Automatic optimization method and system based on intelligent data memory
Through the construction of multimodal data acquisition and dynamic relationship map, combined with time decay factor, the shortcomings of traditional methods in multimodal information processing are solved, intelligent understanding of interactive data and accurate answer generation, and the intelligentization of the system is improved and efficiency.
Patent Information
- Application Number
- CN202410800997.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-20
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-06-20
AI Technical Summary
Traditional automatic optimization methods based on intelligent data memory databases are difficult to effectively process complex and changeable interactive data, especially when multimodal information (text, voice, image) is not effective and difficult to flexibly update and expand, resulting in a lack of targeting answers. Sex and accuracy.
Multimodal interactive data is collected through sensor arrays, multimodal semantic correlation analysis and semantic refinement are carried out, dynamic relationship maps and multi-level memory banks are constructed, time decay factors are introduced for short-term memory tracking and long-term memory association, and scenario perception and answer generation are realized .
It improves the ability to understand interactive data, realizes flexible organization and update of knowledge, enhances the accuracy and intelligence of answers, and helps intelligent assistants manage knowledge and provide information services more systematically.
Smart Images

Figure CN118503390B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an automatic optimization method and system based on an intelligent data memory library. Background Art
[0002] In recent years, with the rapid development of machine learning and artificial intelligence, automatic optimization methods based on intelligent data memory have gradually emerged and have shown great potential in many fields. The core idea of this method is to use historical data and machine learning algorithms to build an intelligent data memory, and automatically optimize system performance by analyzing the information in the memory. Traditional methods usually rely on rules or statistical models, which are difficult to handle complex and changeable interactive data, especially when facing the fusion of multimodal information (text, voice, image). The effect is not good. Traditional knowledge bases usually adopt a fixed structure and are difficult to update and expand flexibly. When generating answers, traditional methods often ignore the current interactive scenario, resulting in a lack of pertinence and accuracy in the answers. Summary of the invention
[0003] Based on this, it is necessary to provide an automatic optimization method and system based on an intelligent data memory to solve at least one of the above technical problems.
[0004] To achieve the above object, an automatic optimization method based on an intelligent data memory library comprises the following steps:
[0005] Step S1: collecting original interaction data through a sensor array to obtain an original interaction data set; extracting initial interaction events according to the original interaction data set to obtain initial interaction event data; and encapsulating the initial interaction event data into event data to obtain interaction event data;
[0006] Step S2: performing multimodal semantic association analysis on the interaction event data to obtain multimodal semantic association data; performing semantic extraction on the multimodal semantic association data to obtain knowledge metadata;
[0007] Step S3: performing relationship strength analysis on the knowledge metadata to obtain relationship strength data; performing relationship extraction processing based on the relationship strength data to obtain a dynamic relationship graph;
[0008] Step S4: sorting the importance of nodes in the dynamic relationship graph to obtain importance sorting data; performing multi-dimensional hierarchical construction processing based on the importance sorting data to obtain multi-level knowledge organization structure data; performing hierarchical storage based on the multi-level knowledge organization structure data to obtain a multi-level memory library;
[0009] Step S5: Perform short-term memory tracking and time decay factor calculation through the sensor array to obtain short-term memory data and time decay factor; perform long-term memory association on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory association data; perform situation perception according to the long-term memory association data to obtain structured situation data;
[0010] Step S6: Decode the scenario information of the structured scenario data to obtain scenario information decoding data; perform memory network retrieval on the scenario information decoding data according to the multi-level memory bank and the time attenuation factor to obtain an initial candidate answer list; perform problem scenario analysis based on the scenario information decoding data to obtain problem scenario parameters; perform question answer matching judgment on the problem scenario parameters according to the initial candidate answer list to obtain a candidate answer list; generate answers based on the candidate answer list to obtain accurate answer data.
[0011] The present invention can comprehensively record user interaction information through multimodal data collection and event extraction, and structure it into event data, providing a basis for subsequent semantic understanding and knowledge construction. Through multimodal semantic association analysis and semantic refinement, user intentions and interactive content can be deeply understood, key knowledge points can be extracted, and materials can be provided for building knowledge graphs. Through relationship strength analysis and relationship extraction, a dynamically updated knowledge graph can be constructed to reflect the semantic association and evolution process between knowledge points, providing support for knowledge reasoning and question-answering. Through node importance sorting and multi-dimensional hierarchical construction, the knowledge graph can be organized into a structured, hierarchical multi-level memory library to improve knowledge storage and retrieval efficiency. By introducing a time decay factor for short-term memory tracking and long-term memory association, the relevance weight of the information is adaptively adjusted according to the time distance of the interaction, and recent data is selectively focused on in the semantic association and reasoning process, so as to obtain long-term memory associations and semantic relationships that are closer to the current scenario. Through the steps of situation information decoding, memory network retrieval, question scenario analysis, and answer matching, the answer that best meets user needs can be retrieved and screened from the multi-level memory library, and accurate and easy-to-understand answer data can be finally generated. Therefore, the present invention provides an automatic optimization method and system based on an intelligent data memory library, which improves the ability to understand interactive data by effectively integrating and processing multimodal information such as text, voice, and images. Dynamic relationship graphs and multi-level memory libraries are used to achieve flexible organization and updating of knowledge. Through short-term memory tracking and long-term memory association, the perception of interactive scenarios is achieved and the accuracy of answers is improved. This method aims to achieve intelligent understanding of interactive data, knowledge organization, and accurate answer generation by simulating human memory and learning processes.
[0012] Applying this automatic optimization method based on intelligent data memory to intelligent assistants can help them manage their own knowledge more systematically, use experience more efficiently, understand user context more intelligently, and provide information services more comprehensively, ultimately improving the intelligence level of the system and the efficiency of problem solving.
