Science and technology intelligence agent collaborative decision-making method and system fusing multi-modal perception

Abstract

This invention provides a collaborative decision-making method and system for science and technology intelligence agents that integrates multimodal perception, relating to the field of science and technology intelligence processing technology. First, multiple science and technology intelligence perception information is acquired, and a preliminary screening of perception deviations is performed to generate an initial list of perception deviations. Then, based on the list, the perception acquisition process parameters corresponding to the discrepancies are traced, and cross-agent dynamic calibration rules are formulated to obtain a calibrated and optimized set of science and technology intelligence perceptions. Next, a cross-agent joint mining function is invoked to extract multi-level correlation features, generating initial decision suggestions. These initial decision suggestions are sent to other processing carriers, whose suggestions are received and analyzed. Conflict points are identified, the perception set is supplemented and optimized, and adjusted decision suggestions are generated. Finally, all adjusted decision suggestions are summarized, conflict resolution and content fusion are performed, and a final collaborative decision-making scheme is generated, achieving efficient collaboration among multiple agents.

Description

Technical Field

[0001] This invention relates to the field of scientific and technological information processing technology, and more specifically, to a collaborative decision-making method and system for scientific and technological information intelligent agents that integrates multimodal perception. Background Technology

[0002] In the field of science and technology intelligence processing, with the rapid development of information technology, the sources of science and technology intelligence information are becoming increasingly diversified, covering multiple modalities such as text, images, and audio. However, existing science and technology intelligence processing methods have many problems in dealing with multimodal information.

[0003] On the one hand, perceptual information from different sources may exhibit perceptual biases due to factors such as acquisition equipment and environment. Existing technologies often lack effective perceptual bias screening mechanisms, making it difficult to accurately identify biased perceptual information and leading to inaccurate subsequent analysis results. On the other hand, when multiple agents collaboratively process scientific and technological intelligence, the perceptual information acquired by each agent may be missing or incomplete. Existing methods struggle to achieve dynamic calibration across agents, failing to effectively supplement and optimize missing perceptual content, thus affecting the comprehensiveness and accuracy of scientific and technological intelligence processing. Furthermore, when generating decision recommendations, suggestions from different processing vehicles may conflict. Existing technologies lack effective conflict identification and resolution mechanisms, making it difficult to generate high-quality collaborative decision-making solutions and failing to meet the demands for intelligent processing of scientific and technological intelligence in complex technological environments. 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 collaborative decision-making method for a technology intelligence intelligent agent that integrates multimodal perception, the method comprising:

[0005] Multiple science and technology intelligence perception information are acquired, and a preliminary screening of perception deviations is performed on each science and technology intelligence perception information. Perception information pairs with different content descriptions under the same perception scenario are identified, and an initial list of perception deviations is generated. The science and technology intelligence perception information includes corresponding perception information identifiers, information content, and perception dimension tags.

[0006] Based on the initial list of perception deviations, the parameters of the perception acquisition process corresponding to the discrepancies are traced, the correlation between parameter differences and perception deviations is analyzed, a perception deviation source tracing report is generated, and cross-agent dynamic calibration rules are formulated in conjunction with the perception deviation source tracing report. The missing perception content is supplemented and optimized according to the cross-agent dynamic calibration rules to obtain the calibrated and optimized set of scientific and technological intelligence perception.

[0007] The cross-agent joint mining function is invoked to extract multi-level correlation features from the calibrated and optimized science and technology intelligence perception set, and integrate them to form a multi-level correlation feature set.

[0008] Initial decision suggestions are generated based on a multi-level associated feature set. The initial decision suggestions are sent to other science and technology intelligence perception information processing carriers. The initial decision suggestions from other science and technology intelligence perception information processing carriers are received and analyzed. Conflict points in the decision suggestions are identified. Based on the conflict points, the corresponding missing perception information items are located. The optimized science and technology intelligence perception set is supplemented, optimized, and calibrated to generate adjusted decision suggestions.

[0009] All revised decision recommendations are compiled, and conflict resolution and content integration are carried out to generate a final collaborative decision-making solution.

[0010] Furthermore, embodiments of the present invention also provide a collaborative decision-making system for a science and technology intelligence intelligent agent that integrates multimodal perception, 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 aforementioned collaborative decision-making method for a technology intelligence agent that integrates multimodal perception 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, the processor executing the machine-executable instructions, causing the computer device to execute the above-described collaborative decision-making method for a technology intelligence intelligent agent that integrates multimodal perception.

[0013] Based on the above, by acquiring multiple pieces of scientific and technological intelligence perception information and conducting preliminary screening for perception deviations, discrepancies in perception information pairs are accurately identified, generating an initial list of perception deviations. Based on this initial list, the perception acquisition process parameters corresponding to the discrepancies are traced, and the correlation between parameter differences and perception deviations is analyzed. Cross-agent dynamic calibration rules are formulated, effectively supplementing and optimizing missing perception content, resulting in a calibrated and optimized set of scientific and technological intelligence perception information. This improves the completeness and accuracy of the perception information. Cross-agent joint mining functions are invoked to extract multi-level correlation features, integrating them to form a multi-level correlation feature set. This deeply explores the inherent correlations within scientific and technological intelligence. Initial decision suggestions are generated based on the multi-level correlation feature set. Through interaction with other scientific and technological intelligence perception information processing carriers, conflict points in the decision suggestions are identified, the perception set is supplemented and optimized, and adjusted decision suggestions are generated. Finally, all adjusted decision suggestions are summarized, conflict resolution and content fusion are performed, and a final collaborative decision-making scheme is generated, achieving efficient collaboration among multiple agents. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the execution flow of the collaborative decision-making method for science and technology intelligence agents that integrates multimodal perception, provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of exemplary hardware and software components of the collaborative decision-making system for a science and technology intelligence intelligent agent that integrates multimodal perception, 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 collaborative decision-making method for a science and technology intelligence agent that integrates multimodal perception, provided in one embodiment of the present invention. The following is a detailed description of this collaborative decision-making method for a science and technology intelligence agent that integrates multimodal perception.

[0017] Step S110: Obtain multiple science and technology intelligence perception information, conduct preliminary screening of perception deviations for each science and technology intelligence perception information, identify perception information pairs with different content descriptions under the same perception scenario, and generate an initial list of perception deviations. The science and technology intelligence perception information includes corresponding perception information identifiers, information content, and perception dimension tags.

[0018] In this embodiment, we take sentiment analysis technology in natural language processing within the field of artificial intelligence as an example. Assume there are multiple different sentiment analysis systems that collect science and technology intelligence perception information from different sources. For example, System A collects user comments on a certain technology product from social media platforms, System B collects discussions on a certain technology from professional technology forums, and System C extracts descriptions of technology events from news reports. This information from different sources constitutes the science and technology intelligence perception information, and each piece of information has a corresponding perception information identifier to uniquely identify it. The information content includes specific text content, sentiment tendencies, etc. Perception dimension tags can be product evaluation dimensions, technology discussion dimensions, event description dimensions, etc.

[0019] Step S111: Extract the perception scene description content from each piece of scientific and technological intelligence perception information. The perception scene description content includes the environmental conditions for information collection, the collection time range, the status of the collection object, and the model of the collection device.

[0020] For System A, which collects comments from social media platforms, the environmental conditions for information collection might include network stability and platform rule restrictions; the collection timeframe could be a specific period, such as comments within a week; the status of the collected objects could include user activity and comment length; and the collection device model could be understood as the hardware devices such as servers used by System A. For System B, which collects discussions from professional technology forums, the environmental conditions for information collection might include the forum's atmosphere and the level of discussion; the collection timeframe could be a specific period, such as discussions within a month; the status of the collected objects could include the professional background of the technical personnel and the depth of the discussion; and the collection device model could be the forum's server. For System C, which extracts descriptions from news reports, the environmental conditions for information collection might include the credibility of the news source and the timeliness of the report; the collection timeframe could be news reports within a specific period; the status of the collected objects could include the news's theme and the angle of the report; and the collection device model could be the software tools such as web crawlers used by System C.

[0021] Step S112: Group scientific and technological intelligence perception information with completely consistent perception scene description content into the same perception scene group, assign a unique scene group identifier to each perception scene group, and record the perception information identifiers of all scientific and technological intelligence perception information in each group.

[0022] For example, all comments about a certain technology product collected by System A from a certain social media platform within a week can be grouped into a perception scenario group, assigned the scenario group identifier "Social Media Product Comments Group," and the perception information identifiers of all comments within this group can be recorded. Similarly, all discussions about a certain technology collected by System B from a certain professional technology forum within a month can be grouped into a perception scenario group, assigned the scenario group identifier "Professional Forum Technology Discussion Group," and the perception information identifiers can be recorded. All descriptions of a certain technology event extracted by System C from a certain news source within a certain time period can be grouped into a perception scenario group, assigned the scenario group identifier "News Report Event Description Group," and the perception information identifiers can be recorded.

[0023] Step S113: For each perception scenario group, extract the information content of all scientific and technological intelligence perception information within the group, and break it down into multiple perception content sub-items according to the logical structure of the information content. Each perception content sub-item corresponds to a specific perception conclusion.

[0024] Taking a "social media product review group" as an example, extract the information content of all reviews within the group, such as the text content and sentiment. Then, break it down into different perceptual content sub-items according to a logical structure, such as product advantages, product disadvantages, and user satisfaction in the reviews. Each perceptual content sub-item corresponds to a specific perceptual conclusion, such as whether the product advantages meet expectations, whether the product disadvantages seriously affect usability, and the level of user satisfaction.

[0025] Step S114: Compare the same perception content sub-items of different science and technology intelligence perception information within the same perception scene group, identify perception content sub-item pairs with different expressions, and record the perception information identifiers of the two science and technology intelligence perception information corresponding to the difference perception content sub-item pairs.

[0026] In the "Social Media Product Review Group," comparing the product advantages sub-items in different reviews may reveal discrepancies between the advantages mentioned in some reviews and those in others. For example, one review might mention "easy to use," while another might mention "powerful features." These discrepancies in perceived content sub-items will be identified and their corresponding perceived information tags recorded.

[0027] Step S115: Perform a difference feature description on the perceptual content sub-items with different expressions. The difference feature description includes the difference expression type, the perceptual parameters involved in the difference, and the semantic distance of the difference content.

[0028] Regarding the differences in the aforementioned product advantages, the types of differences can be: different focuses in the description; the perceptual parameters involved in the differences are the product advantages themselves; and the semantic distance between the differences can be measured by calculating the similarity between the two descriptions. For example, a similarity algorithm from natural language processing, such as the cosine similarity algorithm, can be used to calculate the similarity between the descriptions "easy to operate" and "powerful functions." If the similarity is low, it indicates a large semantic distance between the differences.

