Information technology retrieval consultation system and method

Abstract

The invention relates to the technical field of data retrieval, in particular to an information technology retrieval consultation system and method. The method comprises the following steps: acquiring user input data; extracting modal key information of the user input data to obtain multi-modal key semantic information data; performing data preprocessing on the multi-modal key semantic information data to generate standard user input data; performing historical retrieval information retrieval on the standard user input data, and when a historical retrieval record is retrieved, performing accurate portrait construction on the standard user input data to generate a user accurate retrieval portrait; and when no historical record is retrieved, performing retrieval reasoning on the standard user input data to generate a user fuzzy retrieval portrait. By combining multi-modal information extraction, accurate and fuzzy portrait construction, cross-context inheritance, dynamic knowledge fusion and sentiment analysis optimization, the problems of low retrieval precision, incomplete information coverage and poor user experience in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the field of data retrieval technology, and in particular to an information technology retrieval consultation system and method. Background Art

[0002] Initially, information retrieval mainly relied on manual operations and simple keyword matching. Human searchers helped users find the information they needed through indexing and classification. The retrieval methods of this period were simple, relying on manual knowledge and experience, with low retrieval efficiency and poor accuracy. With the rise of computer technology and the Internet, information retrieval methods have undergone fundamental changes. The emergence of search engine technology, such as the early AltaVista and Google, based on complex algorithm models, realized automated information retrieval by crawling and indexing web page content. Users can obtain faster and more accurate search results through simple keyword input. With the development of big data and artificial intelligence technology, information retrieval is not limited to the search of static web pages, but also extends to multi-dimensional information analysis. Technologies based on natural language processing (NLP), machine learning and deep learning are gradually applied to information retrieval systems, improving the level of intelligence in information retrieval. The application of technologies such as personalized recommendation, semantic search and intelligent question and answer makes retrieval consultation more accurate and efficient. However, in traditional retrieval methods, each query is often processed independently, and the previous context information cannot be effectively inherited, resulting in a lack of relevance in continuous queries. At the same time, when performing knowledge retrieval, it relies on a single data source or static information, resulting in limitations in retrieval results. Summary of the invention

[0003] Based on this, it is necessary to provide an information technology retrieval consultation system and method to solve at least one of the above technical problems.

[0004] To achieve the above purpose, an information technology search and consultation method is provided, the method comprising the following steps:

[0005] Step S1: obtaining user input data; extracting modal key information of the user input data to obtain multimodal key semantic information data; performing data preprocessing on the multimodal key semantic information data to generate standard user input data;

[0006] Step S2: Perform historical search information retrieval on the standard user input data. When the existing historical search records are retrieved, an accurate portrait is constructed for the standard user input data to generate a user accurate search portrait; when no existing historical records are retrieved, search reasoning is performed on the standard user input data to generate a user fuzzy search portrait; a user input context map is constructed, and cross-query context inheritance is performed on the user input context map based on the user accurate search portrait and the user fuzzy search portrait to generate similarity search inheritance data;

[0007] Step S3: Obtain external information data; perform dynamic knowledge retrieval on the user input data according to the similarity retrieval inheritance data to generate dynamic knowledge retrieval data; perform multi-knowledge source fusion on the dynamic knowledge retrieval data and cross-domain similarity retrieval data to generate fusion retrieval data; perform retrieval display adjustment on the fusion retrieval data through a preset sentiment analysis model to generate the final hybrid retrieval display data;

[0008] Step S4: Perform hybrid retrieval feedback optimization on the final hybrid retrieval display data to execute the information technology retrieval consultation optimization operation.

[0009] By extracting multi-modal key information (such as text, voice, image, etc.) input by the user and performing data preprocessing, the present invention can uniformly standardize the user input in different forms, ensuring the consistency and accuracy of subsequent processing, thereby laying a solid foundation for the entire retrieval process. By retrieving historical records or inferring to generate an accurate or fuzzy retrieval portrait of the user, the user's needs can be described personalized, improving the retrieval accuracy. Especially when there is no historical record, a fuzzy retrieval portrait is generated through inference to ensure the system's processing ability for new users or a small amount of data. At the same time, the inheritance mechanism across query contexts can continuously optimize the context relevance of the user input, enhancing the relevance and consistency of consecutive queries. Through dynamic knowledge retrieval and cross-domain information fusion, a wider knowledge coverage can be achieved among multiple information sources, avoiding the limitations caused by a single data source or static information. This fusion not only improves the comprehensiveness and accuracy of the retrieval, but also enables more flexible integration of data in different fields, ensuring higher-quality retrieval results. In addition, the sentiment analysis model adjusts the display method according to the user's sentiment tendency, improving the user experience and satisfaction. Through continuous hybrid retrieval feedback optimization, the retrieval strategy can be adjusted according to the user's real-time feedback, dynamically optimizing the information technology retrieval process. This feedback mechanism can ensure that the retrieval system is continuously improved during operation, thereby improving the system's intelligence level and personalized service ability. Therefore, the present invention solves the problems of low retrieval accuracy, incomplete information coverage, and poor user experience in the prior art by combining multi-modal information extraction, precise and fuzzy portrait construction, cross-context inheritance, dynamic knowledge fusion, and sentiment analysis optimization.

[0010] Preferably, step S1 includes the following steps:

[0011] Step S11: Obtain user input data, where the user input data includes text input data, image input data, and voice input data;

[0012] Step S12: Extract modal key information from the text input data, image input data, and voice input data to obtain multi-modal key information extraction data;

[0013] Step S13: Perform unified semantic representation on the multi-modal key information extraction data to generate multi-modal key semantic information data;

[0014] Step S14: Perform data preprocessing on the multi-modal key semantic information data to generate standard user input data, where the data preprocessing includes data missing value prompting, data abnormal format conversion, and data duplicate value elimination.

[0015] By obtaining text, image, and voice data and extracting key information, the present invention can comprehensively understand the intentions and needs of users. The extraction of such multi-modal information enables the system to simultaneously understand text, visual, and voice content and provide more accurate responses. Converting data of different modalities into a unified semantic representation can reduce ambiguity or information loss problems caused by different data sources. The unified semantic representation helps improve the efficiency and accuracy of the system when processing multiple data sources. By handling data missing, abnormal format conversion, and removing duplicate values, the quality of the input data is ensured. This process can avoid system errors caused by data problems and enhance the stability and reliability of the system. Through efficient multi-modal data processing, no matter what form the user's input is, it can be quickly and accurately understood and responded to by the system, thus enhancing the interaction efficiency and user experience.

[0016] Preferably, step S2 includes the following steps:

[0017] Step S21: Perform historical retrieval information retrieval on the standard user input data. When a historical retrieval record is retrieved, the user's historical retrieval data is retrieved, and a user portrait is constructed based on the user's historical retrieval data to generate a user precise retrieval portrait;

[0018] Step S22: When no historical record is retrieved, perform user intention recognition on the standard user input data to obtain user intention recognition data; use the user intention recognition data to perform retrieval reasoning on the standard user input data to generate a user fuzzy retrieval portrait;

[0019] Step S23: Construct a user input context graph for the standard user input data to obtain a user input context graph; perform dynamic update of the graph context complexity weight on the user input context graph based on the user precise retrieval portrait and the user fuzzy retrieval portrait to generate graph context dynamic update data;

[0020] Step S24: Calculate the relevant edge nodes of the graph context dynamic update data, and perform cross-query context inheritance on the user input context graph through the relevant edge nodes to generate similarity retrieval inheritance data.

[0021] By retrieving the historical data of users, the present invention can effectively construct an accurate retrieval profile of users. This personalized recommendation system based on historical records helps to provide more customized services and enhance the user experience. For example, when the user's input is related to the historical search content, the system can quickly provide information related to the user's historical behavior, thereby improving the retrieval efficiency and accuracy. When there is no historical record, by identifying the user's intention and conducting retrieval reasoning, a fuzzy retrieval profile of the user can be generated. Through this process, even without the user's historical data, the system can still construct a preliminary user profile by identifying the intention of the user's current needs, helping the system understand and predict the potential needs of the user. By constructing the context graph of the user input and combining the accurate and fuzzy profiles of the user, the dynamic update of the complexity weight of the context graph is realized. This update not only improves the accuracy and comprehensiveness of the graph, but also can be adjusted in real time according to the changes in the user's behavior and needs, enabling the system to respond to the user more intelligently. By using the relevance edge nodes of the graph context to dynamically update the data, the context inheritance across queries can be realized. This mechanism can establish connections between different queries, enabling the system to understand and utilize the previous query information to provide more consistent and continuous services, further improving the retrieval efficiency and intelligence level.

[0022] Preferably, step S22 includes the following steps:

[0023] Step S221: When no historical record is retrieved, perform keyword part-of-speech extraction on the standard user input data to obtain keyword part-of-speech extraction data; perform sentiment classification on the keyword part-of-speech extraction data to generate keyword sentiment classification data;

[0024] Step S222: Perform tendency analysis on the standard user input data through the keyword sentiment classification data to generate user intention recognition data; perform intention classification on the user intention recognition data to obtain user intention classification data;

[0025] Step S223: Perform behavior matching according to the user intention classification data to obtain user behavior data; perform user preference association reasoning according to the user behavior data to obtain user preference reasoning data;

[0026] Step S224: Perform user clustering on the user based on the user preference reasoning data to obtain a fuzzy retrieval profile of the user.