[0013] Preferably, step S1 comprises the following steps:
[0014] Step S11: collecting original interaction data through a sensor array to obtain an original interaction data set; performing multimodal synchronization on the original interaction data set to obtain multimodal synchronized data;
[0015] Step S12: extracting the perception signal from the multimodal synchronization data to obtain multimodal perception signal data; and enhancing the signal features of the multimodal perception signal data to obtain multimodal signal feature enhanced data;
[0016] Step S13: performing multimodal semantic fusion on the multimodal signal feature enhancement data to obtain multimodal semantic representation data; performing interaction intent analysis on the multimodal semantic representation data to obtain interaction intent data; performing interaction behavior annotation on the interaction intent data to obtain interaction behavior annotation data;
[0017] Step S14: constructing structured events for the multimodal semantic representation data, the interaction intention data, and the interaction behavior annotation data to obtain initial interaction event data;
[0018] Step S15: adding event metadata to the initial interaction event data to obtain interaction event metadata; performing interaction event encapsulation on the initial interaction event data and the interaction event metadata to obtain interaction event data.
[0019] The present invention collects multimodal raw interaction data through a sensor array and synchronizes it in time, which can fully record the interaction process between the user and the system, ensure that different modal data correspond to the same interaction moment, and lay the foundation for subsequent multimodal information processing and semantic understanding. Extracting perceptual signals from multimodal synchronized data and performing feature enhancement can reduce noise interference, improve signal quality, highlight key information, and provide better feature input for subsequent multimodal semantic fusion, which helps to improve the accuracy of semantic understanding. Semantic fusion of multimodal signal features can integrate the advantages of different modal information to obtain a more comprehensive and accurate semantic representation, and then accurately analyze the user's interaction intention, and annotate the interaction behavior, providing richer semantic information for subsequent event construction. Integrating multimodal semantic representation, interaction intention and interaction behavior to construct structured initial interaction event data can express the user's interaction process in a clearer and more standardized way, and provide a structured data basis for subsequent semantic association analysis and knowledge extraction. Adding event metadata to the initial interaction event data and encapsulating it can form a more complete event description that contains richer contextual information, which helps to more fully understand the semantics of the event and provide a more sufficient information basis for subsequent knowledge storage, retrieval, and reasoning.
[0020] Preferably, step S2 comprises the following steps:
[0021] Step S21: performing event semantic deconstruction on the interactive event data to obtain event semantic unit data; performing text semantic analysis on the event semantic unit data to obtain text semantic data; performing voice and intonation analysis on the event semantic unit data to obtain voice and intonation data; performing image scene understanding on the event semantic unit data to obtain image scene data;
[0022] Step S22: performing multimodal semantic association on the text semantic data, the voice intonation data and the image scene data to obtain multimodal semantic association data;
[0023] Step S23: performing semantic role labeling on the multimodal semantic association data to obtain semantic role labeling data; performing semantic relationship extraction based on the semantic role labeling data to obtain a semantic relationship graph; performing key knowledge condensation on the semantic relationship graph to obtain key knowledge data;
[0024] Step S24: encapsulate the key knowledge data, semantic relationship graph and multimodal semantic association data into knowledge metadata to obtain knowledge metadata.
[0025] The present invention can extract rich semantic features from different modal information by semantically deconstructing interactive event data, and performing text semantic analysis, voice intonation analysis and image scene understanding respectively, so as to provide more fine-grained semantic units and more comprehensive feature representation for subsequent multimodal semantic association. By associating text semantics, voice intonation and image scene information, it is possible to identify semantic connections and complementary information between different modal information, form a more comprehensive and in-depth semantic understanding, and help to more accurately understand the relationship between user intentions and events. By performing semantic role labeling and semantic relationship extraction on multimodal semantic association data, it is possible to construct the semantic structure and logical relationship of events, form a semantic relationship map, and help to more clearly represent the causal relationship, time sequence relationship, etc. between events, and provide structured knowledge representation for subsequent knowledge condensation and reasoning. By encapsulating key knowledge, semantic relationship maps and multimodal semantic association data to form structured knowledge metadata, it is possible to integrate semantic information from different sources and at different levels, form a more comprehensive and systematic knowledge representation, and provide a unified data format for subsequent knowledge storage, retrieval and reasoning.
[0026] Preferably, step S3 comprises the following steps:
[0027] Step S31: performing semantic space mapping on the knowledge metadata to obtain high-dimensional semantic space vector representation data; performing semantic distance measurement on the high-dimensional semantic space vector representation data to obtain semantic distance measurement data;
[0028] Step S32: mining potential relationships of knowledge metadata based on semantic distance measurement data to obtain knowledge metadata potential relationship data; inferring relationship types based on knowledge metadata potential relationship data to obtain relationship type data;
[0029] Step S33: evaluating the relationship strength of the knowledge element potential relationship data to obtain relationship strength data; calculating the relationship confidence according to the relationship type data and the relationship strength data to obtain relationship confidence data;
[0030] Step S34: Update the dynamic graph according to the knowledge element potential relationship data and the relationship confidence data to obtain a dynamic relationship graph.