[0029] Step S116: Associate and record the scene group identifier, the difference perception content sub-item pair, the difference feature description and the corresponding perception information identifier to generate a draft of the initial list of perception deviations.

[0030] For example, for the difference-perceived content sub-item pair "easy to use" and "powerful features" in the "social media product review group", the scenario group identifier "social media product review group", the difference-perceived content sub-item pair "easy to use - powerful features", the difference feature description "different description focus, semantic distance calculated by cosine similarity algorithm", and the perception information identifiers of the corresponding two comments are associated and recorded to generate a draft of the initial list of perception biases.

[0031] Step S117: Supplement the draft of the initial list of perception biases with any missing perception scenario groups or difference perception content sub-items, and merge the difference items that are repeatedly recorded in the draft of the initial list of perception biases to form the initial list of perception biases.

[0032] The draft initial list of perceptual biases is reviewed to ensure that no perceptual scenario groups or differential perceptual content sub-items are omitted. Simultaneously, duplicate recorded discrepancies are merged; for example, if multiple differential perceptual content sub-item pairs have the same discrepancy feature description, they are merged into one record. This ultimately forms the complete initial list of perceptual biases.

[0033] Step S120: Based on the initial list of perception deviations, trace the perception acquisition process parameters corresponding to the differences, analyze the correlation between parameter differences and perception deviations, generate a perception deviation source tracing report, formulate cross-agent dynamic calibration rules based on the perception deviation source tracing report, supplement and optimize the missing perception content according to the cross-agent dynamic calibration rules, and obtain the calibrated and optimized science and technology intelligence perception set.

[0034] Step S121: Extract the perception information identifier of the scientific and technological intelligence perception information corresponding to each difference record from the initial list of perception deviations, and call the corresponding perception acquisition process log according to the perception information identifier.

[0035] Taking the discrepancy record of "easy to use" and "powerful features" in the "social media product review group" as an example, the perceptual information identifiers of the two corresponding reviews are extracted from the initial list of perceptual deviations. Then, based on these two perceptual information identifiers, the perceptual information collection process log of System A, which collects reviews from the social media platform, is invoked. The perceptual information collection process log may contain information such as the time and location of review collection, and relevant user information.

[0036] Step S122: Extract the sensing acquisition process parameters from the sensing acquisition process log. The sensing acquisition process parameters include the sampling frequency of the acquisition device, the data preprocessing algorithm type, the record of interference factors in the sensing environment, and the execution of the operation procedures by the acquisition personnel.

[0037] For the case where System A collects comments from social media platforms, the sampling frequency of the collection device could be how often System A retrieves a comment from the platform; the data preprocessing algorithm type could include text cleaning, word segmentation, part-of-speech tagging, etc.; the recording of interference factors in the perception environment could include network fluctuations, platform maintenance time, etc.; the execution of the collection personnel's operating procedures could include whether they collected comments according to the prescribed procedures and whether they correctly tagged the comments, etc.

[0038] Step S123: Compare the sensing and acquisition process parameters of the two scientific and technological intelligence perception information corresponding to the differences, identify the parameter items with different parameter values, and record the name and value range of the different parameter items.

[0039] Comparing the sensory data collection parameters for the comments "easy to use" and "powerful features," some discrepancies in parameter values ​​may be observed. For example, the sampling frequency of the collection devices might differ; one comment was collected when system A had a higher sampling frequency, while the other was collected when the sampling frequency was lower. The type of data preprocessing algorithm may also differ; one comment used a specific word segmentation algorithm, while the other used a different one. Records of interference factors in the sensory environment may also vary; for instance, the network was relatively stable when one comment was collected, while the network fluctuated when the other was collected. The adherence to operational procedures by the data collectors may also differ; for example, the annotations in one comment are accurate, while the annotations in the other are incorrect. Record the parameter items with differing values, including the parameter name and value range.

[0040] Step S124: Analyze the correlation between the difference parameter items and the perceptual bias. By calculating the correlation between the parameter difference degree and the semantic distance of the perceptual bias, identify the key difference parameter items that have a dominant influence on the perceptual bias.

[0041] Taking the sampling frequency of the data acquisition device as an example, the difference in sampling frequency is calculated, which is the difference between the sampling frequencies of two comments. Then, a correlation analysis is performed between the difference in sampling frequency and the semantic distance of perceptual bias. If a high correlation is found between the difference in sampling frequency and the semantic distance of perceptual bias, it indicates that the difference in sampling frequency may have a significant impact on perceptual bias. Similarly, similar analyses are performed on other difference parameters to identify the key difference parameters that have a dominant impact on perceptual bias.

[0042] Step S125: Based on the influence patterns of key difference parameters, generate a perceptual deviation source tracing report. The perceptual deviation source tracing report includes a list of key difference parameters, a description of the influence patterns of the parameters, and suggestions for deviation improvement.

[0043] Based on the analysis results of key difference parameters, a perceptual bias source tracing report is generated. This report includes a list of key difference parameters, outlining the parameters that have the dominant influence on perceptual bias. It also describes the influence patterns of each key difference parameter; for example, a higher sampling frequency may collect more comments, but it may also lead to inconsistent comment quality; different data preprocessing algorithms may affect the comment analysis results, etc. Finally, based on the influence patterns of the key difference parameters, suggestions for bias improvement are proposed, such as optimizing the sampling frequency of the data collection equipment to improve comment quality while maintaining the quantity of comments; and selecting more suitable data preprocessing algorithms to improve the accuracy of comment analysis, etc.

[0044] Step S126: Based on the deviation improvement direction suggestions in the perception deviation source tracing report, formulate cross-agent dynamic calibration rules for each perception dimension. The cross-agent dynamic calibration rules include calibration source selection criteria, calibration content generation algorithm, calibration weight allocation method, and calibration result verification criteria.

[0045] Taking product evaluation as an example, dynamic calibration rules across agents are formulated based on suggestions for deviation improvement. The selection criteria for calibration sources can include choosing highly credible and authoritative review sources, such as professional testing organizations or renowned technical experts. The calibration content generation algorithm can generate calibrated review content based on the review content from the calibration sources. The calibration weight allocation method can assign different weights to different calibration sources based on their credibility and authority. The calibration result verification standard can be achieved by comparing the review analysis results before and after calibration to evaluate whether the calibration effect has met expectations.

[0046] Step S127: Send the cross-agent dynamic calibration rules to the corresponding science and technology intelligence perception information processing carrier. The science and technology intelligence perception information processing carrier determines the calibration source for this perception dimension according to the calibration source selection standard and requests the corresponding standard perception content from the calibration source.

[0047] The established cross-agent dynamic calibration rules are sent to System A and other technology intelligence sensing information processing carriers. System A determines the calibration source for the product evaluation dimension based on the calibration source selection criteria, for example, selecting the comments of a professional evaluation organization as the calibration source. Then, System A requests the corresponding standard sensing content from the calibration source, that is, requests the comments of that professional evaluation organization on a certain technology product.

[0048] Step S128: According to the calibration content generation algorithm in the cross-agent dynamic calibration rules, the standard perception content provided by the calibration source is fused and optimized with the missing perception content to generate supplementary and optimized perception content.

[0049] System A uses a calibration content generation algorithm to fuse and optimize the standard perceived content provided by the calibration source with missing perceived content from its own collected reviews. For example, if a review from the calibration source mentions a product advantage that is not mentioned in System A's reviews, System A can add this advantage to its own review and apply appropriate weighting based on the calibration weighting method. This ultimately generates supplementary and optimized perceived content.

[0050] Step S129: Integrate the supplemented and optimized perception content with the initial science and technology intelligence perception information, remove duplicate perception content sub-items, and select conflicting perception content sub-items according to the calibration weight allocation method to generate calibrated and optimized science and technology intelligence perception information.

[0051] System A integrates the supplemented and optimized perceived content with the initially collected comments. During integration, duplicate perceived content items are removed; for example, if a comment from the calibration source already includes a benefit from a comment collected by System A, that benefit is removed from System A's comments. For conflicting perceived content items, a selection is made according to the calibration weighting method. For example, if the evaluation of a product feature in the comments from the calibration source differs from the evaluation in the comments collected by System A, the evaluation from the comment with higher credibility and authority is selected as the final evaluation, based on the calibration weighting method. This ultimately generates calibrated and optimized scientific and technological intelligence perceived information.

[0052] Step S130: Invoke the cross-agent joint mining function to extract multi-level related features from the calibrated and optimized science and technology intelligence perception set, and integrate them to form a multi-level related feature set.

[0053] Step S131: Send the calibrated and optimized science and technology intelligence perception set to the cross-agent joint mining carrier. The cross-agent joint mining carrier classifies all science and technology intelligence perception information in the calibrated and optimized science and technology intelligence perception set into perception dimensions and divides the science and technology intelligence perception information into multiple single-dimensional perception information subsets according to the perception dimension labels.

[0054] Taking sentiment analysis technology from natural language processing as an example, the calibrated and optimized set of perceived scientific and technological intelligence is sent to a cross-agent joint mining platform. The platform categorizes all perceived scientific and technological intelligence information in the set according to perception dimension labels, such as product evaluation, technical discussion, and event description. Perceived scientific and technological intelligence information belonging to the same perception dimension label is grouped into a single-dimensional subset. For example, all perceived scientific and technological intelligence information about product evaluation is grouped into a single-dimensional subset for the product evaluation dimension, all perceived scientific and technological intelligence information about technical discussion is grouped into a single-dimensional subset for the technical discussion dimension, and all perceived scientific and technological intelligence information about event description is grouped into a single-dimensional subset for the event description dimension.

[0055] Step S132: For each single-dimensional perception information subset, extract all perception content sub-items of scientific and technological intelligence perception information in the single-dimensional perception information subset, count the frequency of occurrence of the same perception content sub-items, and mark the perception content sub-items whose occurrence frequency reaches the preset frequency threshold as the core perception sub-items of that perception dimension.

[0056] Taking a single-dimensional perceived information subset of the product evaluation dimension as an example, we extract all perceived content items from the subset, including product advantages, disadvantages, and user satisfaction. Then, we count the frequency of each perceived content item. If the frequency of a perceived content item reaches a preset frequency threshold, it is marked as a core perceived item of the product evaluation dimension. For example, if the preset frequency threshold is 100 times, and the perceived content item "easy to operate" appears 120 times in the single-dimensional perceived information subset of the product evaluation dimension, it is marked as a core perceived item of the product evaluation dimension.

[0057] Step S133: Analyze the logical relationships between core perception items in the same single-dimensional perception information subset, identify core perception item pairs with causal or accompanying relationships, record the relationship type and the scenario conditions in which the relationship occurs, and generate single-dimensional relationship feature entries.

