction sample group. The research direction introduction sample group includes the core content and development status of each research direction. Intelligence samples with technical terminology explanations as the description object are selected from the information integrity screening subset to form a core terminology explanation sample group. The core terminology explanation sample group includes definition samples of basic terms, professional terms and emerging terms. The sample groups for the overview of the technical field, the introduction of the research direction, and the explanation of the core terms were merged, and duplicate samples were removed to form a subset of intelligence samples for the initial exploration phase. Add stage identifiers and topic tags to each sample in the intelligence sample subset of the initial exploration phase to facilitate rapid identification and retrieval by the science and technology intelligence agent during the interaction process.

6. The method for training a science and technology intelligence agent based on reinforcement learning according to claim 1, characterized in that, The invoked technology intelligence agent interacts with the phased intelligence sample clusters in the reinforcement learning exploration space in stages, forming a cross-stage agent behavior sequence, including: Configure the initial state of the science and technology intelligence agent to correspond to the initial exploration stage of the reinforcement learning exploration space, and load the initial exploration stage intelligence sample subset from the staged intelligence sample cluster. The science and technology intelligence agent is invoked to select basic behaviors from the set of behaviors available in the initial exploration phase, executes the selected basic behaviors on a subset of intelligence samples in the initial exploration phase, and records the execution process of the basic behaviors and the corresponding sample response information. Monitor the basic behavior execution progress of the science and technology intelligence agent. When the conditions for transitioning from the initial exploration phase to the deep analysis phase are met, switch the state of the science and technology intelligence agent to the deep analysis phase and load the deep analysis phase intelligence sample subset from the phased intelligence sample cluster. The technology intelligence agent is invoked to select advanced behaviors from the set of behaviors available in the deep analysis stage. The selected advanced behaviors are then executed on a subset of intelligence samples from the deep analysis stage. The execution process of the advanced behaviors and the corresponding sample response information are recorded, and the basic behavior records from the initial exploration stage are also associated with them. Monitor the progress of advanced behaviors of the science and technology intelligence agent. When the conditions for transitioning from the deep analysis stage to the result output stage are met, switch the state of the science and technology intelligence agent to the result output stage and load the result output stage intelligence sample subset from the staged intelligence sample cluster. The technology intelligence agent is invoked to select terminal behaviors from the set of possible behaviors in the result output stage, executes the selected terminal behaviors on a subset of intelligence samples in the result output stage, records the execution process of the terminal behaviors and the corresponding sample response information, and associates them with the advanced behavior records in the deep analysis stage. The connection points between the initial exploration phase and the deep analysis phase, and between the deep analysis phase and the result output phase, are marked in the agent's behavior sequence. This clarifies the causal relationships between different phases of behavior, enabling the agent's behavior sequence to fully reflect the cross-phase interaction process of the science and technology intelligence agent in the reinforcement learning exploration space.

7. The method for training a science and technology intelligence agent based on reinforcement learning according to claim 6, characterized in that, The technology intelligence agent selects advanced behaviors from the set of behaviors available in the deep analysis phase, executes the selected advanced behaviors on a subset of intelligence samples from the deep analysis phase, records the execution process of the advanced behaviors and the corresponding sample response information, and simultaneously associates them with the basic behavior records from the initial exploration phase, including: Extract the advanced behavior list of the optional set of behaviors in the deep analysis stage of reinforcement learning exploration space, and show the behavior description and execution object of the sample correlation calculation behavior, temporal feature extraction behavior and information credibility assessment behavior to the science and technology intelligence agent; Based on the content characteristics of the currently loaded deep analysis stage intelligence sample subset, the science and technology intelligence agent selects the first advanced behavior that matches the sample content characteristics from the advanced behavior list. If the sample content contains multiple related data, the sample correlation calculation behavior is selected first. The science and technology intelligence agent is invoked to perform pairwise comparisons on multiple samples in the intelligence sample subset of the in-depth analysis phase, extract the common technical terms, research methods and experimental conclusions among the samples, calculate the proportion of these common elements as the correlation parameter, and require the correlation parameter to reach the preset correlation threshold. The execution process of the sample correlation calculation is recorded. The execution process of the sample correlation calculation includes the compared sample pairs, the extracted common elements and the calculated correlation parameters. At the same time, the response changes of the samples in the intelligence sample subset during the deep analysis stage are recorded. The response changes of the samples in the intelligence sample subset during the deep analysis stage are the state of the samples after the correlation parameters are marked. The science and technology intelligence agent is invoked to process the samples with time series attributes in the intelligence sample subset of the in-depth analysis stage, arrange the data points in the samples in chronological order, and extract the trend characteristics, periodic characteristics and mutation characteristics of the data points as a function of time. The execution process of time series feature extraction is recorded. The execution process of time series feature extraction includes the division method of time series, the extracted trend feature parameters, periodic feature parameters and mutation feature parameters. At the same time, the response changes of samples in the intelligence sample subset in the deep analysis stage are recorded. The response changes of samples in the intelligence sample subset in the deep analysis stage are the state of the samples after they are labeled with time series features. The technology intelligence agent is invoked to process samples containing viewpoints or conclusions in the intelligence sample subset of the in-depth analysis stage. The source authority, data support strength and consistency with other samples of the samples are analyzed to generate a credibility level description. The credibility level description is divided into multiple levels according to preset standards. Record the execution process of information credibility assessment behavior, which includes the basis for source analysis, the criteria for judging the strength of data support and the consistency comparison results. At the same time, record the response changes of samples in the intelligence sample subset during the in-depth analysis stage. The response changes of samples in the intelligence sample subset during the in-depth analysis stage are the state of the samples after they are marked with credibility level. Retrieve basic behavior records from the initial exploration phase, identify basic behaviors related to samples in the current in-depth analysis phase during the initial exploration phase, including sample unit calling behavior or keyword matching behavior for the same topic, and establish association markers between basic behaviors related to samples in the initial exploration phase and advanced behaviors in the in-depth analysis phase. The advanced behavior execution process, sample response information, and association identifiers with the basic behavior records of the initial exploration phase are integrated to form a complete interaction record of the deep analysis phase.