[0027] Through keyword part-of-speech extraction and sentiment classification of the standard user input data in step S221, the present invention can help the system more accurately understand the user's sentiment tendency and context. For example, sentiment classification helps to distinguish whether the user is in a positive, negative or neutral mood, which is crucial for providing appropriate feedback or services. The system can adjust the tone and content of the response based on this sentiment information, making the interaction more personalized and user-friendly. By performing a tendency analysis on the keyword sentiment classification data, user intention recognition data is further generated. This process helps to accurately identify the user's core needs and purposes and provide a response that better meets the user's expectations. Through intention recognition, the system can avoid generating results that do not match the user's actual needs and improve the relevance and quality of the service. Based on the user intention classification data, behavior matching is performed to deduce the user's behavior pattern. In addition, through preference association reasoning, the system can further understand the user's preferences and interests. This step helps to enhance the system's recognition of the user's personalized needs, enabling the system to accurately predict the user's future needs and provide more appropriate suggestions or content based on this. By clustering the user preference reasoning data, users can be divided into different groups according to the similarity of their behaviors and needs, generating a fuzzy retrieval portrait of the user. This process helps to construct a more detailed user portrait, enabling the system to provide customized retrieval results based on these portraits and improve the user experience. Especially when there is a lack of accurate historical records, it can still provide services with a high degree of accuracy. By integrating the analysis of multiple steps, step S22 provides the system with a deeper understanding of the user. The system can better understand the user's sentiment, intention, behavior and preferences, and thus automatically generate the most relevant retrieval portrait when the user inputs information, optimizing the accuracy of recommendations and searches.

[0028] Preferably, step S24 includes the following steps:

[0029] Step S241: Based on the graph context dynamic update data, perform initial node screening on the user input context graph to obtain initial screening nodes; use the graph algorithm to perform boundary point correlation scoring on the initial screening nodes to obtain the correlation score of each initial screening node;

[0030] Step S242: Screen out the edge nodes with the highest correlation according to the correlation score of each initial screening node and mark them as correlation edge nodes; perform edge connection node screening on the initial screening nodes through the correlation edge nodes to obtain adjacent associated screening nodes;

[0031] Step S243: Perform node directionality analysis on the adjacent associated screening nodes to generate adjacent associated screening node directionality data; based on the adjacent associated screening node directionality data, perform context node marking on the adjacent associated screening nodes to obtain upper context associated nodes and lower context associated nodes;

[0032] Step S244: Calculate the node information entropy value of the correlation edge node through the above-mentioned associated node and the following associated node to obtain the standard node information entropy value; perform cross-query context inheritance on the user input context graph through the standard node information entropy value to generate similarity retrieval inheritance data.

[0033] Through step S241 of the present invention, initial node screening is performed based on the graph context dynamic update data, and the nodes are scored for correlation, which can effectively identify the nodes most relevant to the user input. This screening and scoring mechanism based on the graph algorithm helps the system focus on the nodes with the most information value, improving the efficiency and accuracy of data processing. Step S242 screens out the most relevant edge nodes according to the correlation scores of each initially screened node and performs edge connection node screening. This process can establish a more accurate node connection relationship and optimize the structure of the graph by identifying the key associated edge nodes, improving the accuracy of information retrieval. Step S243 generates the directional data of the adjacent associated screened nodes through node directional analysis and performs context node marking. This analysis ensures that the system can deeply understand the context relationship between nodes, thereby distinguishing the associated information between the above text and the following text. By marking these context nodes, the system can more flexibly adjust the extraction and use of information, enhancing the understanding of user needs. Step S244 realizes cross-query context inheritance by calculating the information entropy value of the standard node. This process can quantify the uncertainty of node information and dynamically update the graph according to this information entropy value, enabling similar queries to share and inherit relevant context data, thereby improving the efficiency and accuracy of similarity retrieval. Through the above steps, the system can dynamically adjust and optimize the retrieval process according to the context information input by the user. Especially when dealing with complex or variable queries, it can perform similarity inheritance and information optimization more intelligently, which enables the system to provide more efficient, accurate and personalized retrieval results and improve the user experience.

[0034] Preferably, performing cross-query context inheritance on the user input context graph through the node information entropy value includes:

[0035] Perform context topology structure analysis on the user input context graph to obtain context structure topology data; extract semantic inheritance relationship data from the context structure topology data to generate semantic inheritance relationship data;

[0036] Perform query pattern matching on the semantic inheritance relationship data through the node information entropy value to generate query pattern matching data; perform propagation path analysis on the user input context graph according to the query pattern matching data to generate context query propagation path data;

[0037] Verify the cross-node inheritance rules for the context query propagation path data using the standard node information entropy value. When the information entropy value of a node in the context query propagation path is less than or equal to the standard node information entropy value, similarity retrieval inheritance data is generated; otherwise, the corresponding context query propagation path is excluded.

[0038] The present invention analyzes the topological structure of the context graph input by the user to obtain context structure topological data. Through this analysis, the system can identify the relationships and dependencies between nodes, laying a foundation for subsequent extraction of semantic inheritance relationships. The generation of semantic inheritance relationship data ensures that context information can be effectively transmitted and continued during the query process, thereby providing the system with more accurate context understanding. Matching the query pattern with the semantic inheritance relationship data through the node information entropy value helps the system identify the query intent input by the user and construct a matching pattern on this basis. This matching mechanism enables the system to quickly determine the association between the user's current query intent and the historical query pattern, providing an accurate basis for subsequent context propagation path analysis. By generating context query propagation path data, the system can better handle the information flow and association between multiple queries, thereby improving the retrieval efficiency. The standard node information entropy value is used to verify the information entropy value of the nodes in the context query propagation path, thereby determining whether to execute the cross-node inheritance rules. If the information entropy value of a certain node is lower than the standard node information entropy value, then this node can be considered to conform to the inheritance rules, thereby generating similarity retrieval inheritance data; otherwise, the system will exclude this node to avoid irrelevant or low-correlation nodes from affecting the query results. This dynamic verification mechanism optimizes the accuracy of information inheritance, ensuring that only relevant and high-value node information is transmitted and inherited. Through cross-query context inheritance, the system can effectively maintain and continue the context consistency in the query. Especially when multiple queries have similar information, the system can automatically inherit the existing context relationships, avoiding repeated calculations or information loss. This mechanism improves the accuracy of the system response, especially in complex queries or multi-round conversations, and can accurately understand the user's intent and provide more accurate results. Through dynamic adjustment based on the node information entropy value, the system can flexibly adapt to different types of queries, ensuring the effective propagation of information in the query path, while irrelevant nodes or error information are excluded. This enhances the system's adaptability in dealing with complex queries and various scenarios, enabling the system to generate similarity retrieval results more intelligently according to the context environment.

[0039] Preferably, step S3 includes the following steps:

[0040] Step S31: Obtain external information data using an external domain knowledge base;

[0041] Step S32: Perform cross-domain similarity retrieval on the user input data based on the similarity retrieval inheritance data to obtain cross-domain similarity retrieval data; perform retrieval domain analysis on the cross-domain similarity retrieval data to generate retrieval domain analysis data;

[0042] Step S33: Perform dynamic knowledge retrieval on the cross-domain similarity retrieval data based on the retrieval domain analysis data to generate dynamic knowledge retrieval data; perform multi-knowledge source fusion on the dynamic knowledge retrieval data and the cross-domain similarity retrieval data to generate fusion retrieval data;

[0043] Step S34: Perform user input context sentiment analysis on the standard user input data through a preset sentiment analysis model to generate user input context sentiment data; use the user input context sentiment data to adjust the retrieval display of the fusion retrieval data to generate the final hybrid retrieval display data.

[0044] In the present invention, through step S31, the external information data is obtained by using the external domain knowledge base, and the system can quickly integrate external professional knowledge, improving the depth and breadth of information processing. This integration not only enables the system to better understand and process the user input data, but also provides more comprehensive background information for subsequent cross-domain similarity retrieval. In step S32, cross-domain similarity retrieval is performed on the user input data through the similarity retrieval inheritance data to ensure that the user input information can be compared and analyzed across different domains. This can find information related to the user query but existing in other domains, broadening the retrieval perspective. In addition, the generation of retrieval domain analysis data helps the system understand the diversity and domain characteristics of information sources, thereby performing retrieval and processing more accurately. In step S33, based on the retrieval domain analysis data, dynamic knowledge retrieval is performed, and fusion retrieval data is generated through multi-knowledge source fusion. This step can not only respond to changes in user input in real time, but also provide more accurate answers according to the knowledge fusion of different domains. By fusing different knowledge sources, the system can obtain more comprehensive and accurate information, avoiding the limitations of a single source and improving the multi-dimensionality and accuracy of the answers. In step S34, sentiment analysis is performed on the user input context through a sentiment analysis model, which can capture the user's emotional tendency when expressing. According to the sentiment analysis results, the retrieval display of the fusion retrieval data is adjusted to ensure that the final retrieval display data is not only accurate, but also meets the user's emotional needs. This sentiment-based display optimization can improve the user's interaction experience, especially in application scenarios that require emotional resonance, increasing the intelligence and humanization of the system. Through the above steps, the system can cross domains, fuse multiple knowledge sources, and provide richer retrieval results. At the same time, the addition of sentiment analysis makes the displayed data more in line with the user's emotional tendency, thus enhancing the overall user experience. This method ensures the accuracy, breadth, and personalization of information retrieval, enabling users to quickly obtain answers that meet their needs, and also being more in line with the context and emotional needs of the situation.

[0045] Preferably, the user input context sentiment analysis of the standard user input data through a preset sentiment analysis model includes:

[0046] Performing sentiment dictionary matching on the standard user input data to generate sentiment dictionary matching data; extracting sentiment vocabulary from the standard user input data according to the sentiment dictionary matching data to obtain sentiment vocabulary data;

[0047] Performing sentiment polarity analysis based on the sentiment vocabulary data and judging the sentiment tendency in the standard user input data to obtain sentiment polarity data; performing context sentiment enhancement on the sentiment polarity data through a preset sentiment analysis model to generate context sentiment enhancement data;

[0048] Evaluating the sentiment intensity of the standard user input data by using the context sentiment enhancement data to generate user input context sentiment data.