[0031] The present invention uses semantic space mapping to represent knowledge metadata as a high-dimensional vector, which can quantify semantic information and measure the semantic similarity between knowledge elements through semantic distance measurement, providing a quantitative basis for potential relationship mining. According to the semantic distance measurement results, the potential semantic relationship between knowledge elements is mined, and the specific relationship type is inferred based on the semantic information of the knowledge element itself, which can expand the knowledge graph (as the amount of data increases, the scale of the graph will also expand, which is consistent with the growth of the model scale in scalingLaw), discover new knowledge associations, and improve the richness of knowledge representation. The strength of the mined potential relationship is evaluated, and the relationship confidence is calculated in combination with the relationship type, which can effectively distinguish the importance and credibility of the relationship, filter out the potential relationship with low confidence, and improve the accuracy and reliability of the knowledge graph. According to the high-confidence potential relationship and relationship type, the dynamic relationship graph is updated, and the structure and content of the knowledge graph can be continuously improved, so that it can better reflect the semantic association between knowledge and support deeper knowledge reasoning and application.
[0032] Preferably, step S4 comprises the following steps:
[0033] Step S41: sorting the knowledge element nodes in the dynamic relationship graph by node importance to obtain importance sorting data;
[0034] Step S42: performing semantic feature clustering according to the dynamic relationship graph to obtain node semantic feature clustering data; performing relationship type division according to the dynamic relationship graph and importance ranking data to obtain a node relationship subgraph set;
[0035] Step S43: constructing a multi-dimensional hierarchical structure of the node relationship subgraph set according to the node semantic feature clustering data to obtain multi-level knowledge organization structure data;
[0036] Step S44: encapsulating knowledge units according to the multi-level knowledge organization structure data to obtain knowledge unit data; establishing multi-level indexes for the multi-level knowledge organization structure data to obtain a multi-level index table;
[0037] Step S45: performing memory fragment storage on the knowledge unit data to obtain memory fragment storage data; performing memory fragment association on the memory fragment storage data to obtain memory network data;
[0038] Step S46: Integrate the multi-level memory base according to the memory network data and the multi-level index table to obtain a multi-level memory base.
[0039] The present invention can distinguish the importance of different knowledge elements by ranking the importance of knowledge element nodes in the dynamic relationship map, helping the system to give priority to knowledge elements with high importance in the subsequent knowledge retrieval and reasoning process, thereby improving the efficiency and accuracy of information processing. Clustering is performed according to semantic features, knowledge element nodes with similar semantics are divided together, and node relationship sub-graphs are divided according to relationship types, so that complex knowledge graphs can be decomposed into multiple semantically and structurally relatively independent sub-parts, which is convenient for hierarchical management and efficient retrieval. According to the semantic feature clustering results and the relationship type division results, the node relationship sub-graphs are organized into a multi-level knowledge organization structure, which can build a knowledge system with clear hierarchy and structure, and improve the comprehensibility and interpretability of knowledge. The knowledge units in the multi-level knowledge organization structure are encapsulated, and a multi-level index table is established, so that the knowledge information can be stored in a more structured manner, and it is supported to quickly retrieve the target knowledge unit according to multiple dimensions such as semantic categories and relationship types, thereby improving the efficiency of knowledge access. Storing knowledge unit data in the form of memory fragments and establishing associations between memory fragments can simulate the fragmentation and association characteristics of human brain memory, build a knowledge storage method that is closer to the human brain's cognitive model, and support more flexible knowledge association and reasoning. Integrating memory network data and multi-level index tables to build a multi-level memory library can combine the advantages of the two storage methods to achieve effective organization, storage and access to massive knowledge information, support rapid retrieval, semantic association, knowledge reasoning and other application needs, and provide a basis for building intelligent question-answering systems and knowledge service systems.
[0040] Preferably, step S5 comprises the following steps:
[0041] Step S51: Tracking the interaction flow through the sensor array to obtain interaction flow data; timestamping the interaction flow data and performing short-term memory caching to obtain timestamp data and short-term memory cache data;
[0042] Step S52: performing subject focus identification on the short-term memory cache data to obtain short-term memory data; performing time decay factor calculation according to the timestamp data to obtain the time decay factor;
[0043] Step S53: performing long-term memory retrieval on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory retrieval data;
[0044] Step S54: performing long-term memory association on the long-term memory retrieval data and the short-term memory data to obtain a candidate association data set; performing time weighting on the candidate association data set according to the time decay factor to obtain time weight data; performing semantic relationship reasoning on the candidate association data set according to the time weight data to obtain semantic relationship reasoning data; performing long-term memory screening on the candidate association data set according to the semantic relationship reasoning data to obtain long-term memory association data;
[0045] Step S55: Perform dynamic scenario update modeling based on the long-term memory association data and the semantic relationship reasoning data to obtain a dynamic scenario model; perform scenario model abstraction processing on the dynamic scenario model to obtain structured scenario data.
[0046] The present invention can more accurately record the time when an event occurs by introducing a timestamp mark when tracking the interaction flow, and provide support for subsequent operations such as time decay calculation and time weighting. By performing short-term memory caching on the interaction flow data, the user's interaction information can be temporarily stored within a certain time range to avoid the loss of important information. By adding a time decay factor during long-term memory retrieval, the relevance of the memory data can be reasonably adjusted according to the time of the interaction, so that the memory data closer to the current interaction time obtains a higher weight and better reflects the user's immediate intention. Time-weighted processing is performed on the candidate association data set, so that more time-closed data can be considered during semantic relationship reasoning, so that the reasoning result is more in line with the current scenario. By assigning time weights, the influence of data at different time points on the current semantic relationship reasoning can be quantified. By using semantic relationship reasoning data during long-term memory screening, long-term memories closely related to the current scenario can be selected more accurately to form high-quality long-term memory association data.