8. The method for training a science and technology intelligence agent based on reinforcement learning according to claim 1, characterized in that, The process of constructing a multi-stage reward function based on the multi-stage evolution characteristics of the intelligence task, generating stage reward signals according to the agent's behavior sequence, and updating the decision parameters of the science and technology intelligence agent using the stage reward signals includes: To address the initial exploration stage attribute in the multi-stage evolution characteristics of intelligence missions, a reward sub-function for the initial exploration stage is constructed. The input of the reward sub-function for the initial exploration stage is the basic behavior of the science and technology intelligence agent in the initial exploration stage and the corresponding sample response information. The output of the reward sub-function for the initial exploration stage is the reward value corresponding to the effect of the basic behavior. The reward value corresponding to the effect of the basic behavior is positively correlated with the increase of the key information identification coverage. To address the deep analysis stage attribute in the multi-stage evolution characteristics of intelligence missions, a reward sub-function for the deep analysis stage is constructed. The input of the reward sub-function for the deep analysis stage is the advanced behavior of the science and technology intelligence agent in the deep analysis stage and the corresponding sample response information. The output of the reward sub-function for the deep analysis stage is the reward value corresponding to the effect of the advanced behavior. The reward value corresponding to the effect of the advanced behavior is positively correlated with the improvement of the completeness of intelligence association mining. To address the result output stage attribute in the multi-stage evolution characteristics of intelligence missions, a reward sub-function for the result output stage is constructed. The input of the reward sub-function for the result output stage is the terminal behavior of the science and technology intelligence agent in the result output stage and the corresponding sample response information. The output of the reward sub-function for the result output stage is the reward value corresponding to the effect of the terminal behavior. The reward value corresponding to the effect of the terminal behavior is positively correlated with the improvement of the matching degree between the conclusion and the application scenario. Set an inter-stage reward correlation coefficient. The inter-stage reward correlation coefficient is used to quantify the degree of influence of the reward value of the previous stage on the reward value of the next stage, so that the reward value of the subsequent stage can reflect the cumulative effect of the previous behavior. The reward sub-functions of the initial exploration phase, the deep analysis phase, the result output phase, and the reward correlation coefficients between phases are integrated to form a multi-stage reward function; The agent's behavior sequence is split into the initial exploration phase, the deep analysis phase, and the result output phase. The behavior and sample response information of each phase are extracted and input into the corresponding phase reward sub-function to obtain the initial reward value of the initial exploration phase, the initial reward value of the deep analysis phase, and the initial reward value of the result output phase. Based on the inter-stage reward correlation coefficient, the initial reward values ​​of the initial exploration stage, the initial reward values ​​of the in-depth analysis stage, and the initial reward values ​​of the result output stage are adjusted to obtain a stage reward signal that includes the inter-stage cumulative effect. The stage reward signal is transmitted in reverse chronological order to the decision-making model of the science and technology intelligence agent to calculate the gradient change for each behavior. The parameter weights in the decision-making model related to the behavior in the initial exploration stage, the behavior in the deep analysis stage, and the behavior in the result output stage are adjusted based on the gradient change. For behaviors where the stage reward signal is greater than the preset reward threshold, the corresponding parameter weight increase value is calculated according to the preset ratio. The adjusted parameter weights are integrated to form updated decision parameters, thus completing the process of updating the decision parameters of the science and technology intelligence intelligent agent.

9. A training system for a scientific and technological intelligence agent based on reinforcement learning, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the reinforcement learning-based scientific and technological intelligence agent training method according to any one of claims 1 to 8 by executing the machine-executable instructions.

10. A computer program product, characterized in that, The computer program product includes machine-executable instructions stored in a computer-readable storage medium. A processor of a computer device reads the machine-executable instructions from the computer-readable storage medium and executes the machine-executable instructions, causing the computer device to perform the reinforcement learning-based scientific and technological intelligence agent training method as described in any one of claims 1 to 8.

Images