[0049] Through sentiment dictionary matching and sentiment vocabulary extraction in the present invention, the system can efficiently capture the sentiment information in the text and reduce the risk of misjudgment. The matching of the sentiment dictionary can ensure that no sentiment-related vocabulary is missed in the analysis process, thereby improving the accuracy of sentiment analysis. Context sentiment enhancement can more accurately understand and interpret the sentiment tendency by combining a preset model and the specific context of the user input. Compared with pure sentiment vocabulary analysis, context-based sentiment enhancement can effectively identify implicit sentiments (such as ironic, sarcastic, etc. sentiment expressions), thereby enhancing the ability to handle complex contexts. Through sentiment intensity evaluation, the system can not only judge whether the sentiment of the user input is positive, negative or neutral, but also further evaluate the strength of the sentiment, which helps the system to make more accurate and personalized responses according to the intensity of the user sentiment. The system can adjust the response method according to the sentiment data input by the user. For example, when detecting that the user sentiment is relatively negative, the system can be adjusted to more soothing or positive incentive language to provide a more user-friendly interaction experience. This enhancement of sentiment intelligence helps to improve the user's satisfaction and interaction quality with the system. The results of sentiment analysis can be used to optimize scenarios such as recommendation systems, advertising placement, and user services. For example, in customer support or social media monitoring, the system can adjust according to the user's sentiment feedback, improve service strategies, and increase user loyalty and satisfaction.

[0050] Preferably, step S4 includes the following steps:

[0051] Step S41: Collecting user consultation feedback on the final mixed retrieval display data to obtain user consultation feedback data, where the user consultation feedback collection includes click feedback collection and dwell time feedback collection;

[0052] Step S42: Optimize the final hybrid retrieval display data based on the user consultation feedback data to perform information technology retrieval consultation optimization operations.

[0053] Through step S41 of the present invention, user consultation feedback is collected for the final hybrid retrieval display data. The system can comprehensively understand the interaction mode between users and the display data, specifically including click feedback and dwell time feedback. These feedbacks provide valuable user behavior data, helping the system analyze users' interest points and areas of concentrated attention, so as to more accurately adjust and optimize the retrieval display content. Click feedback reflects users' interest in certain content, while dwell time can provide the degree of users' engagement and in-depth reading of the content. Step S42 optimizes the final hybrid retrieval display data based on the collected user consultation feedback data. This optimization process can adjust the relevance and display mode of search results, enabling the system to better adapt to users' needs and preferences. For example, the system can adjust the ranking according to the click volume, or optimize the order and form of content display according to the dwell time, thereby improving the accuracy and attractiveness of retrieval results. Through optimization based on user feedback, the system can gradually improve the retrieval algorithm and display strategy, providing a retrieval experience that better meets users' needs. The long-term feedback loop can help the system continuously adjust and enhance the relevance and personalization of retrieval, thereby effectively improving user satisfaction and the system's user stickiness. Through feedback optimization of the final display data, the system can continuously adjust its information technology retrieval strategy, improve the retrieval algorithm, making the system more efficient, accurate, and intelligent when processing information queries. This can not only improve the retrieval efficiency but also provide more customized consultation optimization according to the feedback of different users, further enhancing the intelligent level of the information retrieval system. This optimization process enables the system to dynamically adapt to the needs and behavior changes of different users. By continuously monitoring and collecting user behavior data, the system can more flexibly adjust the display strategy, respond to users' needs in real time, and thus maintain an efficient and personalized service.

[0054] In this specification, an information technology retrieval consultation system is provided for performing the above-mentioned information technology retrieval consultation method. The information technology retrieval consultation system includes:

[0055] A data processing module, configured to obtain user input data; extract modal key information of the user input data to obtain multi-modal key semantic information data; perform data preprocessing on the multi-modal key semantic information data to generate standard user input data;

[0056] A retrieval inheritance module is used to perform historical retrieval information retrieval on standard user input data. When historical retrieval records are found, an accurate portrait of the user is constructed based on the standard user input data to generate an accurate user retrieval portrait. When no historical records are found, retrieval reasoning is performed on the standard user input data to generate a fuzzy user retrieval portrait. A user input context graph is constructed, and cross-query context inheritance is performed on the user input context graph based on the accurate user retrieval portrait and the fuzzy user retrieval portrait to generate similarity retrieval inheritance data.

[0057] A fusion retrieval module is used to obtain external information data. Perform dynamic knowledge retrieval on user input data according to the similarity retrieval inheritance data to generate dynamic knowledge retrieval data. Perform multi-knowledge source fusion on the dynamic knowledge retrieval data and cross-domain similarity retrieval data to generate fusion retrieval data. Adjust the retrieval display of the fusion retrieval data through a preset sentiment analysis model to generate the final hybrid retrieval display data.

[0058] A retrieval feedback module is used to optimize the hybrid retrieval feedback of the final hybrid retrieval display data to perform information technology retrieval consultation optimization operations.

[0059] The beneficial effects of the present invention are as follows: By obtaining user input and extracting multi-modal key information, different forms of user input (such as text, images, voice, etc.) are converted into a standardized data format to ensure the consistency and processability of the input data. This preprocessing process helps the subsequent retrieval system to more accurately understand user needs and provides a reliable data basis for precise retrieval and reasoning. By using historical retrieval records and reasoning to generate accurate or fuzzy retrieval portraits of users, personalized needs analysis is achieved. When there is no historical record, a fuzzy portrait is generated through reasoning to ensure that the system can handle new users or situations with a small amount of data. At the same time, through cross-query context inheritance technology, the relevance in continuous queries is improved, ensuring context consistency during multiple interactions, and the retrieval accuracy is further enhanced through similarity retrieval inherited data. Through dynamic knowledge retrieval and cross-domain information fusion, relevant information can be obtained from multiple knowledge sources, improving the breadth and depth of retrieval. The fusion of multiple knowledge sources avoids the problem of information silos and increases the system's coverage of multi-domain data. At the same time, the introduction of an emotion analysis model enables the retrieval results to be adjusted according to the user's emotional tendency, thereby enhancing the personalization of the retrieval results and the user experience. Through retrieval feedback optimization, this module can dynamically adjust the retrieval strategy and results according to the user's feedback, improving the intelligence and adaptability of the retrieval system. The continuous optimization process ensures that the system is continuously improved, improving the relevance, accuracy, and real-time nature of query results, while enhancing the system's learning ability, enabling future retrievals to better meet user needs. Therefore, the present invention solves the problems of low retrieval accuracy, incomplete information coverage, and poor user experience in the prior art by combining multi-modal information extraction, accurate and fuzzy portrait construction, cross-context inheritance, dynamic knowledge fusion, and emotion analysis optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 is a schematic flow chart of the steps of an information technology retrieval consulting method;

[0061] Figure 2 is Figure 1 a detailed implementation step flow chart of step S2 in

[0062] Figure 3 is Figure 1 a detailed implementation step flow chart of step S3 in

[0063] The realization, functional characteristics, and advantages of the object of the present invention will be further described with reference to the accompanying drawings in combination with embodiments. DETAILED IMPLEMENTATION MANNER

[0064] The technical method of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0065] In addition, the accompanying drawings are only schematic diagrams of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0066] It should be understood that although terms such as "first" and "second" may be used here to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used here includes any and all combinations of one or more of the listed related items.

[0067] To achieve the above object, please refer to Figures 1 to 3 , an information technology retrieval and consultation method, the method comprising the following steps:

[0068] Step S1: Obtain user input data; extract the modal key information of the user input data to obtain multi-modal key semantic information data; perform data preprocessing on the multi-modal key semantic information data to generate standard user input data;

[0069] Step S2: Perform a historical retrieval information retrieval on the standard user input data. When a historical retrieval record is retrieved, an accurate portrait of the user is constructed for the standard user input data to generate an accurate user retrieval portrait; when no historical record is retrieved, a retrieval inference is performed on the standard user input data to generate a fuzzy user retrieval portrait; construct a user input context graph, and perform cross-query context inheritance on the user input context graph according to the accurate user retrieval portrait and the fuzzy user retrieval portrait to generate similarity retrieval inheritance data;

[0070] Step S3: Obtain external information data; perform dynamic knowledge retrieval on the user input data according to the similarity retrieval inheritance data to generate dynamic knowledge retrieval data; perform multi-knowledge source fusion on the dynamic knowledge retrieval data and cross-domain similarity retrieval data to generate fusion retrieval data; perform retrieval display adjustment on the fusion retrieval data through a preset sentiment analysis model to generate the final hybrid retrieval display data;

[0071] Step S4: Perform hybrid retrieval feedback optimization on the final hybrid retrieval display data to execute the information technology retrieval consultation optimization operation.

[0072] The present invention can uniformly standardize different forms of user input by extracting multi-modal key information (such as text, voice, image, etc.) of the user input and performing data preprocessing, ensuring the consistency and accuracy of subsequent processing, thereby laying a solid foundation for the entire retrieval process. By retrieving historical records or inferring to generate an accurate or fuzzy retrieval portrait of the user, the user's needs can be described personalized, improving the retrieval accuracy. Especially when there is no historical record, a fuzzy retrieval portrait is generated through inference to ensure the system's processing ability for new users or small amounts of data. At the same time, the inheritance mechanism across query contexts can continuously optimize the context relevance of the user input, improving the relevance and consistency of consecutive queries. Through dynamic knowledge retrieval and cross-domain information fusion, a wider knowledge coverage can be achieved among multiple information sources, avoiding the limitations caused by a single data source or static information. This fusion not only improves the comprehensiveness and accuracy of the retrieval but also enables more flexible integration of data in different fields, ensuring higher-quality retrieval results. In addition, the sentiment analysis model adjusts the display method according to the user's sentiment tendency, improving the user experience and satisfaction. Through continuous hybrid retrieval feedback optimization, the retrieval strategy can be adjusted according to the user's real-time feedback, dynamically optimizing the information technology retrieval process. This feedback mechanism can ensure that the retrieval system is continuously improved during operation, thereby improving the system's intelligence level and personalized service ability. Therefore, the present invention solves the problems of low retrieval accuracy, incomplete information coverage, and poor user experience in the prior art by combining multi-modal information extraction, accurate and fuzzy portrait construction, cross-context inheritance, dynamic knowledge fusion, and sentiment analysis optimization.