[0058] Within the single-dimensional perceived information subset of the product evaluation dimension, analyze the logical relationships between core perceived sub-items. For example, a causal relationship is found between the core perceived sub-items "easy to operate" and "high user satisfaction," meaning that easy operation may lead to high user satisfaction. Record the above association type as a causal relationship and the scenario conditions under which the association occurs, such as a specific product usage scenario. Generate a single-dimensional association feature entry, which includes the core perceived sub-item pair "easy to operate - high user satisfaction," the association type "causal relationship," and the scenario condition "a specific product usage scenario."

[0059] Step S134: After completing the extraction of all single-dimensional associated features, the different single-dimensional perceptual information subsets are combined in pairs to form cross-dimensional combination pairs. Each cross-dimensional combination pair contains two single-dimensional perceptual information subsets of different perceptual dimensions.

[0060] After extracting all single-dimensional related features, including product evaluation, technical discussion, and event description dimensions, different single-dimensional perceptual information subsets are paired. For example, the single-dimensional perceptual information subsets of the product evaluation dimension and the technical discussion dimension are combined into one cross-dimensional pair; the single-dimensional perceptual information subsets of the product evaluation dimension and the event description dimension are combined into another cross-dimensional pair; and the single-dimensional perceptual information subsets of the technical discussion dimension and the event description dimension are combined into a third cross-dimensional pair. Each cross-dimensional pair contains two single-dimensional perceptual information subsets from different perceptual dimensions.

[0061] Step S135: For each cross-dimensional combination pair, extract the core perception sub-items of all scientific and technological intelligence perception information in the two single-dimensional perception information subsets, compare the perception scene description content corresponding to the core perception sub-items, and select the core perception sub-item pairs whose perception scene description content overlaps with the preset overlap ratio.

[0062] Taking the cross-dimensional combination of product evaluation and technical discussion dimensions as an example, we extract the core perception items of all scientific and technological intelligence perception information in the single-dimensional perception information subset of the product evaluation dimension, such as "easy to operate" and "powerful functions," and the core perception items of all scientific and technological intelligence perception information in the single-dimensional perception information subset of the technical discussion dimension, such as "algorithm innovation" and "performance improvement." Then, we compare the perception scenario descriptions corresponding to these core perception items. For example, "easy to operate" corresponds to the product's usage scenario, and "algorithm innovation" corresponds to the technology's research and development scenario. If the overlap ratio of the perception scenario descriptions corresponding to two core perception items reaches a preset overlap ratio, these two core perception item pairs are selected. For example, if the preset overlap ratio is 50%, and more than 50% of the perception scenario descriptions corresponding to "easy to operate" and "algorithm innovation" are the same, then the core perception item pair "easy to operate - algorithm innovation" is selected.

[0063] Step S136: Analyze the relationship between the selected core perception sub-item pairs, determine the cross-dimensional association strength by calculating the semantic association degree of the sub-item content and the synchronization of perception time, and generate cross-perception dimension association feature entries. The cross-perception dimension association feature entries include association strength value, association scene overlap degree and association stability description.

[0064] For the selected core perception sub-item pair "Easy to Operate - Algorithm Innovation," the relationship between them is analyzed. The semantic relevance of the sub-item content is calculated, for example, using semantic similarity algorithms in natural language processing to calculate the similarity between the content of "Easy to Operate" and "Algorithm Innovation." Simultaneously, the synchronicity of their perception time is analyzed, i.e., whether the collection time of the scientific and technological intelligence perception information corresponding to these two sub-items is similar. Based on the semantic relevance and the synchronicity of perception time, the cross-dimensional association strength is determined. For example, if the semantic relevance is high and the perception time synchronicity is good, the cross-dimensional association strength is high. Cross-perception dimension association feature entries are generated, which include association strength values, association scene overlap, and association stability descriptions. For example, an association strength value of 0.8 indicates a high cross-dimensional association strength; an association scene overlap of 60% indicates that the overlap ratio of the perception scene descriptions corresponding to the two core perception sub-items is 60%; and the association stability description can be "the association relationship is relatively stable within a certain time period."

[0065] Step S137: Classify and organize single-dimensional related feature entries and cross-perception dimension related feature entries, sort the entries according to the priority of perception dimensions, and add related feature identifiers and corresponding perception information identifiers of scientific and technological intelligence perception information to each entry.

[0066] The entries with single-dimensional related features and those with cross-perceptual-dimensional related features are categorized and organized. Entries are then sorted according to the priority of the perceptual dimensions; for example, product evaluation may have a higher priority than technical discussion, and technical discussion may have a higher priority than event description. Related feature identifiers are added to each entry for easy identification and management. Simultaneously, corresponding perceptual information identifiers for scientific and technological intelligence are added to each entry for easy traceability and querying.

[0067] Step S138: Delete entries with duplicate content in single-dimensional and cross-perceptual dimension related feature entries, select entries with logical conflicts in single-dimensional and cross-perceptual dimension related feature entries according to the correlation strength value, and integrate the remaining entries to form a multi-level related feature set.

[0068] The single-dimensional and cross-perceptual-dimensional related feature entries, after being categorized and organized, are checked, and duplicate entries are removed. For entries with logical conflicts, such as two cross-dimensional related feature entries describing the relationship between the same pair of core perceptual sub-items differently, they are selected based on their association strength value. Entries with higher association strength values ​​are retained, while those with lower values ​​are deleted. Finally, the remaining entries are integrated to form a multi-level related feature set.

[0069] Step S140: Generate initial decision suggestions based on the multi-level associated feature set, send the initial decision suggestions to other science and technology intelligence perception information processing carriers, receive and analyze the initial decision suggestions from other science and technology intelligence perception information processing carriers, identify conflict points in the decision suggestions, locate the corresponding missing perception information items according to the conflict points, supplement, optimize and calibrate the optimized science and technology intelligence perception set, and generate adjusted decision suggestions.

[0070] Step S141: Extract related feature entries from the multi-level related feature set that are related to the preset decision scope. The preset decision scope is determined based on the preset responsibility area. When extracting, priority is given to selecting entries whose related strength value reaches the preset related strength standard.

[0071] In the context of sentiment analysis technology within natural language processing, the pre-defined responsibility area might be the product development department, and its pre-defined decision-making scope might be about the direction of product improvement. The process involves extracting relevant feature items related to the product improvement direction from a multi-level set of related features. For example, extracting feature items related to product functionality, performance, and user experience. During extraction, priority is given to items with a correlation strength value that meets a pre-defined correlation strength standard. This pre-defined correlation strength standard can be set according to actual conditions, for example, to 0.7. Only feature items with a correlation strength value greater than or equal to 0.7 will be extracted.

[0072] Step S142: Map the extracted associated feature items to a decision direction, and convert the association relationship corresponding to each associated feature item into a specific decision direction suggestion. The decision direction suggestion includes the decision target type, decision implementation priority, and decision risk warning.

[0073] Taking the associated characteristic item "easy to operate - high user satisfaction" as an example, its corresponding association is converted into specific decision-making direction suggestions. The decision goal type can be improving user satisfaction, and the decision implementation priority can be determined based on the association strength value; for example, decisions with higher association strength values ​​have higher implementation priority. Decision risk warnings can include potential technical challenges, increased costs, and other risks that may be encountered in the process of improving ease of operation.

[0074] Step S143: Construct an initial decision suggestion framework based on the decision direction suggestion. The initial decision suggestion framework includes a decision goal module, a decision basis module, a decision implementation path module, and a decision risk response module. The content of the decision basis module is generated based on the core perception sub-items of the extracted related feature entries.

[0075] An initial decision-making suggestion framework is constructed based on the suggested decision-making direction. The decision-making objective module contains specific decision-making objectives, such as improving user satisfaction. The decision-making basis module is generated based on the core perception sub-items of extracted related feature entries. For example, the decision-making basis module could include the core perception sub-item "easy to operate," indicating that the basis for improving user satisfaction is the correlation between ease of operation and user satisfaction. The decision implementation path module contains the specific steps and methods for achieving the decision-making objectives. For example, improving ease of operation can be achieved by optimizing the product's interface design and simplifying the operation process, thereby increasing user satisfaction. The decision risk management module contains measures to address decision-making risks. For example, proactive technological research and development and reasonable cost control can be used to address potential technical challenges and increased costs encountered during the process of improving ease of operation.

[0076] Step S144: Fill in the specific content of the initial decision recommendation framework, add a detailed description to each module, the decision goal module describes the specific results that the decision needs to achieve, the decision implementation path module contains the phased operation steps, and the decision risk response module contains the response measures for potential risks.

[0077] Each module of the initial decision-making framework is described in detail. The Decision Objectives module describes the specific results the decision aims to achieve, such as increasing user satisfaction to over 90%. The Decision Implementation Path module includes phased operational steps; for example, the first phase involves market research to understand user needs and expectations regarding product ease of use; the second phase optimizes the product interface design and simplifies the operation process; and the third phase involves product testing and evaluation, collecting user feedback, and making adjustments and improvements based on that feedback. The Decision Risk Management module includes measures to address potential risks; for example, to address technical challenges, a professional technical team can be established to strengthen technological research and development and innovation; to address increased costs, costs can be controlled through optimizing resource allocation and reducing procurement costs.

[0078] Step S145: Add processing carrier identifier, generation time, and a list of related feature entries to the initial decision suggestion to form the initial decision suggestion.

[0079] Add a processing carrier identifier to the initial decision recommendation, indicating which science and technology intelligence sensing information processing carrier generated it. Add the generation time to understand the timeliness of the decision recommendation. Add a list of related feature entries, listing the related feature entries on which the decision recommendation is based, facilitating subsequent tracing and querying. Finally, a complete initial decision recommendation is formed.

[0080] Step S146: Send the initial decision suggestion to all other science and technology intelligence perception information processing carriers through a preset communication channel, and at the same time receive the initial decision suggestion sent by other science and technology intelligence perception information processing carriers, which includes the processing carrier identifier, generation time and a list of related feature entries.

[0081] Initial decision recommendations are sent to all other technology intelligence sensing and information processing carriers via pre-defined communication channels, such as network communication protocols. Simultaneously, initial decision recommendations from other technology intelligence sensing and information processing carriers, containing carrier identifiers, generation times, and lists of related characteristic entries, are received. For example, initial decision recommendations generated by the product development department are sent to other departments such as marketing and customer service, while initial decision recommendations from these departments are also received.

[0082] Step S147: Compare the content of all received initial decision suggestions, identify content with completely contradictory statements or conflicting steps, and mark them as points of conflict in decision suggestions.

[0083] Step S1471: Classify all received initial decision suggestions according to decision domain, group initial decision suggestions whose decision target modules belong to the same decision domain into the same decision suggestion group, and assign a decision domain identifier to each decision suggestion group.