[0047] Preferably, step S6 comprises the following steps:
[0048] Step S61: decoding the structured scenario data to obtain scenario information decoding data;
[0049] Step S62: performing multi-level index positioning on the scene information decoding data according to the multi-level memory library to obtain memory fragment positioning data; performing semantic similarity sorting on the memory fragment positioning data according to the time decay factor to obtain answer semantic similarity sorting data;
[0050] Step S63: performing relationship path matching on the multi-level memory library according to the scenario information decoding data to obtain relationship path matching data;
[0051] Step S64: using the memory fragment link relationship data in the multi-level memory library to perform memory network traversal according to the answer semantic similarity sorting data and the relationship path matching data and perform time threshold control according to the time decay factor to obtain an initial candidate answer list;
[0052] Step S65: performing problem scenario analysis according to the scenario information decoding data to obtain problem scenario parameters;
[0053] Step S66: performing question answer matching calculation on the question scenario parameters according to the initial candidate answer list to obtain a question answer matching value; when the question answer matching value is greater than or equal to a preset matching threshold, the initial candidate answer list is used as a candidate answer list and step S69 is executed; when the question answer matching value is less than the preset matching threshold, step S67 is executed;
[0054] Step S67: Expand external knowledge according to the question scenario parameters to obtain external knowledge data; filter external information on the external knowledge data to obtain a list of external candidate answers;
[0055] Step S68: perform answer fusion sorting on the initial candidate answer list and the external candidate answer list to obtain a candidate answer list;
[0056] Step S69: Generate answers based on the candidate answer list to obtain accurate answer data.
[0057] The present invention can convert abstract scenario information into specific user intentions, question backgrounds, etc. by decoding structured scenario data, providing a clearer target for subsequent answer retrieval and matching. By using multi-level indexes to quickly locate memory fragments containing answers and sorting them according to semantic similarity, the answer search range can be narrowed, the retrieval efficiency can be improved, and answers related to the current scenario semantics can be given priority. Through relationship path matching, knowledge paths related to the current scenario can be found, such as the causal chain of events, the time development sequence, etc., to provide richer context information for answer retrieval and improve the accuracy and relevance of answers. Combining the results of semantic similarity sorting and relationship path matching, traversing the association relationship between memory fragments can more comprehensively consider various relevant information and find the candidate answer that best matches the current scenario. In-depth analysis of the problem scenario and extraction of key parameters such as time, place, user intention, emotion, etc. can more accurately characterize the characteristics of the problem and provide more refined evaluation indicators for subsequent answer matching. According to the problem scenario parameters, the matching degree of the candidate answers is calculated and the matching threshold is set, which can effectively filter out answers that do not meet the requirements, improve the reliability of the answers, and provide a judgment basis for whether external knowledge expansion is needed. When the appropriate answer cannot be found in the memory bank, external knowledge expansion can be used to obtain more extensive knowledge information, make up for the shortcomings of the memory bank, and improve the answer coverage of the system. By integrating and sorting the answers retrieved from the internal memory bank and the answers obtained from external knowledge expansion, the reliability and relevance of information from different sources can be comprehensively considered to obtain a more comprehensive and accurate list of candidate answers. Based on the final list of candidate answers, answers can be generated. The appropriate answer presentation method can be selected according to the user's needs and preferences, and the answers can be integrated and optimized to ultimately generate more accurate and easier-to-understand answer data.
[0058] Preferably, step S65 includes the following steps:
[0059] Step S651: Perform time information analysis based on the scenario information decoded data to obtain problem time information data; perform subject area analysis based on the scenario information decoded data to obtain problem subject area data;
[0060] Step S652: Perform user intention analysis based on the context information decoded data to obtain problem user intention data; perform emotion analysis based on the context information decoded data to obtain problem emotion data;
[0061] Step S653: Extract question scenario parameters based on question time information data, question subject domain data, question user intention data, and question emotion data; obtain question scenario parameters.
[0062] The present invention can extract the time range and subject category of the question by analyzing the time information and subject area of the scenario information decoding data, so as to help the system locate the information related to the question more accurately when retrieving and matching the answer, eliminate the interference of irrelevant information, and improve the accuracy and efficiency of the answer. By analyzing the user's intention and emotion, the user's question can be understood more accurately, such as distinguishing whether the user is asking for information, seeking help or expressing emotion, as well as the user's emotional state, so as to select a more appropriate answer generation strategy, such as providing a more detailed explanation, a more friendly suggestion or a more empathetic response, and improve user satisfaction. By integrating information such as time information, subject area, user intention and emotional emotion into the problem scenario parameters, the characteristics of the problem can be more comprehensively characterized, and more refined evaluation indicators can be provided for subsequent answer matching, helping the system to select answers that are more matched with the question in terms of time, subject, intention, emotion, etc., and improving the accuracy of the answer and user experience.
[0063] Preferably, step S69 includes the following steps:
[0064] Step S691: parsing the candidate answer list to obtain candidate answer parsing data;
[0065] Step S692: Refining the answer content of the candidate answer analysis data to obtain refined answer data;
[0066] Step S693: performing answer multimodal enhancement on the refined answer data to obtain answer multimodal enhanced data;
[0067] Step S694: The refined answer data is refined according to the answer multimodal enhancement data to obtain accurate answer data.
[0068] The present invention can convert different forms of answer information into structured data by parsing candidate answers and extracting key information, such as text content, source, related pictures, etc., so as to facilitate the subsequent processing and analysis of the system, such as answer refinement, multimodal enhancement, etc. Refining the content of candidate answers, removing redundant information, and refining the core content can make the answers more concise and easy to understand, improve the user's reading experience, and facilitate the system to perform subsequent multimodal enhancement and answer integration. Performing multimodal enhancement on the refined answers, such as converting text answers into voice broadcasts, or generating related pictures and videos based on the answer content, can make the answers more vivid and vivid, improve user participation and interest, and enhance user experience. According to the answers after multimodal enhancement, the refined answers are finally integrated and optimized, such as selecting a suitable answer presentation method according to the user's preferences, or adjusting the expression style of the answer according to different interactive scenarios, etc., so that the final answer can be more accurate, more in line with the user's needs and expectations, and improve user satisfaction.