[0073] In the embodiment of the present invention, referring to Figure 1 as shown, it is a step flow schematic diagram of an information technology retrieval consultation method of the present invention. In this example, the information technology retrieval consultation method includes the following steps:

[0074] Step S1: Obtain user input data; extract the modal key information of the user input data to obtain multi-modal key semantic information data; perform data preprocessing on the multi-modal key semantic information data to generate standard user input data;

[0075] In the embodiments of the present invention, the user input data can come from multiple sources and modalities, such as text, speech, images, videos, sensor data, etc. The system first receives the user input data through different input interfaces (such as text boxes, speech recognition, image uploads, etc.). For each input modality, different data collection methods are adopted: collecting the text data input by the user on the interface, converting the speech data into text data through speech recognition technology, and obtaining the image or video file provided by the user through image / video collection tools. Integrate the user input data of different modalities into a unified processing system. At this time, the data has not been preprocessed or parsed and is usually stored in its original format, such as text, image files, speech audio files, etc. Through natural language processing (NLP) technology, extract the key information in the text, such as keyword extraction, entity recognition, syntactic analysis, sentiment analysis, etc. After converting the speech into text using speech recognition technology, further use a semantic understanding model to extract the key information in the speech, such as extracting the theme, sentiment, etc. Use image processing technology and computer vision algorithms (such as convolutional neural network, CNN) to extract key features or objects from the image. For example, image recognition, object detection, scene analysis, etc.; video input involves image processing for each frame and analysis of the inter-frame relationship. Integrate the key information extracted from different modalities. Use multimodal learning technology (such as multimodal neural networks, cross-modal mapping models, etc.) to combine the information of text, images, speech, and other modalities to form multimodal key semantic information data. For example, if the user input contains both text descriptions and images, the system can combine the visual features of the image with the semantic information of the text to provide a richer context understanding. For text in Chinese or other languages, first perform word segmentation to split the sentence into individual words. Remove common words with no practical meaning, such as "de", "le", etc. Label the part-of-speech information for each word, such as noun, verb, adjective, etc. Convert the text into a vector representation (such as using pre-trained models like Word2Vec, BERT, etc.) for subsequent model processing. Standardize the data of different modalities to ensure that they are within the same range, have consistent dimensions and formats. For example, convert images and text into vectors of the same length through predefined models, and speech and sensor data can be converted into numerical vectors through corresponding feature extraction steps. According to the system requirements, organize the preprocessed multimodal data into standard user input data in a unified format, which can be a unified multi-dimensional array, vector set, database record, etc. For example, the standard input data is a comprehensive data structure containing text vectors, image feature vectors, speech feature vectors, etc., for use by subsequent analysis and decision-making modules.

[0076] Step S2: Conduct a historical retrieval information retrieval on the standard user input data. When a historical retrieval record is retrieved, construct an accurate portrait of the user for the standard user input data to generate an accurate retrieval portrait of the user. When no historical record is retrieved, conduct a retrieval inference on the standard user input data to generate a fuzzy retrieval portrait of the user. Construct a user input context graph, and perform cross-query context inheritance on the user input context graph based on the accurate retrieval portrait of the user and the fuzzy retrieval portrait of the user to generate similarity retrieval inheritance data;

[0077] In the embodiments of the present invention, historical retrieval is performed on standard user input data (including user input in modalities such as text, image, and voice). The historical retrieval data includes the user's past behavior data, search records, interaction data, etc. Using technologies such as natural language processing (NLP), speech recognition, and image matching, the standard user input data is matched with the historical data. If a historical record is retrieved, that is, the standard user input data matches the historical retrieval data (there is a matching query record), then step S2.2 is entered. If no matching historical record is found, that is, the standard user input data has no matching item in the historical record, then step S2.3 is entered. Based on the historical retrieval records and user behavior data, analyze the user's search interests, preferences, past retrieval paths, etc. Combining the user's behavior in the historical record (such as clicks, browsing duration, browsing content, etc.), create an accurate user profile for the user. Extract the user's basic attributes from the historical record, such as age, gender, interest tags, usage habits, etc., and integrate this information into the user profile. Analyze the user's potential needs, intentions, and purposes based on the historical record, and add emotional, preference, or intention features related to the user's current needs to the user profile. During the construction of the accurate profile, combine the user's past query records, behavior patterns, and preference data to construct a personalized and accurate user retrieval profile. This profile includes: the user's interest areas and preferences, frequently queried keywords, topics or fields, the user's emotional tendency towards specific content (such as positive or negative emotions), relevant context information (such as time, location, device type, etc.). When the system fails to find a match in the historical retrieval records, the system will perform reasoning based on the current standard user input data, and use reasoning models (such as machine learning and deep learning models) to infer the user's potential needs and interests. By analyzing the context information of the user input (such as the time, topic, emotion, etc. of the current session) and the user's behavior pattern, infer the user's needs or query direction. Based on the data characteristics of the user's current input (such as the semantic similarity of the query, the result of sentiment analysis, etc.), perform fuzzy matching in the known data set or external domain to speculate the user's interests or intentions. In the absence of historical retrieval records, the reasoning process will generate a fuzzy user profile, called the "user fuzzy retrieval profile". According to the standard user input data and its historical record or reasoning result, construct a graph structure containing various context information. The nodes include: keywords, themes, emotions, behavior patterns, etc. input by the user. In the graph, the edges connecting different nodes represent various relationships, such as semantic relationships (such as synonyms, related words, etc.), time relationships (such as the time of the user's last query), context relationships (such as the content of similar queries), etc. Optimize the graph to ensure that the relationships between different modal data input by the user and the historical behavior can be accurately reflected. For example, data input in text, image, and voice can be fused together through technologies such as embedding vectors and graph neural networks to form a multi-modal context graph.Utilize the information in the user's precise retrieval portrait and the user's fuzzy retrieval portrait to perform cross-query inheritance on the context graph input by the user, so as to obtain more accurate query results. Set inheritance rules according to the user's historical query records and inference portrait. For example, if the user has a high interest in a certain field, the nodes related to that field will be weighted in the context graph and passed on to subsequent queries. During the cross-query inheritance process, the system will update the graph nodes according to the precise portrait and the fuzzy portrait. For example, the system will infer the user's continuous interest in certain topics, thereby enhancing the weights of these topics in the current query and promoting the display of relevant content. Calculate the similarity between the current user query and the historical query or other user queries by analyzing the similarity between the historical query and the current input, or based on the information in the user portrait. Based on similarity metrics (such as cosine similarity, Euclidean distance, etc.), match the user input data with the historical records or other relevant data to generate similarity retrieval inheritance data. Provide relevant recommendations, query results, or content displays for the user according to the similarity retrieval inheritance data. For example, if the user's query highly matches the historical data, the system can directly display the relevant results in history; if there is no historical data, recommendations will be made according to the similarity inheritance results.

[0078] Step S3: Obtain external information data; perform dynamic knowledge retrieval on the user input data according to the similarity retrieval inheritance data to generate dynamic knowledge retrieval data; perform multi-knowledge source fusion on the dynamic knowledge retrieval data and the cross-domain similarity retrieval data to generate fusion retrieval data; perform retrieval display adjustment on the fusion retrieval data through a preset sentiment analysis model to generate the final hybrid retrieval display data;

[0079] In the embodiments of the present invention, by selecting external information data sources, including open datasets, third-party databases, industry reports, online APIs, etc., these data sources include various types of data such as text, images, audio, and video. External data is obtained through methods such as web scraping, API calls, and database queries. The obtained data can be structured (such as in JSON or XML format) or unstructured (such as web content, documents, etc.). Using user input data and similarity retrieval inherited data as the input for retrieval, combined with information such as context graphs and user portraits for dynamic retrieval. Through a preset retrieval engine (such as deep learning models based on inverted index, TF-IDF, BERT, etc.), according to the user's query and historical retrieval records, more accurate retrieval is performed through semantic search. Through a cross-domain multi-level knowledge base, dynamic retrieval is performed on the user input data, and the retrieval results can include literature, reports, expert answers, external API data, etc. in known fields. Organize the retrieved relevant data according to certain sorting rules, involving keyword matching, semantic relevance scoring, etc. The dynamic retrieval results can include content such as text, video links, documents, and expert recommendations. According to factors such as the relevance of the data, the urgency of the user's needs, and preferences, screen, sort, and deduplicate the retrieval results to ensure that the most relevant and valuable information is provided to the user. During the knowledge retrieval process, information from different fields and different sources has different forms and structures. First, identify these knowledge sources and label them. For example, retrieval results from external databases, recommendation data from user behavior analysis, inference results from historical data, etc. Use methods such as weighted fusion, priority fusion, and consensus fusion to combine data from different knowledge sources. Specifically, it can: assign different weights to each data source according to the reliability, relevance, and user preferences of different knowledge sources. Give priority to displaying information from high-confidence data sources (such as expert recommendations, official literature, etc.). When multiple data sources provide consistent information, enhance the credibility of this information and give priority to displaying it. Use machine learning (such as random forest, XGBoost, etc.) or deep learning models (such as neural networks, graph convolutional networks, etc.) to automatically fuse multiple data sources and output the final fused data. Perform semantic alignment on data from different knowledge sources to ensure their consistency in the same context and avoid information chaos. Generate the final retrieval results based on the fused data, which include relevant data from multiple knowledge sources and maintain the semantic consistency between each data source. Use a sentiment dictionary (for example, a sentiment-based thesaurus or a sentiment analysis tool) to perform sentiment analysis on the fused retrieval data and identify the sentiment tendency (such as positive, negative, neutral) in the data. On the basis of identifying sentiment words, correct the sentiment vocabulary through a context sentiment analysis model to improve the accuracy of sentiment analysis. For example, some words express different sentiment tendencies in different contexts, so the sentiment results need to be further corrected according to the context.Evaluate the sentiment tendency and sentiment intensity of each retrieval result to ensure that the sentiment data that meets the user's needs is highlighted during display. For example, since the user prefers positive feedback, the weight of positive data can be increased through sentiment intensity evaluation. For data with negative sentiment, the system can lower its priority or attach sentiment adjustment during display (such as reminding the user of negative content). According to the sentiment analysis results, adjust the display order and content of the final fused retrieval data. Through sentiment filtering, present the data that meets the user's current sentiment needs to the user. For example, display data with positive sentiment or filter out negative content according to the user's sentiment needs. During display, the display order of the content can be dynamically adjusted, or even filtered according to the results of sentiment analysis to enhance the user experience. Design a clear and concise display structure based on the fused retrieval results and the adjustments of sentiment analysis. Multiple display methods (such as card type, list type, grid type, etc.) can be used for display according to the data type and user preference. For different modalities of data (such as text, image, video), adopt different display strategies. For example, for text content, use a concise summary; for image or video content, display thumbnails and provide clickable detailed content. Present the final mixed retrieval display data to the user through a preset sorting algorithm (such as sorting adjusted based on sentiment analysis, sorting adjusted based on user portraits, etc.).