[0084] In the context of sentiment analysis technology in natural language processing, all initial decision suggestions are categorized according to different decision domains. For example, decision suggestions related to product feature improvement are grouped into the product feature improvement decision suggestion group and assigned the decision domain label "product feature improvement"; decision suggestions related to marketing strategies are grouped into the marketing promotion decision suggestion group and assigned the decision domain label "marketing promotion", and so on.

[0085] Step S1472: For each decision suggestion group, extract the core description of the decision objective module of all initial decision suggestions in the decision suggestion group. The core description includes the core results that the decision needs to achieve and key evaluation indicators.

[0086] Taking the product feature improvement decision suggestion group as an example, we extract the core statements of the decision objective module for all decision suggestions within the group. For instance, the core statement of the decision objective module for a certain decision suggestion is "Add new features to the product, so that the user usage rate of the new features reaches more than 50%". Here, "Add new features to the product" is the core result that the decision needs to achieve, and "the user usage rate of the new features reaches more than 50%" is the key evaluation indicator.

[0087] Step S1473: Compare the core statements of decision objectives of different initial decision suggestions within the decision suggestion group, calculate the semantic similarity between the statements, and mark the core statement results as conflict points if the semantic similarity is lower than the preset semantic similarity threshold and the core statement results are opposite. Record the processing carrier identifiers and core statement contents of the conflicting parties.

[0088] In the product feature improvement decision suggestion group, the core statements of the decision objectives of different decision suggestions are compared. Semantic similarity calculation methods from natural language processing, such as the cosine similarity algorithm, are used to calculate the similarity between the statements. The preset semantic similarity threshold is 0.6. If the semantic similarity of the core statements of the decision objectives of two decision suggestions is less than 0.6, and the core results are opposite—for example, one decision suggestion is "add new features to the product," and the other is "reduce existing features of the product"—then it is marked as a point of conflict in decision objectives. The processing entities of both parties in the conflict are also recorded; for example, the processing entity for the decision suggestion to "add new features to the product" is the product development department, and the processing entity for the decision suggestion to "reduce existing features of the product" is the marketing department, along with the core statements of both parties in the conflict.

[0089] Step S1474: After completing the comparison of the decision target module, extract the step sequence of the decision implementation path module of all initial decision suggestions in the decision suggestion group. The step sequence includes phased operation steps and the execution order of the steps.

[0090] After comparing the decision target modules of the product feature improvement decision suggestion group, the step sequence of the decision implementation path module for all decision suggestions within the group is extracted. For example, the step sequence of the decision implementation path module for a certain decision suggestion is "market research, feature design, technology development, product testing, and online promotion", where "market research" is the first stage operation step, "feature design" is the second stage operation step, and so on, and each step has its specific execution order.

[0091] Step S1475: Compare the step sequences of different initial decision suggestions. First, count the consistency of the number of steps, and then check the consistency of the operation content at the same step position. If there is a difference in the number of steps and key operation steps are missing, or the operation content of steps at the same position is completely opposite, mark it as a conflict point in the decision implementation path.

[0092] Within the product feature improvement decision recommendation group, the step sequences of different decision recommendations are compared. First, the number of steps is checked for consistency. If there is a difference in the number of steps, and a key operational step is missing—for example, one decision recommendation includes the crucial step of "user feedback collection," while another does not—it is marked as a conflict point in the decision implementation path. If the number of steps is the same, the operational content at the same step position is checked for consistency. If the operational content at the same position is completely opposite—for example, in the "feature design" step, one decision recommendation is "design a simple and easy-to-use functional interface," while another is "design a complex but powerful functional interface"—this is also marked as a conflict point in the decision implementation path.

[0093] Step S1476: Assign unified numbers to the conflict points of decision objectives and conflict points of decision implementation paths, and record the decision suggestion group identifier, conflict type, and initial decision suggestion reference information of both parties to the conflict for each conflict point.

[0094] The marked conflict points in the decision-making objectives and conflict points in the decision-making implementation paths are uniformly numbered. For example, the conflict point in the decision-making objective is numbered "Conflict Point 1", and the conflict point in the decision-making implementation path is numbered "Conflict Point 2". The decision suggestion group identifier corresponding to each conflict point is recorded. For example, the decision suggestion group identifier corresponding to "Conflict Point 1" is "Product Function Improvement". The conflict type is recorded. For example, the conflict type of "Conflict Point 1" is "Decision-Making Objective Conflict", and the conflict type of "Conflict Point 2" is "Decision-Making Implementation Path Conflict". The initial decision suggestion reference information of both parties in the conflict is recorded. For example, the initial decision suggestion reference information of both parties in the conflict of "Conflict Point 1" is "Decision suggestion from the product development department to add new product functions" and "Decision suggestion from the marketing department to reduce existing product functions".

[0095] Step S1477: Mark the content with opposite risk assessment results in the decision risk response module as decision risk response conflict points, and add decision risk response conflict points to the decision suggestion conflict point list.

[0096] In the decision risk response module of each decision recommendation, check whether there are contradictory risk assessment results. For example, one decision recommendation may believe that product feature improvement may face technical difficulties that could lead to delays in the development schedule, while another decision recommendation believes that, under the current level of technology, product feature improvement does not face any technical difficulties that could lead to delays in the development schedule. Mark the contradictory risk assessment results as decision risk response conflict points and add them to the decision recommendation conflict point list.

[0097] Step S1478: Determine the severity of the conflict according to the scope of its impact on the overall decision-making process, sort all marked conflict points according to their severity, and form the final list of conflict points for decision recommendations.

[0098] The severity of conflict is determined by assessing its impact on the overall decision-making process. For example, if a conflict point involves the core functionality and market positioning of a product, its severity is considered high; conversely, if a conflict point in the implementation path only involves differences in some non-critical steps, its severity is relatively low. All marked conflict points are then ranked from highest to lowest severity to create a final list of conflict points for decision recommendations, enabling targeted conflict resolution and decision adjustments in the future.

[0099] Step S148: For each point of conflict in decision recommendations, trace the list of related feature entries corresponding to the decision recommendations of both parties in the conflict, analyze the differences in the referenced entries, and determine whether the conflict stems from the absence of any related feature entry or the difference in the correlation strength value of the referenced entries.

[0100] Step S1481: Extract the reference list of related feature entries of the initial decision suggestions of the conflicting parties, compare the related feature identifiers in the two lists, identify the related feature identifiers that exist only in one party's list, and mark them as potentially missing related feature identifiers.

[0101] Taking a conflicting decision suggestion within a product feature improvement decision suggestion group as an example, let's assume the conflicting parties are the product development department and the marketing department, respectively. We extract the reference list of related feature items from the product development department's decision suggestion, assuming it includes related feature markers such as "user needs - feature improvement direction" and "technical feasibility - feature implementation difficulty." We also extract the reference list of related feature items from the marketing department's decision suggestion, assuming it includes related feature markers such as "market competition - feature differentiation needs" and "user feedback - feature satisfaction." Comparing the related feature markers in these two lists, we find that the related feature marker "user needs - feature improvement direction" only exists in the product development department's decision suggestion reference list; therefore, we mark it as a potentially missing related feature marker.

[0102] Step S1482: Based on the potential missing association feature identifier, query the corresponding association feature entry from the multi-level association feature set, and obtain the perception dimension label and core perception sub-item description of the association feature entry.

[0103] Based on the potentially missing associated feature identifier "User Needs - Functional Improvement Direction", the corresponding associated feature item is queried from the multi-level associated feature set. The query shows that the perception dimension label of this associated feature item is "User Needs Dimension", and the core perception sub-item description is "Users' specific needs and expectations for product functions".

[0104] Step S1483: Analyze the role of the associated feature item in the conflict decision-making recommendations. By reviewing the references to the associated feature item in the decision basis module, determine whether the associated feature item is a key basis for supporting the decision conclusion of the conflicting parties.

[0105] By examining the references to the "User Needs - Function Improvement Direction" related feature item in the Decision Basis module of the Product Development Department's decision-making recommendations, it was found that the Decision Basis module mentions "Based on user needs analysis, the direction for improving product functions is determined to be adding new functions to meet users' diverse needs for product functions." This indicates that the related feature item is a key basis for supporting the product development department's decision-making conclusions.

[0106] Step S1484: If the associated feature entry is a key basis and is only cited by one party, it is determined that the conflict stems from the other party's failure to obtain the associated feature entry, i.e., there is a missing associated feature entry.

[0107] Since the "User Needs - Function Improvement Direction" related feature item is a key basis for supporting the product development department's decision-making conclusions and is only cited by the product development department, while the marketing department's decision-making suggestions do not cite this item, the conflict of judgment stems from the marketing department's failure to obtain this related feature item, that is, there is a missing related feature item.

[0108] Step S1485: Based on the perception dimension label of the associated feature item, determine the corresponding perception dimension, and then based on the core perception sub-item description, determine the specific core perception sub-item under the perception dimension.

[0109] Based on the perception dimension label "User Needs Dimension" of the associated feature item "User Needs - Function Improvement Direction", the corresponding perception dimension is determined to be the User Needs Dimension. Then, based on the core perception sub-item description "Specific needs and expectations of users for product functions", the specific core perception sub-items under this perception dimension are determined to be the specific needs and expectations of users for product functions.

[0110] Step S1486: Query whether the core perception sub-item under the perception dimension is included in the self-calibrated and optimized science and technology intelligence perception set. If it is not included, record the core perception sub-item and the corresponding perception dimension as a missing perception information item.

[0111] A review of the calibrated and optimized technology intelligence perception set of the marketing department revealed that it did not include the core perception sub-item "users' specific needs and expectations for product functions" under the user needs dimension. Therefore, this core perception sub-item and its corresponding perception dimension, user needs dimension, were recorded as missing perception information items.

[0112] Step S1487: If the core perception sub-item is included in the self-calibrated and optimized science and technology intelligence perception set but is not extracted as an associated feature item, analyze the reason for not extracting it. If it is because the association strength value does not reach the preset extraction strength threshold, then record the core perception sub-item and the corresponding association strength calculation parameters as perception information optimization items, and include the perception information optimization items in the category of perception information missing items.

[0113] Suppose, in another scenario, the calibrated and optimized STEM intelligence perception set of the marketing department contains a core perception sub-item under the user needs dimension: "users' specific needs and expectations for product functions." However, this sub-item is not extracted as a related feature item. Further analysis reveals that the reason for this failure to extract it is that the correlation strength between this core perception sub-item and other perception sub-items does not reach the preset extraction strength threshold. In this case, the core perception sub-item and its corresponding correlation strength calculation parameters are recorded as perception information optimization items, and these optimization items are included in the category of missing perception information items.