[0069] Preferably, the present invention further provides an automatic optimization system based on an intelligent data memory, which is used to execute the automatic optimization method based on an intelligent data memory as described above, and the automatic optimization system based on an intelligent data memory comprises:
[0070] The interaction capture module is used to collect original interaction data through the sensor array to obtain an original interaction data set; extract initial interaction events based on the original interaction data set to obtain initial interaction event data; and encapsulate the initial interaction event data into event data to obtain interaction event data;
[0071] The semantic extraction module is used to perform multimodal semantic association analysis on the interactive event data to obtain multimodal semantic association data; perform semantic extraction on the multimodal semantic association data to obtain knowledge metadata;
[0072] The relationship weaving module is used to analyze the relationship strength of knowledge metadata to obtain relationship strength data; perform relationship extraction processing based on the relationship strength data to obtain a dynamic relationship graph;
[0073] The memory precipitation module is used to sort the importance of nodes in the dynamic relationship graph to obtain importance sorting data; perform multi-dimensional hierarchical construction processing based on the importance sorting data to obtain multi-level knowledge organization structure data; perform hierarchical storage based on the multi-level knowledge organization structure data to obtain a multi-level memory library;
[0074] The situational awareness module is used to perform short-term memory tracking and time decay factor calculation through the sensor array to obtain short-term memory data and time decay factor; perform long-term memory association on the multi-level memory bank according to the short-term memory data and time decay factor to obtain long-term memory association data; perform situational awareness according to the long-term memory association data to obtain structured situational data;
[0075] The problem scenario retrieval module is used to decode the scenario information of the structured scenario data to obtain the scenario information decoding data; perform memory network retrieval on the scenario information decoding data according to the multi-level memory bank and the time attenuation factor to obtain an initial candidate answer list; perform problem scenario analysis based on the scenario information decoding data to obtain problem scenario parameters; perform question answer matching judgment on the problem scenario parameters according to the initial candidate answer list to obtain a candidate answer list; generate answers according to the candidate answer list to obtain accurate answer data.
[0076] The present invention can comprehensively record user interaction information through multimodal data collection and event extraction, and structure it into event data, providing a basis for subsequent semantic understanding and knowledge construction. Through multimodal semantic association analysis and semantic refinement, user intentions and interactive content can be deeply understood, key knowledge points can be extracted, and materials can be provided for building knowledge graphs. Through relationship strength analysis and relationship extraction, a dynamically updated knowledge graph can be constructed to reflect the semantic association and evolution process between knowledge points, providing support for knowledge reasoning and question-answering. Through node importance sorting and multi-dimensional hierarchical construction, the knowledge graph can be organized into a structured, hierarchical multi-level memory library to improve knowledge storage and retrieval efficiency. By introducing a time decay factor for short-term memory tracking and long-term memory association, the relevance weight of the information is adaptively adjusted according to the time distance of the interaction, and recent data is selectively focused on in the semantic association and reasoning process, so as to obtain long-term memory associations and semantic relationships that are closer to the current scenario. Through the steps of situation information decoding, memory network retrieval, question scenario analysis, and answer matching, the answer that best meets user needs can be retrieved and screened from the multi-level memory library, and accurate and easy-to-understand answer data can be finally generated. Therefore, the present invention provides an automatic optimization method and system based on an intelligent data memory library, which improves the ability to understand interactive data by effectively integrating and processing multimodal information such as text, voice, and images. Dynamic relationship graphs and multi-level memory libraries are used to achieve flexible organization and updating of knowledge. Through short-term memory tracking and long-term memory association, the perception of interactive scenarios is achieved and the accuracy of answers is improved. This method aims to achieve intelligent understanding of interactive data, knowledge organization, and accurate answer generation by simulating human memory and learning processes.
[0077] Applying this automatic optimization method based on intelligent data memory to intelligent assistants can help them manage their own knowledge more systematically, use experience more efficiently, understand user context more intelligently, and provide information services more comprehensively, ultimately improving the intelligence level of the system and the efficiency of problem solving. BRIEF DESCRIPTION OF THE DRAWINGS
[0078] Figure 1 A schematic diagram of the steps of an automatic optimization method based on an intelligent data memory;
[0079] Figure 2 for Figure 1 Detailed implementation steps of step S5;
[0080] Figure 3 for Figure 1 Detailed implementation steps of step S6 in FIG.
[0081] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION
[0082] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.
[0083] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying 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 implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.
[0084] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. 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 herein includes any and all combinations of one or more of the listed associated items.
[0085] To achieve this, please refer to Figures 1 to 3, an automatic optimization method based on intelligent data memory, comprising the following steps:
[0086] Step S1: collecting original interaction data through a sensor array to obtain an original interaction data set; extracting initial interaction events according to the original interaction data set to obtain initial interaction event data; and encapsulating the initial interaction event data into event data to obtain interaction event data;
[0087] Step S2: performing multimodal semantic association analysis on the interaction event data to obtain multimodal semantic association data; performing semantic extraction on the multimodal semantic association data to obtain knowledge metadata;
[0088] Step S3: performing relationship strength analysis on the knowledge metadata to obtain relationship strength data; performing relationship extraction processing based on the relationship strength data to obtain a dynamic relationship graph;
[0089] Step S4: sorting the importance of nodes in the dynamic relationship graph to obtain importance sorting data; performing multi-dimensional hierarchical construction processing based on the importance sorting data to obtain multi-level knowledge organization structure data; performing hierarchical storage based on the multi-level knowledge organization structure data to obtain a multi-level memory library;
[0090] Step S5: Perform short-term memory tracking and time decay factor calculation through the sensor array to obtain short-term memory data and time decay factor; perform long-term memory association on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory association data; perform situation perception according to the long-term memory association data to obtain structured situation data;
[0091] Step S6: Decode the scenario information of the structured scenario data to obtain scenario information decoding data; perform memory network retrieval on the scenario information decoding data according to the multi-level memory bank and the time attenuation factor to obtain an initial candidate answer list; perform problem scenario analysis based on the scenario information decoding data to obtain problem scenario parameters; perform question answer matching judgment on the problem scenario parameters according to the initial candidate answer list to obtain a candidate answer list; generate answers based on the candidate answer list to obtain accurate answer data.