[0080] Step S4: Optimize the hybrid retrieval feedback of the final hybrid retrieval display data to perform information technology retrieval consultation optimization operations.

[0081] In the embodiments of the present invention, the content clicked by the user in the displayed data is recorded. The click behavior reflects the user's attention and relevance to specific content, which can be collected through a click tracking system (such as an event listener). The time the user stays on a specific display item is recorded. The length of the stay time is an intuitive reflection of the user's interest in the displayed content, especially for the stay time on detailed information or long documents. The stay time data can be collected through the user behavior tracking tool of the front-end application. Through simple sentiment scoring (such as like, dislike, approve, oppose, etc.), the emotional feedback of the user on the specific displayed content is collected to supplement the data of clicks and stay times. If the system allows, the user can also provide more specific opinions in the form of feedback forms, questionnaires, real-time scoring, etc. to help the system understand the user's needs. The user click feedback, stay time feedback, and emotional feedback are integrated together to generate user consultation feedback data. Feedback scores are generated for each display item according to different dimensions of the feedback data (such as the number of clicks, stay time, emotional tendency), and based on this, the user's interest and satisfaction with the retrieval results are analyzed. Analyze the display items clicked by the user and their relevance, and identify which display items have a high click-through rate. For these display items with high clicks, the system can increase their display weight to increase their appearance frequency in future retrieval results. Associate the stay time with the relevant content of the display item. The content with a long stay time represents the user's in-depth attention to this content. The system can mark the display items with a long stay time to improve their display priority. By analyzing the emotional feedback (such as the distribution of positive and negative emotions), the emotional preferences of the user are identified. For example, when the user shows obvious negative emotions towards certain types of data, the system can adjust the display frequency of this type of data accordingly. Through a weighting mechanism, different weights are assigned to the click feedback, stay time feedback, and emotional feedback to ensure that important feedback factors have a greater impact on the optimization process. For example, higher weights can be given to the display items with longer stay times and click feedback, so as to strengthen the display of such content in future retrievals. Perform dynamic optimization according to the feedback data and fine-tune the current hybrid retrieval display data.

[0082] Preferably, step S1 includes the following steps:

[0083] Step S11: Obtain user input data, where the user input data includes text input data, image input data, and voice input data;

[0084] Step S12: Extract modal key information from the text input data, image input data, and voice input data to obtain multi-modal key information extraction data;

[0085] Step S13: Perform unified semantic representation on the multi-modal key information extraction data to generate multi-modal key semantic information data;

[0086] Step S14: Perform data preprocessing on the multi-modal key semantic information data to generate standard user input data, where the data preprocessing includes data missing value prompting, data abnormal format conversion, and data duplicate value elimination.

[0087] In the embodiment of the present invention, by collecting the input data of the user, which includes three types of data: the text input data is obtained through text boxes, speech recognition transcription, etc., the image input data is obtained by real-time collection of the pictures uploaded by the user or the camera, and the voice input data is converted from the sound input from the microphone into text through speech recognition technology. Key information such as keywords, entities, and emotions in the text input data is extracted through natural language processing (NLP) technology. Key features such as objects, scenes, colors, and shapes in the image are extracted using computer vision technology (such as convolutional neural networks). Keywords, emotional colors, and emotional states in the speech are extracted through speech recognition and audio processing technology. For text, image, and voice data, a unified semantic representation technology (such as a multi-modal fusion model based on deep learning) is used to map information in different modalities into a common semantic space, which can uniformly process the semantics of various types of data, enabling information in different modalities to work together and complement each other. If there are missing values in any input data, the system will issue a warning and prompt the user to supplement it. For input data with inconsistent formats, the system will automatically perform format conversion, such as converting non-standard date formats into a unified format, or cleaning special characters in the text. For duplicate data in the user input, the system will automatically detect and remove redundant duplicate items to ensure the uniqueness and accuracy of the data.

[0088] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes:

[0089] Step S21: Perform historical retrieval information retrieval on the standard user input data. When a historical retrieval record is retrieved, the user's historical retrieval data is retrieved, and a user profile is constructed based on the user's historical retrieval data to generate an accurate user retrieval profile.

[0090] Step S22: When no historical record is retrieved, perform user intention recognition on the standard user input data to obtain user intention recognition data; use the user intention recognition data to perform retrieval reasoning on the standard user input data to generate a fuzzy user retrieval profile.

[0091] Step S23: Construct a context graph for the standard user input data to obtain a user input context graph; dynamically update the context complexity weight of the user input context graph based on the accurate user retrieval profile and the fuzzy user retrieval profile to generate context graph dynamic update data.

[0092] Step S24: Calculate the correlation edge nodes of the atlas context dynamic update data, and perform cross-query context inheritance on the user input context atlas through the correlation edge nodes to generate similarity retrieval inheritance data.

[0093] In the embodiment of the present invention, by obtaining standard user input data, which includes search keywords, user interests, browsing history, and other personal preference data. Using the standard user input data, historical retrieval queries are performed. Historical data is stored in a database (such as the user's previous search records, behavior data, known preferences, etc.). When there are historical retrieval records in the query results, the relevant historical data is retrieved and analyzed, and this data will be used to construct a user profile. According to the historical retrieval data, a precise user retrieval profile is generated using machine learning or statistical models, which specifically reflects the user's long-term interests, search behaviors, and historical query information related to the user. If no relevant information is retrieved in the historical records, user intent recognition is performed based on the standard user input data. Intent recognition can adopt natural language processing (NLP) methods, such as intent classification, sentiment analysis, etc., to identify the user's basic needs (for example, the user is looking for a certain product, service, information, or answer). By analyzing the user intent, an inference mechanism is used to infer the standard user input data, and a fuzzy user retrieval profile is generated in combination with the inference results, which can be a preliminary profile based on the user's current input and predicted needs. Based on the standard user input data and its corresponding background information (such as time, location, user environment, etc.), a user input context atlas is constructed. This step involves converting the input data and context information (for example, the time, location, and device of the user's query) into an atlas structure, which can be achieved through a graph database or a graph neural network. Using the precise user retrieval profile and the fuzzy user retrieval profile, dynamic update of the complexity weights of the atlas context is performed. The purpose of this step is to dynamically adjust the structure of the atlas and the weights of each node and edge according to the user's input and profile to reflect the user's current focus, interests, and query purposes. Based on the atlas context dynamic update data, the correlation edge nodes between each node and edge are calculated. This step can be calculated through atlas algorithms (such as PageRank, HITS algorithm, etc.) to calculate the relationships and importance between nodes. Through the calculated correlation edge nodes, cross-query context inheritance is performed, that is, relevant query histories or similar context information are inherited into the new query, which will generate similarity retrieval inheritance data, enabling the user to use the previous query information for more precise retrieval in subsequent retrieval processes.

[0094] Preferably, step S22 includes the following steps:

[0095] Step S221: When no historical record is retrieved, perform keyword part-of-speech extraction on the standard user input data to obtain keyword part-of-speech extraction data; perform sentiment classification on the keyword part-of-speech extraction data to generate keyword sentiment classification data;

[0096] Step S222: Perform a tendency analysis on the standard user input data through the keyword sentiment classification data to generate user intention recognition data; classify the user intention recognition data to obtain user intention classification data;

[0097] Step S223: Perform behavior matching based on the user intention classification data to obtain user behavior data; perform user preference association reasoning based on the user behavior data to obtain user preference reasoning data;

[0098] Step S224: Cluster users based on the user preference reasoning data to obtain a user fuzzy retrieval profile.

[0099] In the embodiments of the present invention, by performing part-of-speech tagging on standard user input data, natural language processing (NLP) technology is used to analyze the part of speech of each word. For example, using a dependency parsing tool, nouns, verbs, adjectives, etc. in the input data are extracted. Next, representative and high-frequency keywords are selected from them to form "keyword part-of-speech extraction data". For example, verbs or nouns are highly important for user intentions. Sentiment analysis is performed on the extracted keywords. Through a pre-trained sentiment classification model (such as LSTM or BERT), the sentiment tendency is judged according to the data input by the user, such as positive, negative or neutral. The analysis results are formed into "keyword sentiment classification data" to provide a preliminary understanding of the user's emotions and support for subsequent user intention recognition and behavior matching. Using the sentiment classification data, tendency analysis is performed on the data input by the user to identify the user's sentiment tendency and potential intentions. For example, a positive sentiment indicates interest, and a negative sentiment represents complaints or dissatisfaction. By combining sentiment and part of speech, "user intention recognition data" is generated to describe the sentiment and goals of the user input. The user intentions are classified according to the intention characteristics in the user input. For example, according to the results of sentiment analysis and the recognition of key terms, user intentions can be classified into "purchase intention", "query intention", "complaint intention", etc., to generate "user intention classification data", which provides a basis for subsequent behavior matching and personalized service recommendation. According to the user's intention classification data, corresponding historical behavior data is matched. The user's historical behavior data (such as click records, search history, purchase records, etc.) can be used to find behavior patterns related to the current intention. Through algorithms (such as collaborative filtering, content recommendation algorithms, etc.), "user behavior data" is generated to understand the user's current needs. Reasoning is performed based on the user behavior data, and machine learning or data mining techniques (such as decision trees, clustering analysis, etc.) are used to deduce the user's potential preferences to generate "user preference inference data", which provides an accurate basis for personalized recommendation systems, advertising placement, or user service customization. Based on the "user preference inference data", clustering analysis is performed on the users. Through unsupervised learning algorithms (such as K-means, DBSCAN, etc.), users are divided into different categories. Fuzzy matching is performed on different user categories to construct a "user fuzzy retrieval portrait", which not only includes the user's explicit preferences, but also reflects the user's potential interests and needs.