[0114] Step S1488: Organize all identified missing and optimized perception information items into a perception supplementation requirement list, and record the perception dimension, core perception sub-item content, and supplementation priority for each missing item.

[0115] All identified missing and optimized perceptual information items are compiled into a perceptual supplementation requirement list. For example, the perceptual supplementation requirement list records the marketing department's missing perceptual information item "Users' specific needs and expectations for product functions (user needs dimension)", which has a high priority for supplementation; it also records the perceptual information optimization item "Users' specific needs and expectations for product functions (user needs dimension, correlation strength calculation parameters)", which has a medium priority for supplementation.

[0116] Step S149: If the conflict originates from the absence of associated feature entries, locate the perceptual dimension and core perceptual sub-item corresponding to the missing entry and determine it as a missing perceptual information item.

[0117] If the conflict stems from the absence of related feature items—for example, a related feature item is cited in one decision suggestion but not in another—then it's necessary to locate the perceptual dimension and core perceptual sub-item corresponding to the missing item. For instance, a decision suggestion to increase product prices cites the related feature item "product scarcity - price increase," while a decision suggestion to decrease product prices does not. Therefore, it's necessary to locate the perceptual dimension and core perceptual sub-item corresponding to this "product scarcity - price increase" related feature item. By querying the multi-level related feature set, it can be determined that the perceptual dimension corresponding to the "product scarcity - price increase" related feature item is the market demand dimension, and the core perceptual sub-item is product scarcity. Product scarcity is thus identified as the missing perceptual information item.

[0118] Step S150: Summarize all adjusted decision recommendations, resolve conflicts and integrate content among the adjusted decision recommendations, and generate the final collaborative decision solution.

[0119] Step S151: Establish a collaborative decision-making fusion platform, send the adjusted decision suggestions from all science and technology intelligence perception information processing carriers to the collaborative decision-making fusion platform, and the collaborative decision-making fusion platform performs format standardization processing on the received adjusted decision suggestions to ensure that the module structure of each adjusted decision suggestion is consistent.

[0120] In the context of sentiment analysis technology within natural language processing, a collaborative decision-making fusion platform is needed to achieve collaborative decision-making among various technology intelligence perception and information processing carriers. This platform is responsible for receiving adjusted decision suggestions from all technology intelligence perception and information processing carriers and standardizing their format. For example, it unifies the structure and content format of the decision objective module, decision basis module, decision implementation path module, and decision risk response module in the decision suggestions generated by different processing carriers, facilitating subsequent conflict resolution and content fusion.

[0121] Step S152: Classify the adjusted decision suggestions according to the decision domain, group the adjusted decision suggestions of the same decision domain into the same fusion group, and assign a fusion group identifier and decision domain label to each fusion group.

[0122] Based on the decision objective module in the adjusted decision recommendations, all recommendations are categorized according to decision domain. For example, decision recommendations related to product development are grouped into the product development domain integration group, those related to marketing into the marketing domain integration group, and those related to customer service into the customer service domain integration group, etc. A unique integration group identifier is assigned to each integration group for easy identification and management. Simultaneously, a decision domain label is added to each integration group to clearly identify the decision domain to which that integration group belongs.

[0123] Step S153: For each fusion group, extract the core expressions of the decision target module of all adjusted decision suggestions within the fusion group, cluster them by calculating semantic similarity, classify the core expressions with semantic similarity higher than the preset clustering similarity threshold into the same target cluster, and select the core expression with the highest frequency in the target cluster as the unified decision target of the fusion group.

[0124] Step S1531: Extract the content of the decision objective module of each adjusted decision suggestion, remove redundant descriptive statements in the decision objective module, and retain the statements containing the core decision results and key evaluation indicators as the core statement of the decision objective.

[0125] Taking the integration group in the product development field as an example, for an adjusted decision recommendation within the group, the decision objective module content is "to improve product performance and quality by improving product technology and design to meet market demands and user expectations, thereby increasing product market share and user satisfaction." After removing redundant descriptive statements, "improving product performance and increasing market share" is retained as the core statement of the decision objective. "Improving product performance" is the core decision outcome, and "increasing market share" is the key evaluation indicator.

[0126] Step S1532: Perform text preprocessing on all core statements of decision objectives, including word segmentation, stop word removal, and semantic feature vector conversion, converting each core statement of decision objectives into a fixed-dimensional semantic feature vector.

[0127] Text preprocessing is performed on the core statements of all decision objectives within the product development integration group. First, word segmentation is performed; for example, "improve product performance" is segmented into "improve," "product," and "performance." Then, stop words such as "of" and "and" are removed. Next, semantic feature vector transformation is performed using word vector models from natural language processing, such as Word2Vec, to convert each core statement of decision objectives into a fixed-dimensional semantic feature vector. For example, "improve product performance" is converted into a 100-dimensional semantic feature vector.

[0128] Step S1533: Use a clustering algorithm to cluster the semantic feature vectors, calculate the cosine similarity between any two semantic feature vectors, and group vectors with a cosine similarity higher than the preset clustering similarity threshold into the same cluster to form multiple target clusters.

[0129] Clustering algorithms, such as K-means, are used to cluster the semantic feature vectors transformed from the core expressions of decision-making objectives within the product development integration group. The cosine similarity between any two semantic feature vectors is calculated, with a preset clustering similarity threshold of 0.7. For example, if there are three semantic feature vectors transformed from the core expressions of decision-making objectives, vector A, vector B, and vector C, and the cosine similarity between vector A and vector B is 0.8, between vector A and vector C is 0.6, and between vector B and vector C is 0.75, then vector A and vector B are grouped into the same cluster, forming one target cluster; vector C forms a separate target cluster.

[0130] Step S1534: Count the number of core statements of decision objectives contained in each target cluster, and determine the target cluster with the most statements as the main target cluster.

[0131] In the product development field integration group, the number of core decision-making objective statements contained in each target cluster is counted. Assuming that, after counting, one target cluster contains 5 core decision-making objective statements and another target cluster contains 3 core decision-making objective statements, then the target cluster containing 5 statements is determined as the primary target cluster.

[0132] Step S1535: If there are multiple main target clusters containing the same number of expressions, calculate the average vector of all semantic feature vectors in each main target cluster, compare the average vector with the preset core target vector of the decision domain, and select the target cluster corresponding to the average vector with the highest similarity as the final main target cluster.

[0133] If there are two primary target clusters with the same number of statements within the product development domain fusion group, for example, both clusters contain four core decision-making target statements, calculate the average vector of all semantic feature vectors within each primary target cluster. Assuming the semantic feature vectors in the first primary target cluster are vectors D, E, F, and G, calculate their average vector as vector M; and the semantic feature vectors in the second primary target cluster are vectors H, I, J, and K, calculate their average vector as vector N. Presuppose the core target vector in the product development domain is vector O. Calculate the similarity between vector M and vector O, and the similarity between vector N and vector O. Assuming the similarity between vector M and vector O is 0.9, and the similarity between vector N and vector O is 0.8, then select the target cluster corresponding to vector M as the final primary target cluster.

[0134] Step S1536: Select the core expression of the decision target with the highest frequency from the final main target cluster. If there are multiple expressions with the same frequency, calculate the similarity between the multiple expressions and the average vector of the cluster, and select the expression with the highest similarity as the unified decision target of the fusion group.

[0135] In the final identified primary target clusters, the core statement of the decision objective that appears most frequently is selected as the unified decision objective for the product development domain integration group. Assuming that the core statement of the decision objective "improving product performance" appears 3 times and the core statement of the decision objective "enhancing product functionality" also appears 3 times in the primary target clusters, the similarity between these two statements and the average vector of the primary target clusters is calculated. Assuming that the similarity between "improving product performance" and the average vector is 0.95, and the similarity between "enhancing product functionality" and the average vector is 0.9, then "improving product performance" is selected as the unified decision objective for this integration group.

[0136] Step S1537: Supplement the core requirement descriptions of the decision domain that are missing in the unified decision objective, extract relevant descriptions from other objective clusters and add them to the unified decision objective to form a unified decision objective description.

[0137] After determining "improving product performance" as the unified decision-making objective for the product development domain integration group, it's necessary to check whether the unified decision-making objective is missing any core requirement statements for this decision domain. For example, the core requirement for product development might also include "reducing product costs," but this statement is not included in the unified decision-making objective. In this case, relevant statements are extracted from other objective clusters. Assuming that the statement "optimize product manufacturing processes and reduce product costs" exists in another objective cluster, it is added to the unified decision-making objective, forming the unified decision-making objective description "improve product performance, optimize product manufacturing processes, and reduce product costs."

[0138] Step S154: Extract the decision basis module content of all adjusted decision recommendations within the fusion group, integrate the same or similar decision basis, remove duplicate statements, and supplement and integrate the basis content that is different but not conflicting to form a unified decision basis for the fusion group.

[0139] Within the product development integration group, the decision-making basis module content of all adjusted decision recommendations is extracted. For example, some decision recommendations are based on "market demand analysis shows that users have high requirements for product performance," while others are based on "technological development trends show that product functions need continuous enhancement." These decision bases are then integrated, merging identical or similar bases and removing duplicate statements. For example, the similar bases "market demand analysis shows that users have high requirements for product performance" and "user survey results show that product performance is a key focus for users" are merged into "both market demand analysis and user survey results show that users have high requirements for product performance." For bases that differ but do not conflict, supplementary integration is performed. For example, one decision recommendation is based on "technical feasibility study shows that improving product performance is feasible," while another is based on "cost-benefit analysis shows that enhancing product functions is cost-acceptable." These two bases differ but do not conflict; they can be supplemented and integrated into "technical feasibility study shows that improving product performance is feasible, and cost-benefit analysis shows that enhancing product functions is cost-acceptable." This ultimately forms the unified decision-making basis for the product development integration group.

[0140] Step S155: Analyze the decision implementation path module of the adjusted decision recommendations within the fusion group, select the implementation path with the most complete step logic and the most detailed operation description as the basic path, and then integrate the reasonable supplementary steps from other adjusted decision recommendations to optimize and form the unified implementation path of the fusion group.

[0141] Taking the product development integration group as an example, we analyze the decision implementation path modules of all adjusted decision recommendations within the group. Each decision implementation path module may contain different steps and operational descriptions. The implementation path with the most complete logical steps and the most detailed operational descriptions is selected as the base path. For example, a decision implementation path module for a certain adjusted decision recommendation may include steps such as product requirements analysis, technical solution design, product development, and product testing, with detailed operational descriptions for each step. Other decision implementation path modules may have incomplete steps or insufficient operational descriptions; therefore, this decision implementation path is selected as the base path. Then, we integrate reasonable supplementary steps from other adjusted decision recommendations. For example, other decision recommendations may propose adding a user feedback process during product development or adopting stricter testing standards during product testing. These supplementary steps are integrated into the base path to optimize and form a unified implementation path for the product development integration group.