[0092] The present invention can comprehensively record user interaction information through multimodal data collection and event extraction, and structure it into event data, providing a basis for subsequent semantic understanding and knowledge construction. Through multimodal semantic association analysis and semantic refinement, user intentions and interactive content can be deeply understood, key knowledge points can be extracted, and materials can be provided for building knowledge graphs. Through relationship strength analysis and relationship extraction, a dynamically updated knowledge graph can be constructed to reflect the semantic association and evolution process between knowledge points, providing support for knowledge reasoning and question-answering. Through node importance sorting and multi-dimensional hierarchical construction, the knowledge graph can be organized into a structured, hierarchical multi-level memory library to improve knowledge storage and retrieval efficiency. By introducing a time decay factor for short-term memory tracking and long-term memory association, the relevance weight of the information is adaptively adjusted according to the time distance of the interaction, and recent data is selectively focused on in the semantic association and reasoning process, so as to obtain long-term memory associations and semantic relationships that are closer to the current scenario. Through the steps of situation information decoding, memory network retrieval, question scenario analysis, and answer matching, the answer that best meets user needs can be retrieved and screened from the multi-level memory library, and accurate and easy-to-understand answer data can be finally generated. Therefore, the present invention provides an automatic optimization method and system based on an intelligent data memory library, which improves the ability to understand interactive data by effectively integrating and processing multimodal information such as text, voice, and images. Dynamic relationship graphs and multi-level memory libraries are used to achieve flexible organization and updating of knowledge. Through short-term memory tracking and long-term memory association, the perception of interactive scenarios is achieved and the accuracy of answers is improved. This method aims to achieve intelligent understanding of interactive data, knowledge organization, and accurate answer generation by simulating human memory and learning processes.
[0093] Applying this automatic optimization method based on intelligent data memory to intelligent assistants can help them manage their own knowledge more systematically, use experience more efficiently, understand user context more intelligently, and provide information services more comprehensively, ultimately improving the intelligence level of the system and the efficiency of problem solving.
[0094] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of the automatic optimization method based on the intelligent data memory of the present invention. In this example, the automatic optimization method based on the intelligent data memory includes the following steps:
[0095] Step S1: collecting original interaction data through a sensor array to obtain an original interaction data set; extracting initial interaction events according to the original interaction data set to obtain initial interaction event data; and encapsulating the initial interaction event data into event data to obtain interaction event data;
[0096] In the embodiment of the present invention, a sensor array (such as a microphone, a camera, etc.) is used to collect raw data of user interaction, including voice, text, images, etc., to form a raw interaction data set. Then, the system identifies key interaction events from the raw interaction data according to a preset event recognition rule or model, such as a user asking a question, a user clicking a button, etc., to form initial interaction event data. Finally, the system performs structured processing on the initial interaction event data, adding information such as event type, timestamp, participants, etc., to form complete interaction event data.
[0097] Step S2: performing multimodal semantic association analysis on the interaction event data to obtain multimodal semantic association data; performing semantic extraction on the multimodal semantic association data to obtain knowledge metadata;
[0098] In the embodiment of the present invention, the multimodal information in the interactive event data is analyzed by imitating the information processing mechanism of the human brain (the human brain itself perceives and processes information in a multimodal manner), which is also similar to the association mechanism of the human brain, such as associating voice intonation with text emotion, or associating user expressions with interactive behaviors, to form semantic association data containing richer semantic information. Then, the system abstracts and refines the semantic association data, extracts key knowledge points, such as the causal relationship, time sequence relationship, semantic similarity relationship, etc. between events, and forms structured knowledge metadata.
[0099] Step S3: performing relationship strength analysis on the knowledge metadata to obtain relationship strength data; performing relationship extraction processing based on the relationship strength data to obtain a dynamic relationship graph;
[0100] In an embodiment of the present invention, knowledge metadata is analyzed, and indicators such as semantic similarity, co-occurrence frequency, and dependency between knowledge elements are calculated to evaluate the strength of the relationship between knowledge elements and form relationship strength data. Then, based on the relationship strength data, a dynamic relationship graph is constructed to represent the semantic association and evolution process between knowledge elements.
[0101] Step S4: sorting the importance of nodes in the dynamic relationship graph to obtain importance sorting data; performing multi-dimensional hierarchical construction processing based on the importance sorting data to obtain multi-level knowledge organization structure data; performing hierarchical storage based on the multi-level knowledge organization structure data to obtain a multi-level memory library;
[0102] In the embodiment of the present invention, the importance of knowledge element nodes is sorted according to the topological structure and node attributes of the dynamic relationship graph to distinguish the importance of different knowledge elements. Then, the system organizes the dynamic relationship graph into a multi-level knowledge organization structure according to the node importance ranking and semantic association relationship, and stores it in a multi-level memory library for efficient retrieval and access.