[0100] Preferably, step S24 includes the following steps:

[0101] Step S241: Based on the graph context dynamic update data, initial node screening is performed on the user input context graph to obtain initial screening nodes; the graph algorithm is used to perform boundary point correlation scoring on the initial screening nodes to obtain the correlation score of each initial screening node;

[0102] Step S242: Screen out the edge nodes with the highest relevance according to the relevance scores of each initial screening node, and mark them as relevant edge nodes; Screen the initial screening nodes through the relevant edge nodes to obtain adjacent associated screening nodes;

[0103] Step S243: Conduct node directivity analysis on the adjacent associated screening nodes to generate adjacent associated screening node directivity data; Based on the adjacent associated screening node directivity data, perform context node marking on the adjacent associated screening nodes to obtain upstream associated nodes and downstream associated nodes;

[0104] Step S244: Calculate the node information entropy value of the relevant edge nodes through the upstream associated nodes and the downstream associated nodes to obtain the standard node information entropy value; Perform cross-query context inheritance on the user input context graph through the standard node information entropy value to generate similarity retrieval inheritance data.

[0105] In the embodiments of the present invention, the initial nodes of the user input context graph are screened by dynamically updating data based on the graph context. By analyzing the user's context data (such as the user's search terms, input text, behavior data, etc.), the relevant initial nodes in the graph are extracted. Graph algorithms (such as graph neural networks, Dijkstra's algorithm, etc.) are used to perform boundary point correlation scoring on the initially screened nodes to measure the correlation between each initial node and the input context. For example, the correlation can be evaluated by calculating the similarity between nodes (such as cosine similarity, Euclidean distance, etc.) to obtain the correlation score of each initially screened node. The nodes most relevant to the user input context are selected according to the correlation score to obtain a set of "initially screened nodes" as the basis for further analysis. According to the correlation score of each initially screened node, the edge nodes with the highest correlation are screened out (i.e., the nodes most strongly associated with the initially screened nodes), and these edge nodes represent other nodes closely related to the initial nodes and are the key to further expanding the graph. These screened edge nodes are marked as "correlation edge nodes" to provide a basis for subsequent association analysis. Based on the screened correlation edge nodes, the edge-connected nodes of the initially screened nodes are screened. That is, the nodes connected by the correlation edge nodes form "adjacent associated screened nodes", and these nodes have a direct or indirect relationship with the initially screened nodes in the graph. The node directionality analysis is performed on the adjacent associated screened nodes to analyze their structural positions and relationship directions in the graph (such as parent-child relationship, front-back relationship, causal relationship, etc.). For example, the directed edges in the graph are used to analyze the relative directionality of the nodes to generate "adjacent associated screened node directionality data", that is, to record the direction information of each adjacent node, such as whether it is an "above-text" node (i.e., having an upward relationship such as causal, historical, time, etc. with the current node) or a "below-text" node (i.e., having an influencing or subsequent relationship with the current node). Based on the directionality data of the adjacent associated screened nodes, the context nodes of the adjacent associated screened nodes are marked, and this step can be performed by setting specific rules (such as time sequence, semantic association, etc.) to mark the nodes as "above-text associated nodes" or "below-text associated nodes". Through such marking, the nodes in the graph are distinguished and classified according to the context relationship. By analyzing the above-text associated nodes and below-text associated nodes, the node information entropy value of the correlation edge nodes is calculated. The information entropy value can be used to measure the amount of information and uncertainty of the nodes in the graph. Generally, the higher the node information entropy value, the richer the information represented by the node and the greater the uncertainty. The standard node information entropy value is calculated to evaluate the importance of the graph nodes and their roles in the graph. Based on the standard node information entropy value, cross-query context inheritance is performed, which means that the information entropy value of the existing nodes is used to reason and match the new user query context to ensure the inheritance of relevant context information from the graph.Through the inheritance process, "similarity retrieval inheritance data" is generated, that is, according to information entropy and node relevance, the nodes and information related to the user query in the graph are extended to provide more accurate context support for subsequent retrievals.

[0106] Preferably, cross-query context inheritance of the user input context graph through the node information entropy value includes:

[0107] Perform context topology structure analysis on the user input context graph to obtain context structure topology data; extract semantic inheritance relationship data from the context structure topology data to generate semantic inheritance relationship data;

[0108] Perform query pattern matching on the semantic inheritance relationship data through the node information entropy value to generate query pattern matching data; perform propagation path analysis on the user input context graph according to the query pattern matching data to generate context query propagation path data;

[0109] Use the standard node information entropy value to verify the cross-node inheritance rules for the context query propagation path data. When the information entropy value of the nodes in the context query propagation path is less than or equal to the standard node information entropy value, similarity retrieval inheritance data is generated; otherwise, the corresponding context query propagation path is excluded.

[0110] In the embodiments of the present invention, by performing a topological structure analysis on the context graph input by the user. Through graph algorithms, the relationships and connection patterns between nodes in the graph are identified, such as parent-child relationships, sibling relationships, dependency relationships, etc., to generate "context structure topological data", which records each node in the graph and its connection method. This topological data can include the hierarchical relationship of nodes, adjacent node information, and the weights of the edges between nodes. Based on the context structure topological data, semantic inheritance relationships are extracted. The semantic inheritance relationship refers to the inheritance and extension at the semantic level between nodes in the context graph. For example, a node represents a certain concept, and its related child nodes inherit some attributes or semantics of this concept to generate "semantic inheritance relationship data", including the semantic connections between the node and its directly or indirectly related nodes. These data help to understand the actual meaning of the nodes in the context graph and the relationships between them. The node information entropy value is used to perform query pattern matching on the "semantic inheritance relationship data". According to the query statement or context information input by the user, the relevant node patterns in the graph are matched to ensure that the nodes in the graph highly match the intention and requirements of the user's query, generating "query pattern matching data", including all nodes that meet the query conditions and their relationships. These data are used for subsequent context propagation path analysis to ensure the accuracy and relevance of the query. According to the generated query pattern matching data, the propagation path of the user input context graph is analyzed. The propagation path analysis refers to starting from the initial node input by the user and analyzing how information flows and spreads in the graph based on the connections and inheritance relationships in the graph, generating "context query propagation path data", which records the propagation order of each node and path and their related information during the query process. This helps to identify which paths can best meet the query requirements. The standard node information entropy value is used to verify the cross-node inheritance rules for the "context query propagation path data". Each node in the propagation path is compared with the standard node information entropy value according to its information entropy value. When the information entropy value of a certain node is less than or equal to the standard node information entropy value, it is considered that the node has high reliability in the context propagation path and can be retained. If the information entropy value of the node is greater than the standard node information entropy value, it is considered that the information volume of the node in the propagation path is low or the noise is large, resulting in inaccurate retrieval, so it should be excluded. If the nodes in the propagation path meet the inheritance rules (that is, the node information entropy value is less than or equal to the standard node information entropy value), then the node is continued to be retained and "similarity retrieval inheritance data" is generated. This data includes the nodes and paths most relevant to the user's query and can provide high-quality and personalized retrieval results. If the node information entropy value of a certain propagation path does not meet the rules, the corresponding context query propagation path is excluded. This process ensures the accuracy and effectiveness of the node information in the query path.

[0111] As an example of the present invention, refer to Figure 3 shown, in this example, step S3 includes:

[0112] Step S31: Obtain external information data by using an external domain knowledge base;

[0113] Step S32: Perform cross-domain similarity retrieval on the user input data according to the similarity retrieval inheritance data to obtain cross-domain similarity retrieval data; perform retrieval domain analysis on the cross-domain similarity retrieval data to generate retrieval domain analysis data;

[0114] Step S33: Perform dynamic knowledge retrieval on the cross-domain similarity retrieval data based on the retrieval domain analysis data to generate dynamic knowledge retrieval data; perform multi-knowledge source fusion on the dynamic knowledge retrieval data and the cross-domain similarity retrieval data to generate fusion retrieval data;

[0115] Step S34: Perform user input context sentiment analysis on the standard user input data through a preset sentiment analysis model to generate user input context sentiment data; use the user input context sentiment data to adjust the retrieval display of the fusion retrieval data to generate the final hybrid retrieval display data.

[0116] In the embodiments of the present invention, external information related to the user input data is obtained by leveraging external domain knowledge bases (such as academic databases, industry-specific data sources, network resources, open APIs, etc.). These knowledge bases can be structured (such as databases) or unstructured (such as texts, articles, reports, etc.). A query mechanism or API call is used to extract relevant information from the external knowledge base, and it is combined with the current user input data for matching and filtering to obtain relevant "external information data". The external information data can include expert suggestions, industry reports, historical data, research literature, etc., providing basic information for subsequent similarity retrieval and knowledge fusion. Cross-domain similarity retrieval is performed on the user input data according to the "similarity retrieval inheritance data". This step utilizes known cross-domain data mapping, semantic matching, and other technologies to compare the user input data with data in multiple domains and find similar data items. For example, if the user's query involves multiple domains (such as medicine and engineering), cross-domain technologies (such as transfer learning, semantic mapping, etc.) are used to find information similar to the user's query from knowledge bases in different domains. Retrieval domain analysis is performed on the obtained "cross-domain similarity retrieval data". That is, it is analyzed which domain the retrieval results belong to, the domain labels and related attributes of each retrieval data are determined, and "retrieval domain analysis data" is generated, which includes information such as the domain, category, and relevance score corresponding to each retrieval result. This step helps to understand the diversity of the retrieval results and which domain information is the most relevant. Based on the "retrieval domain analysis data", dynamic knowledge retrieval is performed on the "cross-domain similarity retrieval data". By dynamically adjusting the query method (for example, through changes in query syntax, adjustment of real-time data feedback, context-related reasoning, etc.), more domain knowledge related to the user query is obtained, and "dynamic knowledge retrieval data" is generated, which will include knowledge points further optimized according to the user's needs and highly relevant to the user input. Multi-knowledge source fusion is performed on the "dynamic knowledge retrieval data" and the "cross-domain similarity retrieval data". The fusion can be carried out through weighted fusion, priority ranking, model fusion, etc., in order to integrate the knowledge of multiple data sources, eliminate information conflicts, enhance the comprehensiveness and accuracy of the data, and generate "fusion retrieval data", which combines retrieval results and knowledge from multiple domains and gives weighted scores according to the relevance of different domains. Using a preset sentiment analysis model, user input context sentiment analysis is performed on the "standard user input data". The sentiment analysis model can be a machine learning-based (such as deep learning models, support vector machines, etc.) or rule-based sentiment classification model, analyzing the user's emotions, intentions, preferences, etc., and generating "user input context sentiment data", including the user's sentiment tendency (such as positive, negative, neutral), sentiment intensity, emotion type (such as joy, anger, etc.) and other data. According to the "user input context sentiment data", the display of the "fusion retrieval data" is adjusted. According to the user's sentiment tendency, the display method of the retrieval results is adjusted.For example, when the user shows positive emotions, retrieve results with strong relevance, good feedback, and more positive content are preferentially presented; when the user shows negative emotions, more soothing and suggestive information is provided to generate "final mixed retrieval display data", which contains search results adjusted by emotion, aiming to improve user satisfaction and optimize the user experience.