[0142] Step S156: Summarize the decision risk response modules of the adjusted decision recommendations within the fusion group, integrate the risk identification results and response measures from different adjusted decision recommendations, and form a unified risk response plan for the fusion group.

[0143] Within the product development integration group, a risk response module is created to summarize all adjusted decision recommendations. Each risk response module may contain different risk identification results and corresponding measures. For example, some recommendations identify technical challenges that could delay product development, with the response being the formation of a dedicated technical team to tackle these challenges. Other recommendations identify changes in market demand that could lead to products failing to meet market needs, with the response being enhanced market research and user feedback. The risk identification results and corresponding measures are then integrated, merging similar or identical risk identification results from different recommendations and comprehensively considering different measures. For instance, the risk identification results of technical challenges and changes in market demand are merged into "Technology and Market Risk," and measures such as forming a dedicated technical team and enhancing market research and user feedback are combined to create a unified risk response plan for the product development integration group.

[0144] Step S157: Integrate the unified decision-making objectives, unified decision-making basis, unified implementation path and unified risk response plan of each fusion group to form a single-domain decision-making sub-plan.

[0145] The unified decision-making objective of the product development integration group—"improving product performance"—the unified decision-making basis—"Market demand analysis and user survey results both indicate that users have high requirements for product performance; technical feasibility studies show that improving product performance is feasible; cost-benefit analysis shows that enhancing product functionality is cost-acceptable"—the unified implementation path—"Product demand analysis, technical solution design, product development, adding user feedback links, product testing, and adopting more stringent testing standards"—and the unified risk response plan—"Technological and market risks, with countermeasures including establishing a professional technical team and strengthening market research and user feedback"—are integrated to form a single-domain decision-making sub-plan for the product development field. Similarly, following the same method, relevant content from other integration groups, such as marketing and customer service, is integrated to form their respective single-domain decision-making sub-plans.

[0146] Step S158: Sort all single-domain decision sub-schemes according to the importance priority of the decision domain. The importance priority is determined based on the contribution of the domain decision to the overall collaborative goal.

[0147] Determine the importance priority of each decision domain. Importance priority can be determined based on the contribution of decisions in that domain to the overall collaborative goal. For example, in the context of sentiment analysis in natural language processing, decisions in product development have a direct impact on product quality and competitiveness, contributing significantly to the overall collaborative goal; therefore, product development might have a higher importance priority. While customer service decisions are also important, their contribution to the overall collaborative goal is relatively smaller; therefore, customer service might have a relatively lower importance priority. Based on the determined importance priority, rank all single-domain decision sub-solutions. For example, rank product development sub-solutions first, and customer service sub-solutions later.

[0148] Step S159: Analyze the relationships between different single-domain decision sub-solutions, identify sub-solutions with dependencies, and adjust the order of sub-solutions to ensure logical coherence in the implementation of sub-solutions.

[0149] For example, step S1591: Extract the initiation condition description and implementation result description from the unified implementation path module of each single domain decision sub-scheme. The initiation condition description includes the prerequisite operations to be completed or the prerequisite data to be obtained before the implementation of the sub-scheme, and the implementation result description includes the output result data or the achieved state after the implementation of the sub-scheme.

[0150] Taking a single-domain decision-making sub-solution in the product development field as an example, the initiation conditions in its unified implementation path module might include "completing market research and obtaining user demand data" and "determining technically feasible solutions," while the implementation results might include "developing a product prototype with new functions" and "the product performance indicators meeting expected requirements." Similarly, for a single-domain decision-making sub-solution in the marketing field, the initiation conditions might include "product development completed, with a product ready for promotion," while the implementation results might include "increasing product market share by 10%."

[0151] Step S1592: Construct a sub-solution association matrix. The rows and columns of the sub-solution association matrix correspond to the identifiers of single-domain decision sub-solutions. The matrix elements represent the association type between the sub-solutions corresponding to the rows and the sub-solutions corresponding to the columns.

[0152] Construct a sub-solution relationship matrix, where the rows and columns correspond to the identifiers of decision-making sub-solutions in single domains such as product development, marketing, and customer service. For example, the first row and first column correspond to sub-solutions in the product development domain, the second row and second column to sub-solutions in the marketing domain, and the third row and third column to sub-solutions in the customer service domain. Matrix elements indicate the relationship type between the sub-solutions in the rows and columns, which can be "outcome dependency," "process dependency," or "no dependency."

[0153] Step S1593: Traverse the initiation condition descriptions of all sub-solutions. If the initiation condition of any sub-solution contains the implementation result of another sub-solution, then mark the former sub-solution as dependent on the latter sub-solution in the sub-solution association matrix, with the association type being result dependency.

[0154] The initiation condition descriptions for all single-domain decision sub-options are iterated. For example, the initiation condition for the marketing sub-option is described as "product development completed, with a marketable product," while the implementation result for the product development sub-option is described as "development of a product prototype with new functions." This indicates that the initiation condition for the marketing sub-option includes the implementation result for the product development sub-option. Therefore, in the sub-option relationship matrix, the matrix element at the intersection of the row corresponding to the marketing sub-option and the column corresponding to the product development sub-option is marked as "outcome dependency," indicating that the marketing sub-option depends on the product development sub-option.

[0155] Step S1594: If the initiation condition of any sub-scheme contains intermediate data in the implementation process of another sub-scheme, then mark the former sub-scheme as dependent on the latter sub-scheme, and the association type is process dependency.

[0156] Suppose the initiation condition for the customer service sub-solution is described as "obtaining user feedback data on the product in the market," while the marketing sub-solution collects user feedback data on the product in the market as intermediate data during its implementation. Then, in the sub-solution relationship matrix, the matrix element at the intersection of the row corresponding to the customer service sub-solution and the column corresponding to the marketing sub-solution is marked as "process dependent," indicating that the customer service sub-solution depends on the marketing sub-solution.

[0157] Step S1595: After completing the marking of all relationships, generate a sub-solution dependency graph. In the sub-solution dependency graph, the nodes are sub-solution identifiers, and the directed edges represent dependencies, pointing from the dependent sub-solution to the dependent sub-solution.

[0158] Based on the sub-solution association matrix, a sub-solution dependency graph is generated. In the sub-solution dependency graph, the identifier of each single-domain decision sub-solution is used as a node; for example, the product development sub-solution, marketing sub-solution, and customer service sub-solution are each represented by three nodes. Directed edges represent dependencies, pointing from the dependent sub-solution to the dependent sub-solution. For example, since the marketing sub-solution depends on the product development sub-solution, there is a directed edge in the sub-solution dependency graph pointing from the product development sub-solution node to the marketing sub-solution node; similarly, since the customer service sub-solution depends on the marketing sub-solution, there is a directed edge pointing from the marketing sub-solution node to the customer service sub-solution node.

[0159] Step S1596: Perform topological sorting based on the sub-solution dependency graph. First, sort the sub-solutions without prerequisite dependencies, then sort the sub-solutions that only depend on the sorted sub-solutions, and so on, to form a preliminary sorting result.

[0160] A topological sort is performed based on the sub-solution dependency graph. First, sub-solutions without any predecessor dependencies (i.e., no directed edges pointing to them) are identified and ranked first. In the example above, the product development sub-solution has no predecessor dependencies, so it is ranked first. Next, sub-solutions that depend only on the ranked sub-solutions are ranked. The marketing sub-solution depends only on the ranked product development sub-solution, so it is ranked second. Finally, the customer service sub-solution depends on the ranked marketing sub-solution, so it is ranked third. The preliminary ranking is: product development sub-solution, marketing sub-solution, and customer service sub-solution.

[0161] Step S1597: Analyze whether there is a circular dependency in the preliminary sorting results. If so, break down the implementation steps of the relevant sub-solutions, adjust the starting conditions, break the circular dependency, and then re-sort.

[0162] Analyze the preliminary ranking results to identify circular dependencies. A circular dependency occurs when a directed cycle exists in the sub-solution dependency graph, meaning that starting from a sub-solution and following a directed edge leads back to that sub-solution itself. For example, if the initiation condition of a marketing sub-solution also includes the implementation result of a customer service sub-solution, and the implementation result of the customer service sub-solution depends on the implementation of the marketing sub-solution, then a circular dependency exists. If a circular dependency exists, the implementation steps of the related sub-solutions need to be broken down, and the initiation conditions adjusted to break the circular dependency. For example, the implementation steps of the marketing sub-solution can be broken down into two stages: the first stage involves market promotion to obtain user feedback data, which is then provided to the customer service sub-solution; the second stage uses the feedback from the customer service sub-solution to further optimize the market promotion strategy. Adjust the initiation conditions of the marketing sub-solution so that it only depends on the implementation result of the product development sub-solution, i.e., the completion of product development. After breaking the circular dependency, perform a new topological sort to obtain a new ranking result.

[0163] Step S1598: Verify the implementation logic of the sub-schemes in the sorting results to ensure that the implementation results of the preceding sub-schemes can meet the activation conditions of the subsequent sub-schemes. If there are any unmet conditions, adjust the sorting position of the sub-schemes or supplement intermediate connecting steps.

[0164] Verify the implementation logic of the sub-solutions in the sorting results. Check whether the implementation results of the preceding sub-solutions can meet the initiation conditions of the subsequent sub-solutions. For example, the implementation result of the product development sub-solution is the development of a product prototype with new functions, while the initiation condition of the marketing sub-solution is the completion of product development and the availability of a marketable product. Verify whether the implementation result of the product development sub-solution meets the initiation conditions of the marketing sub-solution, that is, check whether the developed product prototype meets the conditions for marketability. If there are any discrepancies, such as the product prototype having new functions but having problems with stability and reliability, thus not meeting the conditions for marketability, then the sorting position of the sub-solutions needs to be adjusted. This could involve advancing the subsequent steps of the product development sub-solution or adding intermediate connecting steps, such as adding product stability and reliability testing steps, to ensure that the implementation result of the product development sub-solution meets the initiation conditions of the marketing sub-solution.

[0165] Step S1599: Determine the order of sub-solutions that satisfy dependencies and have logical coherence, and form a list of sorted single-domain decision sub-solutions.

[0166] After verification and adjustment through the above steps, the order of sub-solutions that satisfy dependencies and have logical coherence is determined, forming a list of ordered single-domain decision sub-solutions. For example, the final ordered order might be product development sub-solutions, marketing sub-solutions, and customer service sub-solutions. These are recorded in the list of ordered single-domain decision sub-solutions to provide a basis for subsequent integration of collaborative decision-making solutions.