[0103] Step S5: Perform short-term memory tracking and time decay factor calculation through the sensor array to obtain short-term memory data and time decay factor; perform long-term memory association on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory association data; perform situation perception according to the long-term memory association data to obtain structured situation data;
[0104] In the embodiment of the present invention, the user interaction is continuously tracked by the sensor array to form the interaction flow data with time stamp, which is stored in the short-term memory cache. Then, the cached data is subjected to the subject focus identification and the time decay factor calculation to obtain the short-term memory data and the decay factor. Then, the multi-level long-term memory library is searched according to the short-term memory data and the decay factor to obtain the relevant long-term memory retrieval data. Then, the retrieval data and the short-term memory data are subjected to the association analysis to form the candidate association data set, and the time decay factor is combined to perform the time weighting, and the weighted candidate data set is subjected to the semantic relationship reasoning and screening to obtain the high-quality long-term memory association data. Finally, the association data and the reasoning data are integrated into the dynamic scenario model, and the model is abstracted to form the structured scenario data for the subsequent intelligent application.
[0105] Step S6: Decode the structured scenario data to obtain scenario information decoding data; perform memory network retrieval on the scenario information decoding data according to the multi-level memory bank and the time decay factor to obtain an initial candidate answer list; perform question scenario analysis on the scenario information decoding data to obtain question scenario parameters; perform question answer matching judgment on the question scenario parameters according to the initial candidate answer list to obtain a candidate answer list; generate answers according to the candidate answer list to obtain accurate answer data;
[0106] In the embodiment of the present invention, the structured scenario data is parsed to extract key information, such as user intention, question background, emotional tendency, etc., to form scenario information decoding data. Then, the system searches the multi-level memory library based on the scenario information decoding data to find candidate answers related to the current scenario, and sorts and screens the candidate answers based on the question scenario parameters, and finally generates accurate answer data and returns it to the user.
[0107] Preferably, step S1 comprises the following steps:
[0108] Step S11: collecting original interaction data through a sensor array to obtain an original interaction data set; performing multimodal synchronization on the original interaction data set to obtain multimodal synchronized data;
[0109] Step S12: extracting the perception signal from the multimodal synchronization data to obtain multimodal perception signal data; and enhancing the signal features of the multimodal perception signal data to obtain multimodal signal feature enhanced data;
[0110] Step S13: performing multimodal semantic fusion on the multimodal signal feature enhancement data to obtain multimodal semantic representation data; performing interaction intent analysis on the multimodal semantic representation data to obtain interaction intent data; performing interaction behavior annotation on the interaction intent data to obtain interaction behavior annotation data;
[0111] Step S14: constructing structured events for the multimodal semantic representation data, the interaction intention data, and the interaction behavior annotation data to obtain initial interaction event data;
[0112] Step S15: adding event metadata to the initial interaction event data to obtain interaction event metadata; performing interaction event encapsulation on the initial interaction event data and the interaction event metadata to obtain interaction event data.
[0113] The present invention collects multimodal raw interaction data through a sensor array and synchronizes it in time, which can fully record the interaction process between the user and the system, ensure that different modal data correspond to the same interaction moment, and lay the foundation for subsequent multimodal information processing and semantic understanding. Extracting perceptual signals from multimodal synchronized data and performing feature enhancement can reduce noise interference, improve signal quality, highlight key information, and provide better feature input for subsequent multimodal semantic fusion, which helps to improve the accuracy of semantic understanding. Semantic fusion of multimodal signal features can integrate the advantages of different modal information to obtain a more comprehensive and accurate semantic representation, and then accurately analyze the user's interaction intention, and annotate the interaction behavior, providing richer semantic information for subsequent event construction. Integrating multimodal semantic representation, interaction intention and interaction behavior to construct structured initial interaction event data can express the user's interaction process in a clearer and more standardized way, and provide a structured data basis for subsequent semantic association analysis and knowledge extraction. Adding event metadata to the initial interaction event data and encapsulating it can form a more complete event description that contains richer contextual information, which helps to more fully understand the semantics of the event and provide a more sufficient information basis for subsequent knowledge storage, retrieval, and reasoning.
[0114] In an embodiment of the present invention, the system uses a deployed sensor array (such as a microphone, a camera, a depth sensor, etc.) to record the interaction process between the user and the system in real time, and obtains the original interaction data set, which contains multiple modal data such as voice signals, image sequences, and depth information. Subsequently, the system aligns the multimodal data in time according to timestamps or other synchronization marks to obtain multimodal synchronization data, ensuring that different modal data correspond to the same interaction moment. The system analyzes the multimodal synchronization data and extracts the perception signals of each modality respectively. For example, voice features are extracted from voice signals, visual features (target detection frames, human body posture estimation, etc.) are extracted from image sequences, and spatial position information is extracted from depth information to obtain multimodal perception signal data. Then, the system improves the quality and strengthens the information of the multimodal perception signal data, such as voice noise reduction, image enhancement, feature standardization, etc., to obtain multimodal signal feature enhancement data. The system uses multimodal fusion technology (such as attention mechanism, graph neural network, etc.) to integrate the multimodal signal feature enhancement data to obtain unified multimodal semantic representation data containing richer information. Subsequently, the system parses the user's interaction intention based on the multimodal semantic representation data and combines it with a predefined interaction intention dictionary or intention recognition model to obtain interaction intention data (such as query information, execution instructions, etc.). Finally, the system annotates the interaction intention data according to the interaction intention and the preset interaction behavior label system to obtain interaction behavior annotation data (such as clicking a button, voice input, etc.). According to the multimodal semantic representation data, interaction intention data, and interaction behavior annotation data, the user's interaction process is structured according to the predefined event structure template (such as event subject, event predicate, event object, etc.), and the initial interaction event data is constructed for subsequent processing and analysis. The initial interaction event data is supplemented and improved, and metadata information such as the time, place, and participants of the event is added to obtain interaction event metadata. Finally, the system encapsulates the initial interaction event data and interaction event metadata to form complete interaction event data, providing basic data for subsequent semantic association analysis.