[0117] Preferably, the user input context emotion analysis of the standard user input data by a preset emotion analysis model includes:

[0118] Perform emotion dictionary matching on the standard user input data to generate emotion dictionary matching data; extract emotion vocabulary from the standard user input data according to the emotion dictionary matching data to obtain emotion vocabulary data;

[0119] Perform emotion polarity analysis based on the emotion vocabulary data, and judge the emotion tendency in the standard user input data to obtain emotion polarity data; perform context emotion enhancement on the emotion polarity data by a preset emotion analysis model to generate context emotion enhancement data;

[0120] Use the context emotion enhancement data to evaluate the emotion intensity of the standard user input data to generate user input context emotion data.

[0121] In the embodiments of the present invention, by selecting a suitable sentiment dictionary, such as the commonly used SentiWordNet, sentiment dictionaries (such as Baidu sentiment dictionary in Chinese, or other sentiment dictionaries for specific fields). The sentiment dictionary contains a large number of words with sentiment attributes, and attaches a sentiment label (such as "positive", "negative", "neutral") and its intensity value to each word. Match the words in the standard user input data with the entries in the sentiment dictionary. Through natural language processing techniques (such as word segmentation, part-of-speech tagging, etc.), extract the keywords in the input data, and then compare them with the sentiment dictionary to find the sentiment words. After matching, generate sentiment dictionary matching data, which includes all the matching sentiment words in the input data and their corresponding sentiment polarity labels (such as "positive", "negative") and intensities (such as strong, medium, weak). According to the matched sentiment dictionary matching data, extract the sentiment-related words from the standard user input data. Sentiment words include words expressing emotions, feelings, attitudes, evaluations, etc., such as "happy", "sad", "satisfied", etc., and generate sentiment word data, including all the sentiment words in the input data and their context positions in the sentence. Filter the extracted sentiment words to remove irrelevant words (such as neutral words, function words without sentiment). Through part-of-speech tagging and semantic analysis, ensure that the extracted sentiment words can truly reflect the sentiment state of the user input. Based on the extracted sentiment word data, perform sentiment polarity analysis. Sentiment polarity analysis is the process of determining the sentiment tendency, that is, judging whether a sentiment word belongs to a "positive" or "negative" sentiment. Sentiment analysis algorithms (such as Naive Bayes, Support Vector Machine, deep learning models, etc.) can be used to classify the sentiment words to obtain sentiment polarity data. The sentiment polarity data includes the polarity (positive, negative or neutral) and intensity of each sentiment word. According to the polarity information of the sentiment words, judge the sentiment tendency in the entire standard user input data. Usually, the sentiment tendency of the input data is determined by the weighted average of the sentiment words. If most of the sentiment words tend to be "positive", the overall sentiment is positive, and vice versa, and generate sentiment polarity data, including the classification label of the sentiment tendency (positive, negative, mixed, etc.) and the intensity of this sentiment (such as strong, medium, slight). In sentiment analysis, the sentiment value of a single sentiment word varies with the context, so context needs to be combined for sentiment enhancement. Context sentiment enhancement adjusts the sentiment intensity and polarity of sentiment words through context information. For example, if a sentiment word has a stronger sentiment value in certain specific contexts, its weight can be increased according to the context. Use a preset sentiment analysis model (such as context-aware deep learning models like LSTM, BERT, etc.) to enhance the sentiment polarity data. Through context relationship analysis, increase or adjust the weight of sentiment words. Through context sentiment enhancement, generate enhanced context sentiment enhancement data, which includes the adjusted sentiment intensity and sentiment polarity after context analysis.For example, in positive sentiment sentences, enhance the positive intensity of sentiment words; in negative sentiment sentences, enhance the negative sentiment intensity. Based on the context sentiment enhancement data, conduct sentiment intensity evaluation on the standard user input data. This step is the final quantification and evaluation of the sentiment tendency and sentiment intensity in the user input data. Sentiment intensity evaluation can use methods such as weighted average method, sentiment analysis models (such as deep neural networks), etc., to perform weighted processing on each sentiment word in the input data, and finally obtain the comprehensive sentiment intensity of the entire input data. According to the results of the sentiment intensity evaluation, generate the final user input context sentiment data. This data contains the detailed evaluation results of the sentiment tendency (such as positive, negative, mixed, etc.) and sentiment intensity (such as strong, medium, weak) of the user input. This sentiment data can reflect the user's emotional state in a specific context and provide important emotional information basis for subsequent recommendations, feedback, etc.

[0122] Preferably, step S4 includes the following steps:

[0123] Step S41: Collect user consultation feedback on the final mixed retrieval display data to obtain user consultation feedback data, where user consultation feedback collection includes click feedback collection and dwell time feedback collection;

[0124] Step S42: Optimize the final mixed retrieval display data based on the user consultation feedback data to perform information technology retrieval consultation optimization operations.

[0125] In the embodiments of the present invention, during the interaction process between the user and the system, two main types of feedback data are collected: Record the user's click behavior on different items in the search results, including the content clicked, the click order, the click frequency, etc. The click feedback reflects the user's attention and interest in certain information, and is an important basis for analyzing user needs and optimizing search results. Record the user's stay time on each search result or page. The longer the stay time, it indicates that the user is interested in the content, or it takes more time to understand the content. This data can help analyze which content is most attractive or useful to the user. Use the front-end interface and user interaction tools (such as buttons, links, page jumps, pop-ups, etc.) to obtain click feedback. Use page analysis tools (such as Google Analytics or custom page tracking scripts) to capture the user's stay time. These tools will record the stay duration of each page or content, so as to provide data support for subsequent analysis. The collected click feedback and stay time feedback data need to be preprocessed and cleaned to remove invalid and duplicate records to ensure the accuracy and effectiveness of the data. Format the data to ensure that it conforms to the data structure required for subsequent analysis and is convenient for subsequent processing. According to the user's click feedback and stay time feedback data, optimize the "final hybrid retrieval display data". The goal of this step is to dynamically adjust the search results according to the user's actual behavior to improve relevance and user experience. By analyzing the frequency and order of user clicks, identify which results the user is more interested in. For these frequently clicked results, increase their priority or weight in future retrievals. Items with a high number of clicks can be considered as more attractive content to the user, so more displays are given in subsequent retrievals. According to the time the user stays on the page, analyze which content attracts the user to stay for a long time. These contents can be improved in terms of their exposure rate and attractiveness by increasing their display frequency and improving their presentation form (such as adding summary information, adding recommended tags, etc.). On the basis of multi-knowledge source fusion, combine user behavior data to adjust the retrieval algorithm, and adopt a weighted model based on user feedback. According to data such as the user's click-through rate and stay time, dynamically adjust the weights and sorting rules of each search result item. Introduce a reinforcement learning or online learning model, and continuously receive and analyze user feedback data to optimize the retrieval display strategy in real time. This way can enable the system to gradually adapt to the changes in the user's interests and improve the accuracy and personalization level of the search results. According to the optimized hybrid retrieval results, perform the retrieval consultation optimization operation of information technology. Specifically: In the information retrieval system, readjust the database index and query algorithm so that the optimized content is displayed more preferentially during query; in the user consultation interface, adjust the display strategy, reorder the retrieval items according to the optimization results, and enable the user to first see the results that meet their preferences; according to the user consultation feedback, adjust the node weights in the knowledge graph to reflect the user's actual needs and concerns.

[0126] In this specification, an information technology retrieval and consultation system is provided for performing the above-mentioned information technology retrieval and consultation method. The information technology retrieval and consultation system includes:

[0127] A data processing module for obtaining user input data; extracting modal key information of the user input data to obtain multi-modal key semantic information data; and performing data preprocessing on the multi-modal key semantic information data to generate standard user input data;

[0128] A retrieval inheritance module for performing historical retrieval information retrieval on the standard user input data. When a historical retrieval record is found, an accurate portrait of the user is constructed to generate a user accurate retrieval portrait; when no historical record is found, retrieval reasoning is performed on the standard user input data to generate a user fuzzy retrieval portrait; constructing a user input context graph, and performing cross-query context inheritance on the user input context graph according to the user accurate retrieval portrait and the user fuzzy retrieval portrait to generate similarity retrieval inheritance data;

[0129] A fusion retrieval module for obtaining external information data; performing dynamic knowledge retrieval on the user input data according to the similarity retrieval inheritance data to generate dynamic knowledge retrieval data; performing multi-knowledge source fusion on the dynamic knowledge retrieval data and cross-domain similarity retrieval data to generate fusion retrieval data; and adjusting the retrieval display of the fusion retrieval data through a preset sentiment analysis model to generate final hybrid retrieval display data;

[0130] A retrieval feedback module for optimizing the hybrid retrieval feedback of the final hybrid retrieval display data to perform information technology retrieval and consultation optimization operations.