[0167] Step S1510: Add a scheme overview, implementation sequence description, and scheme verification criteria to the sorted single-domain decision sub-schemes. The scheme overview includes a summary of the core objectives of all single-domain decision sub-schemes. The implementation sequence description describes the initiation conditions and connection requirements of each single-domain decision sub-scheme. The scheme verification criteria include evaluation indicators of the implementation effect of each single-domain decision sub-scheme.

[0168] Add a summary, implementation sequence description, and validation criteria to the sorted single-domain decision sub-solutions. The summary includes a summary of the core objectives of all single-domain decision sub-solutions. For example, if the core objective of product development is to improve product performance, the core objective of marketing is to increase market share, and the core objective of customer service is to improve customer satisfaction, then the summary could be: "Achieve sustainable product development and the company's overall goals by improving product performance, increasing market share, and improving customer satisfaction." The implementation sequence description outlines the initiation conditions and connection requirements for each single-domain decision sub-solution. For example, the initiation condition for a single-domain decision sub-solution in product development is the completion of market demand analysis, and the connection requirement is delivering the developed product to the marketing domain for promotion; the initiation condition for a single-domain decision sub-solution in marketing is the completion of product development, and the connection requirement is feeding market feedback information back to product development for product improvement. The validation criteria include evaluation indicators for the implementation effectiveness of each single-domain decision sub-solution. For example, the implementation effectiveness in product development can be evaluated using product performance indicators, in marketing using market share indicators, and in customer service using customer satisfaction indicators.

[0169] Step S1511: Integrate the overall solution, the ordered single-domain decision sub-solutions, the implementation sequence description, and the solution verification criteria to form the final collaborative decision-making solution.

[0170] The final collaborative decision-making scheme integrates the overall scheme overview, the ranked sub-schemes for each individual domain, the implementation sequence description, and the scheme verification criteria. This final scheme includes the content of all sub-schemes for each individual domain, as well as information on the relationships between them, their implementation sequence, and verification criteria. In practical applications, various science and technology intelligence sensing and information processing carriers can implement specific decisions based on the final collaborative decision-making scheme and evaluate and provide feedback on the implementation effects according to the scheme verification criteria. This allows for timely adjustments to the decision-making scheme, improving the accuracy and effectiveness of the decisions.

[0171] Based on the same inventive concept, please refer to Figure 2 This paper shows a schematic block diagram of the structure of a collaborative decision-making system 100 for a scientific and technological intelligence agent that integrates multimodal perception, provided in an embodiment of this application, for executing the above-described collaborative decision-making method for a scientific and technological intelligence agent that integrates multimodal perception. The collaborative decision-making system 100 for a scientific and technological intelligence agent that integrates multimodal perception may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0172] 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 collaborative decision-making method for a technology intelligence intelligent agent that integrates multimodal perception provided in the foregoing method embodiments.

[0173] 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 collaborative decision-making of a technology intelligence agent fusing multi-modal perception, characterized in that, The method comprises: Obtaining a plurality of sci-tech information perception information, performing preliminary screening of perception deviation on each sci-tech information perception information, identifying pairs of sci-tech information perception information with different content expressions under the same perception scene, generating an initial list of perception deviation, and the sci-tech information perception information comprises corresponding perception information identifier, information content and perception dimension label; Wherein, a plurality of different emotion analysis systems collect sci-tech information perception information from different sources, group sci-tech information perception information with completely consistent perception scene description content into the same perception scene group, assign a unique scene group identifier to each perception scene group, record the perception information identifier of all sci-tech information perception information in each group, and the perception scene description content comprises environmental conditions of information collection, collection time range, collection object state and collection equipment model; Based on the initial list of perception deviation, trace the perception collection process parameters corresponding to the different contents, analyze the correlation between parameter differences and perception deviation, generate a perception deviation traceability report, formulate a cross-agent dynamic calibration rule based on the perception deviation traceability report, supplement and optimize the missing perception content according to the cross-agent dynamic calibration rule, and obtain a calibrated and optimized sci-tech information perception set; Call the cross-agent joint mining function, perform multi-level associated feature extraction on the calibrated and optimized sci-tech information perception set, and integrate to form a multi-level associated feature set; Generate an initial decision suggestion based on the multi-level associated feature set, send the initial decision suggestion to other sci-tech information perception information processing carriers, receive and analyze the initial decision suggestion of other sci-tech information perception information processing carriers, identify conflict points in the decision suggestion, locate the corresponding perception information missing items according to the conflict points, supplement and optimize the calibrated and optimized sci-tech information perception set, and generate an adjusted decision suggestion; Summarize all adjusted decision suggestions, resolve conflicts and fuse contents of the adjusted decision suggestions, and generate a final collaborative decision scheme.

2. The method according to claim 1, wherein, The method comprises: For each perception scene group, extract the information content of all sci-tech information perception information in the group, split the information content into a plurality of perception content sub-items according to the logical structure of the information content, and each perception content sub-item corresponds to a specific perception conclusion; Compare the same perception content sub-items of different sci-tech information perception information in the same perception scene group, identify pairs of perception content sub-items with different expressions, and record the perception information identifiers of the two sci-tech information perception information corresponding to the pair of different perception content sub-items; Differential feature description is performed on the pair of different perception content sub-items, and the differential feature description comprises differential expression type, differential perception parameter involved and semantic distance of differential content; Correlate and record the scene group identifier, the pair of different perception content sub-items, the differential feature description and the corresponding perception information identifier, and generate a draft of the initial list of perception deviation. Supplement the missing perception scene group or difference perception content sub-item in the draft of the initial perception bias list, merge the repeated records in the draft of the initial perception bias list, and form the initial perception bias list.

3. The method of claim 1, wherein the method further comprises: Based on the initial perception bias list, trace the perception acquisition process parameters corresponding to the difference content, analyze the correlation between parameter difference and perception bias, generate a perception bias traceability report, formulate cross-agent dynamic calibration rules based on the perception bias traceability report, supplement and optimize the missing perception content according to the cross-agent dynamic calibration rules, and obtain the calibrated and optimized science and technology intelligence perception set, including: Extract the perception information identifier of the science and technology intelligence perception information corresponding to each difference record from the initial perception bias list, and call the corresponding perception acquisition process log according to the perception information identifier; Extract the perception acquisition process parameters from the perception acquisition process log, which include the sampling frequency of the acquisition equipment, the data preprocessing algorithm type, the interference factor record of the perception environment, and the operation specification execution of the acquisition personnel; Compare the perception acquisition process parameters of the two pieces of science and technology intelligence perception information corresponding to the difference content, identify the parameter items with different values, and record the name and value range of the difference parameter items; Analyze the correlation between the difference parameter items and the perception bias, determine the key difference parameter items that have a dominant influence on the perception bias by calculating the correlation between the parameter difference degree and the semantic distance of the perception bias, and Based on the influence law of the key difference parameter items, generate a perception bias traceability report, which includes a list of key difference parameter items, parameter influence law description, and bias improvement direction suggestions; According to the bias improvement direction suggestions in the perception bias traceability report, formulate cross-agent dynamic calibration rules for each perception dimension, which include calibration source selection criteria, calibration content generation algorithm, calibration weight allocation method, and calibration result verification standard; Send the cross-agent dynamic calibration rules to the corresponding science and technology intelligence perception information processing carrier, determine the calibration source of the perception dimension according to the calibration source selection criteria, and request the standard perception content from the calibration source; According to the calibration content generation algorithm in the cross-agent dynamic calibration rules, fuse and optimize the standard perception content provided by the calibration source with the missing perception content to generate the supplemented and optimized perception content; Integrate the supplemented and optimized perception content with the initial science and technology intelligence perception information, remove duplicate perception content sub-items, and select or discard conflicting perception content sub-items according to the calibration weight allocation method to generate calibrated and optimized science and technology intelligence perception information; Summarize the calibrated and optimized science and technology intelligence perception information of all science and technology intelligence perception information processing carriers to form the calibrated and optimized science and technology intelligence perception set.

4. The method of claim 1, wherein the method further comprises: The calling cross-agent joint mining function performs multi-level correlation feature extraction on the calibrated and optimized science and technology intelligence perception set, integrates to form a multi-level correlation feature set, including: The calibrated and optimized science and technology information perception set is sent to a cross-agent joint mining carrier, and all science and technology information perception information in the calibrated and optimized science and technology information perception set is classified by perception dimension through the cross-agent joint mining carrier, and the science and technology information perception information is divided into multiple single-dimension perception information subsets according to the perception dimension label; For each single-dimension perception information subset, the perception content sub-items of all science and technology information perception information in the single-dimension perception information subset are extracted, the occurrence frequency of the same perception content sub-items is counted, and the perception content sub-items with an occurrence frequency reaching a preset frequency threshold are marked as core perception sub-items of the perception dimension; The logical association between the core perception sub-items in the same single-dimension perception information subset is analyzed, the core perception sub-item pairs with a cause-and-effect relationship or a concomitant relationship are identified, the association type and the scene condition of the association are recorded, and a single-dimension association feature item is generated; After completing the extraction of all single-dimension association features, the different single-dimension perception information subsets are combined two by two to form cross-dimension combination pairs, and each cross-dimension combination pair includes single-dimension perception information subsets of two different perception dimensions; For each cross-dimension combination pair, the core perception sub-items of all science and technology information perception information in the two single-dimension perception information subsets are extracted, the perception scene description contents corresponding to the core perception sub-items are compared, and the core perception sub-item pairs with an overlapping proportion of perception scene description content reaching a preset overlapping proportion are screened out; The association relationship between the screened core perception sub-item pairs is analyzed, the cross-dimension association strength is determined by calculating the semantic association degree of the sub-item content and the synchronicity of the perception time, a cross-perception-dimension association feature item is generated, and the cross-perception-dimension association feature item includes an association strength value, an association scene overlapping degree, and an association stability description; The single-dimension association feature items and the cross-perception-dimension association feature items are classified and arranged, the items are sorted according to the perception dimension priority, an association feature identifier and a perception information identifier of the corresponding science and technology information perception information are added to each item; The items with repeated contents in the single-dimension association feature items and the cross-perception-dimension association feature items are deleted, the items with logical conflicts in the single-dimension association feature items and the cross-perception-dimension association feature items are selected according to the association strength value, and the remaining items are integrated to form a multi-level association feature set.

5. The method of claim 1, wherein the method further comprises: The initial decision suggestion is generated based on the multi-level association feature set, the initial decision suggestion is sent to other science and technology information perception information processing carriers, the initial decision suggestions of the other science and technology information perception information processing carriers are received and analyzed, the conflict points in the decision suggestions are identified, the corresponding perception information missing items are located according to the conflict points, the calibrated and optimized science and technology information perception set is supplemented and optimized, and an adjusted decision suggestion is generated, including: The association feature items related to a preset decision range are extracted from the multi-level association feature set, the preset decision range is determined based on a preset responsibility field, and the items with an association strength value reaching a preset association strength standard are preferentially selected during the extraction; The extracted association feature items are mapped to decision direction, and the association relationship corresponding to each association feature item is converted into a specific decision direction suggestion, which includes a decision target type, a decision implementation priority, and a decision risk prompt; An initial decision suggestion framework is constructed based on the decision direction suggestion, which includes a decision target module, a decision basis module, a decision implementation path module, and a decision risk response module, and the content of the decision basis module is generated based on the core perception sub-items of the extracted association feature items; The specific content of the initial decision suggestion framework is filled in, and detailed descriptions are added to each module. The decision target module describes the specific results that need to be achieved by the decision, the decision implementation path module includes stage operation steps, and the decision risk response module includes countermeasures for potential risks; A processing carrier identifier, a generation time, and an association feature item reference list are added to the initial decision suggestion to form an initial decision suggestion; Through a preset communication channel, the initial decision suggestion is sent to all other science and technology intelligence perception information processing carriers, and the initial decision suggestion with a processing carrier identifier, a generation time, and an association feature item reference list sent by other science and technology intelligence perception information processing carriers is received; The content of all received initial decision suggestions is compared, and the content with completely opposite expressions or step logic conflicts is marked as a decision suggestion conflict point; For each decision suggestion conflict point, the association feature item reference list corresponding to the decision suggestions of the conflict parties is traced back, the differences in the reference items are analyzed, and it is determined that the conflict is caused by the absence of any association feature item or the difference in the association strength value of the reference item; If the conflict is caused by the absence of the association feature item, the perception dimension and the core perception sub-item corresponding to the missing item are located, and the missing item is determined as a perception information missing item; The science and technology intelligence perception information processing carrier holding the core perception sub-item corresponding to the perception information missing item is requested to supplement the perception content, and after receiving the supplemented perception content, the supplemented perception content is added to the calibrated and optimized science and technology intelligence perception set to generate a supplemented and optimized perception set; Based on the supplemented and optimized perception set, the association feature items are re-extracted, the multi-level association feature subset is updated, and the content of the decision target module, the decision implementation path module, and the decision risk response module is adjusted according to the generation process of the initial decision suggestion to form an adjusted decision suggestion.

6. The method of claim 5, wherein the method further comprises: The content comparison of all received initial decision suggestions, the identification of content with completely opposite expressions or step logic conflicts, and the marking of decision suggestion conflict points include: All received initial decision suggestions are classified according to decision fields, and initial decision suggestions with the same decision target module belonging to the same decision field are classified into the same decision suggestion group, and each decision suggestion group is assigned a decision field identifier; For each decision suggestion group, the core expression of the decision target module of all initial decision suggestions in the decision suggestion group is extracted, which includes the core result and key evaluation index that need to be achieved by the decision. Comparing the core expressions of the decision-making objectives of different initial decision-making suggestions in the decision-making suggestion group, calculating the semantic similarity between the expressions, if the semantic similarity is lower than the preset semantic similarity threshold and the core expressions are opposite, marking it as a decision-making objective conflict point, and recording the processing carrier identifier and core expression content of the conflicting parties; After comparing the decision-making objectives in the decision-making suggestion group, extracting the step sequence of the decision-making implementation path module of all initial decision-making suggestions, which includes phased operation steps and step execution order; Comparing the step sequences of different initial decision-making suggestions, first counting the consistency of the number of steps, then checking the consistency of the operation content at the same step position, if there is a difference in the number of steps and the key operation steps are missing, or the operation content of the steps at the same position is completely opposite, it is marked as a decision-making implementation path conflict point; Numbering the decision-making objective conflict points and decision-making implementation path conflict points, recording the decision-making suggestion group identifier, conflict type, and initial decision-making suggestion reference information of the conflicting parties corresponding to each conflict point; Marking the content of the risk assessment results in the decision risk response module as a decision risk response conflict point, and including the decision risk response conflict point in the decision suggestion conflict point list; Determining the conflict severity according to the influence range of the conflict on the overall decision, sorting all marked conflict points according to the conflict severity, and forming the final decision suggestion conflict point list.

7. The method of claim 5, wherein the method further comprises: If the conflict is caused by the absence of an associated feature item, locate the corresponding perception dimension and core perception sub-item of the missing item, and determine it as a perception information missing item, including: Extracting the associated feature item reference list of the initial decision-making suggestions of the conflicting parties, comparing the associated feature identifiers in the two lists, identifying the associated feature identifiers that only exist in one list, and marking them as potential missing associated feature identifiers; According to the potential missing associated feature identifier, query the corresponding associated feature item from the multi-level associated feature set, obtain the perception dimension label and core perception sub-item description of the associated feature item; Analyze the role of the associated feature item in the conflicting decision-making suggestions by checking the reference content of the associated feature item in the decision basis module to determine whether the associated feature item is a key basis for supporting the decision conclusion of the conflict party; If the associated feature item is a key basis and is only referenced by one party, it is determined that the conflict is caused by the other party not obtaining the associated feature item, i.e., there is an associated feature item missing; Based on the perception dimension label of the associated feature item, determine the corresponding perception dimension, and based on the core perception sub-item description, determine the specific core perception sub-item under the perception dimension; Query whether the core perception sub-item under the perception dimension is included in the self-calibration and optimization of the science and technology intelligence perception set, if not, record the core perception sub-item and the corresponding perception dimension as a perception information missing item; If the core perception sub-item is included in the self-calibration and optimization of the science and technology intelligence perception set but not extracted as an associated feature item, analyze the reason for not being extracted, if the associated strength value does not reach the preset extraction strength threshold, record the core perception sub-item and the corresponding associated strength calculation parameter as a perception information optimization item, and include the perception information optimization item in the perception information missing item category. The missing items and the optimization items of the perception information are arranged into a perception supplement demand list, and the corresponding perception dimension, the core perception sub-item content, and the supplement priority of each missing item are recorded.

8. The method of claim 1, wherein the method further comprises: The adjusted decision suggestions are summarized, and conflict resolution and content fusion are performed on the adjusted decision suggestions to generate a final collaborative decision scheme, including: A collaborative decision fusion platform is established, and the adjusted decision suggestions of all technology intelligence perception information processing carriers are sent to the collaborative decision fusion platform. The collaborative decision fusion platform performs format unification processing on the received adjusted decision suggestions to ensure that the module structures of the adjusted decision suggestions are consistent. The adjusted decision suggestions are classified according to decision fields, and the adjusted decision suggestions whose decision target modules belong to the same decision field are classified into the same fusion group. A fusion group identifier and a decision field label are assigned to each fusion group. For each fusion group, the core expressions of the decision target modules of all the adjusted decision suggestions in the fusion group are extracted, and clustering is performed by calculating semantic similarity. The core expressions with a semantic similarity higher than a preset clustering similarity threshold are classified into the same target cluster, and the core expression with the highest occurrence frequency in the target cluster is selected as the unified decision target of the fusion group. The contents of the decision basis modules of all the adjusted decision suggestions in the fusion group are extracted, and the same or similar decision bases are integrated. Repetitive expressions are removed, and the contents of the bases that exist differences but do not conflict are supplemented and integrated to form the unified decision basis of the fusion group. The decision implementation path modules of the adjusted decision suggestions in the fusion group are analyzed, the implementation path with the most complete step logic and the most detailed operation description is selected as the basic path, and the reasonable step supplement contents in other adjusted decision suggestions are integrated to optimize the unified implementation path of the fusion group. The decision risk response modules of the adjusted decision suggestions in the fusion group are summarized, and the risk identification results and response measures in different adjusted decision suggestions are integrated to form the unified risk response scheme of the fusion group. The unified decision target, the unified decision basis, the unified implementation path, and the unified risk response scheme of each fusion group are integrated to form a single-field decision sub-scheme. All the single-field decision sub-schemes are sorted according to the importance priority of the decision fields. The importance priority is determined based on the contribution of the decision in the field to the overall collaborative target. The correlation between different single-field decision sub-schemes is analyzed, and the sub-schemes with a dependency relationship are identified. The sorting order of the sub-schemes is adjusted to make the implementation logic of the sub-schemes coherent. A scheme summary, an implementation sequence description, and a scheme verification standard are added to the sorted single-field decision sub-schemes. The scheme summary includes the core target summary of all the single-field decision sub-schemes. The implementation sequence description describes the start-up conditions and the connection requirements of the single-field decision sub-schemes. The scheme verification standard includes the evaluation indexes of the implementation effect of the single-field decision sub-schemes. The scheme summary, the sorted single-field decision sub-schemes, the implementation sequence description, and the scheme verification standard are integrated to form the final collaborative decision scheme.

9. The method of claim 8, wherein the method further comprises: The method comprises the following steps: for each fusion group, extracting a decision target module core expression of all adjusted decision suggestions in the fusion group, clustering by calculating semantic similarity, and classifying core expressions with a semantic similarity higher than a preset clustering similarity threshold into the same target cluster; selecting a core expression with the highest frequency of occurrence in the target cluster as a unified decision target of the fusion group, and comprising the following steps: extracting a decision target module content of each adjusted decision suggestion, removing redundant description sentences in the decision target module, and retaining expressions containing decision core results and key evaluation indexes as decision target core expressions; performing text preprocessing on all decision target core expressions, including word segmentation, stop word removal, and semantic feature vector conversion, converting each decision target core expression into a fixed-dimension semantic feature vector; using a clustering algorithm to perform clustering processing on the semantic feature vectors, calculating the cosine similarity between any two semantic feature vectors, classifying vectors with a cosine similarity higher than a preset clustering similarity threshold into the same cluster, and forming multiple target clusters; counting the number of decision target core expressions contained in each target cluster, and determining a target cluster containing the most expressions as a main target cluster; if there are multiple main target clusters containing the same number of expressions, calculating the average vector of all semantic feature vectors in each main target cluster, comparing the average vector with a preset core target vector of the decision field, and selecting the target cluster corresponding to the average vector with the highest similarity as the final main target cluster; selecting a decision target core expression with the highest frequency of occurrence from the final main target cluster, and if there are multiple expressions with the same frequency, calculating the similarity between the multiple expressions and the clustering average vector, and selecting the expression with the highest similarity as the unified decision target of the fusion group; supplementing the missing core demand expressions of the decision field in the unified decision target, extracting related expressions from other target clusters and adding them to the unified decision target to form a unified decision target description.

10. A system for collaborative decision making by a fusion of multi-modal perception of technology intelligence agents, characterized in that, comprising: a processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the machine-executable instructions to perform the method of claim 1-9.

Images