[0131] The beneficial effects of the present invention are as follows: By obtaining user input and extracting multi-modal key information, converting different forms of user input (such as text, images, voice, etc.) into a standardized data format, ensuring the consistency and processability of the input data. This preprocessing process helps the subsequent retrieval system to more accurately understand user needs and provides a reliable data basis for precise retrieval and reasoning. By using historical retrieval records and reasoning to generate accurate or fuzzy retrieval portraits of users, personalized requirement analysis is achieved. When there is no historical record, a fuzzy portrait is generated through reasoning to ensure that the system can handle the situation of new users or small amounts of data. At the same time, through cross-query context inheritance technology, the relevance in continuous queries is enhanced, ensuring context consistency during multiple interactions, and the retrieval accuracy is further enhanced through similarity retrieval inheritance data. Through dynamic knowledge retrieval and cross-domain information fusion, relevant information can be obtained from multiple knowledge sources, improving the breadth and depth of retrieval. The fusion of multiple knowledge sources avoids the problem of information silos and increases the system's coverage ability for multi-domain data. At the same time, the introduction of the sentiment analysis model enables the retrieval results to be adjusted according to the user's sentiment tendency, thus enhancing the personalization of the retrieval results and the user experience. Through retrieval feedback optimization, this module can dynamically adjust the retrieval strategy and results according to the user's feedback, improving the intelligence and adaptability of the retrieval system. The continuous optimization process ensures that the system is continuously improved, improving the relevance, accuracy, and real-time nature of the query results, while enhancing the system's learning ability, enabling future retrievals to better meet user needs. Therefore, the present invention solves the problems of low retrieval accuracy, incomplete information coverage, and poor user experience in the prior art by combining multi-modal information extraction, accurate and fuzzy portrait construction, cross-context inheritance, dynamic knowledge fusion, and sentiment analysis optimization.

[0132] Therefore, from any perspective, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, it is intended to encompass all changes that fall within the meaning and scope of the equivalent elements of the application documents within the present invention.

[0133] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features disclosed herein.

Claims

1. An information technology search and consultation method, characterized in that: The following steps are involved: Step S1: obtaining user input data; extracting modal key information of the user input data to obtain multimodal key semantic information data; performing data preprocessing on the multimodal key semantic information data to generate standard user input data; Step S2: Perform historical search information retrieval on the standard user input data. When the existing historical search records are retrieved, an accurate portrait is constructed for the standard user input data to generate a user accurate search portrait; when no existing historical records are retrieved, search reasoning is performed on the standard user input data to generate a user fuzzy search portrait; a user input context map is constructed, and cross-query context inheritance is performed on the user input context map based on the user accurate search portrait and the user fuzzy search portrait to generate similarity search inheritance data; Step S3: Acquire external information data; Perform dynamic knowledge retrieval on user input data according to similarity retrieval inheritance data to generate dynamic knowledge retrieval data; Perform multi-knowledge source fusion on dynamic knowledge retrieval data and cross-domain similarity retrieval data to generate fused retrieval data; perform retrieval display adjustment on the fused retrieval data through a preset sentiment analysis model to generate the final hybrid retrieval display data; Step S4: Perform hybrid search feedback optimization on the final hybrid search display data to perform information technology search consulting optimization operations.

2. The information technology search and consultation method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: obtaining user input data, wherein the user input data includes text input data, image input data and voice input data; Step S12: extracting modal key information from the text input data, the image input data, and the voice input data to obtain multimodal key information extraction data; Step S13: performing unified semantic representation on the multimodal key information extraction data to generate multimodal key semantic information data; Step S14: preprocess the multimodal key semantic information data to generate standard user input data, wherein the data preprocessing includes data missing value prompts, data abnormal format conversion, and data duplicate value removal.

3. The information technology search and consultation method according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: historical search information is searched for the standard user input data. When historical search records are found, the user's historical search data is retrieved, and a user profile is constructed based on the user's historical search data to generate an accurate user search profile. Step S22: when no existing historical records are retrieved, user intent recognition is performed on the standard user input data to obtain user intent recognition data; the user intent recognition data is used to perform search reasoning on the standard user input data to generate a user fuzzy search portrait; Step S23: construct a context graph for standard user input data to obtain a user input context graph; dynamically update the graph context complexity weight of the user input context graph based on the user precise retrieval portrait and the user fuzzy retrieval portrait to generate graph context dynamic update data; Step S24: Calculate the correlation edge nodes of the graph context dynamic update data, and perform cross-query context inheritance on the user input context graph through the correlation edge nodes to generate similarity retrieval inheritance data.

4. The information technology search and consultation method according to claim 3, characterized in that: Step S22 includes the following steps: Step S221: when no existing historical records are retrieved, keyword part-of-speech extraction is performed on the standard user input data to obtain keyword part-of-speech extraction data; sentiment classification is performed on the keyword part-of-speech extraction data to generate keyword sentiment classification data; Step S222: performing a tendency analysis on the standard user input data through the keyword sentiment classification data to generate user intention recognition data; performing intent classification on the user intention recognition data to obtain user intention classification data; Step S223: performing behavior matching according to the user intention classification data, thereby obtaining user behavior data; performing user preference association inference according to the user behavior data, thereby obtaining user preference inference data; Step S224: Cluster users based on user preference inference data to obtain user fuzzy retrieval portraits.

5. The information technology search and consultation method according to claim 4, characterized in that: Step S24 includes the following steps: Step S241: Performing initial node screening on the user input context graph based on the graph context dynamic update data to obtain initial screening nodes; performing boundary point relevance scoring on the initial screening nodes using the graph algorithm to obtain a relevance score for each initial screening node; Step S242: Filter out the edge nodes with the highest correlation according to the correlation score of each initial screening node, and mark them as correlation edge nodes; perform edge connection node screening on the initial screening nodes through the correlation edge nodes to obtain adjacent associated screening nodes; Step S243: performing node directionality analysis on the adjacent associated screening nodes to generate adjacent associated screening node directionality data; performing context node marking on the adjacent associated screening nodes based on the adjacent associated screening node directionality data to obtain a previous associated node and a following associated node; Step S244: Calculate the node information entropy value of the correlation edge node through the previous associated node and the following associated node to obtain the standard node information entropy value; perform cross-query context inheritance on the user input context graph through the standard node information entropy value to generate similarity retrieval inheritance data.

6. The information technology search and consultation method according to claim 5, characterized in that: The cross-query context inheritance of the user input context graph through the node information entropy value includes: Performing context topological structure analysis on the user input context graph to obtain context structure topological data; extracting semantic inheritance relationship from the context structure topological data to generate semantic inheritance relationship data; Perform query pattern matching on semantic inheritance relationship data through node information entropy value to generate query pattern matching data; perform propagation path analysis on user input context graph based on query pattern matching data to generate context query propagation path data; The cross-node inheritance rules of context query propagation path data are verified using the standard node information entropy value. When the information entropy value of the node in the context query propagation path is less than or equal to the standard node information entropy value, similarity retrieval inheritance data is generated; if not, the corresponding context query propagation path is eliminated.

7. The information technology search and consultation method according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: using the external domain knowledge base to obtain external information data; Step S32: performing a cross-domain similarity search on the user input data according to the similarity search inheritance data to obtain cross-domain similarity search data; performing a search field analysis on the cross-domain similarity search data to generate search field analysis data; Step S33: Perform dynamic knowledge retrieval on the cross-domain similar retrieval data based on the retrieval domain analysis data to generate dynamic knowledge retrieval data; perform multi-knowledge source fusion on the dynamic knowledge retrieval data and the cross-domain similar retrieval data to generate fused retrieval data; Step S34: Perform user input context sentiment analysis on standard user input data through a preset sentiment analysis model to generate user input context sentiment data; use the user input context sentiment data to perform retrieval display adjustment on the fused retrieval data to generate final hybrid retrieval display data.

8. The information technology search and consultation method according to claim 7, characterized in that: The user input context sentiment analysis of standard user input data using a preset sentiment analysis model includes: Performing sentiment dictionary matching on standard user input data to generate sentiment dictionary matching data; performing sentiment vocabulary extraction on standard user input data according to the sentiment dictionary matching data to obtain sentiment vocabulary data; Perform sentiment polarity analysis based on sentiment vocabulary data and determine the sentiment tendency in standard user input data, thereby obtaining sentiment polarity data; perform contextual sentiment enhancement on the sentiment polarity data through a preset sentiment analysis model to generate contextual sentiment enhancement data; Contextual sentiment enhancement data is used to evaluate the sentiment intensity of standard user input data and generate user input contextual sentiment data.

9. The information technology search and consultation method according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: collecting user consultation feedback on the final hybrid search display data to obtain user consultation feedback data, wherein the user consultation feedback collection includes click feedback collection and dwell time feedback collection; Step S42: Perform hybrid retrieval optimization on the final hybrid retrieval display data based on the user consultation feedback data to perform information technology retrieval consultation optimization operations.

10. An information technology retrieval and consultation system, characterized in that: Used to execute the information technology search and consultation method as claimed in claim 1, the information technology search and consultation system comprises: The data processing module is used to obtain user input data; extract the modal key information of the user input data to obtain multimodal key semantic information data; perform data preprocessing on the multimodal key semantic information data to generate standard user input data; The retrieval inheritance module is used to retrieve historical retrieval information of standard user input data. When historical retrieval records are found, an accurate portrait of the standard user input data is constructed to generate a user accurate retrieval portrait; when no historical records are found, retrieval reasoning is performed on the standard user input data to generate a user fuzzy retrieval portrait; a user input context map is constructed, and cross-query context inheritance is performed on the user input context map based on the user accurate retrieval portrait and the user fuzzy retrieval portrait to generate similarity retrieval inheritance data; The fusion retrieval module is used to obtain external information data; perform dynamic knowledge retrieval on user input data according to similarity retrieval inheritance data to generate dynamic knowledge retrieval data; perform multi-knowledge source fusion on dynamic knowledge retrieval data and cross-domain similarity retrieval data to generate fusion retrieval data; perform retrieval display adjustment on fusion retrieval data through a preset sentiment analysis model to generate the final hybrid retrieval display data; The retrieval feedback module is used to perform hybrid retrieval feedback optimization on the final hybrid retrieval display data to perform information technology retrieval consulting optimization operations.

Images