[0115] Preferably, step S2 comprises the following steps:
[0116] Step S21: performing event semantic deconstruction on the interactive event data to obtain event semantic unit data; performing text semantic analysis on the event semantic unit data to obtain text semantic data; performing voice and intonation analysis on the event semantic unit data to obtain voice and intonation data; performing image scene understanding on the event semantic unit data to obtain image scene data;
[0117] Step S22: performing multimodal semantic association on the text semantic data, the voice intonation data and the image scene data to obtain multimodal semantic association data;
[0118] Step S23: performing semantic role labeling on the multimodal semantic association data to obtain semantic role labeling data; performing semantic relationship extraction based on the semantic role labeling data to obtain a semantic relationship graph; performing key knowledge condensation on the semantic relationship graph to obtain key knowledge data;
[0119] Step S24: encapsulate the key knowledge data, semantic relationship graph and multimodal semantic association data into knowledge metadata to obtain knowledge metadata.
[0120] The present invention can extract rich semantic features from different modal information by semantically deconstructing interactive event data, and performing text semantic analysis, voice intonation analysis and image scene understanding respectively, so as to provide more fine-grained semantic units and more comprehensive feature representation for subsequent multimodal semantic association. By associating text semantics, voice intonation and image scene information, it is possible to identify semantic connections and complementary information between different modal information, form a more comprehensive and in-depth semantic understanding, and help to more accurately understand the relationship between user intentions and events. By performing semantic role labeling and semantic relationship extraction on multimodal semantic association data, it is possible to construct the semantic structure and logical relationship of events, form a semantic relationship map, and help to more clearly represent the causal relationship, time sequence relationship, etc. between events, and provide structured knowledge representation for subsequent knowledge condensation and reasoning. By encapsulating key knowledge, semantic relationship maps and multimodal semantic association data to form structured knowledge metadata, it is possible to integrate semantic information from different sources and at different levels, form a more comprehensive and systematic knowledge representation, and provide a unified data format for subsequent knowledge storage, retrieval and reasoning.
[0121] In an embodiment of the present invention, the system parses the interactive event data and decomposes each event into multiple event semantic units, such as the subject, object, action, time, place, etc. of the event. For each event semantic unit, the system performs text semantic parsing, voice intonation analysis, and image scene understanding. For example, natural language processing technology is used to extract keywords, entities, emotions, and other information of the text semantic unit to form text semantic data; speech recognition and sentiment analysis technology are used to identify emotions, attitudes, and other information in voice intonation to form voice intonation data; computer vision technology is used to identify objects, characters, relationships, and other information in image scenes to form image scene data. Multimodal fusion technology is used to perform association analysis on text semantic data, voice intonation data, and image scene data to identify semantic connections and complementary information between different modal information. For example, the system can identify the contrast between the sarcastic tone in voice intonation and the semantic content of the text, or associate the facial expressions of characters in the image scene with the emotions of voice intonation, and finally obtain multimodal semantic association data containing more comprehensive and in-depth semantic information. Based on multimodal semantic association data, each semantic unit in the event is role-labeled, such as identifying the initiator, recipient, time, and location of the event. Then, based on the semantic role labeling results, the system extracts the semantic relationship between each semantic unit in the event and constructs a semantic relationship graph to represent the semantic structure and logical relationship of the event. Finally, the system analyzes the semantic relationship graph to extract the core content and key knowledge points of the event to form key knowledge data. The key knowledge data, semantic relationship graph, and multimodal semantic association data are integrated to form structured knowledge metadata. Knowledge metadata encapsulates the semantic information, logical relationships, and multimodal features of the event, providing a basis for subsequent knowledge storage, retrieval, and reasoning.
[0122] Preferably, step S3 comprises the following steps:
[0123] Step S31: performing semantic space mapping on the knowledge metadata to obtain high-dimensional semantic space vector representation data; performing semantic distance measurement on the high-dimensional semantic space vector representation data to obtain semantic distance measurement data;
[0124] Step S32: mining potential relationships of knowledge metadata based on semantic distance measurement data to obtain knowledge metadata potential relationship data; inferring relationship types based on knowledge metadata potential relationship data to obtain relationship type data;
[0125] Step S33: evaluating the relationship strength of the knowledge element potential relationship data to obtain relationship strength data; calculating the relationship confidence according to the relationship type data and the relationship strength data to obtain relationship confidence data;
[0126] Step S34: Update the dynamic graph according to the knowledge element potential relationship data and the relationship confidence data to obtain a dynamic relationship graph.
[0127] The present invention uses semantic space mapping to represent knowledge metadata as a high-dimensional vector, which can quantify semantic information and measure the semantic similarity between knowledge elements through semantic distance measurement, providing a quantitative basis for potential relationship mining. According to the semantic distance measurement results, the potential semantic relationship between knowledge elements is mined, and the specific relationship type is inferred based on the semantic information of the knowledge element itself, which can expand the knowledge graph (as the amount of data increases, the scale of the graph will also expand, which is consistent with the growth of the model scale in scalingLaw), discover new knowledge associations, and improve the richness of knowledge representation. The strength of the mined potential relationship is evaluated, and the relationship confidence is calculated in combination with the relationship type, which can effectively distinguish the importance and credibility of the relationship, filter out the potential relationship with low confidence, and improve the accuracy and reliability of the knowledge graph. According to the high-confidence potential relationship and relationship type, the dynamic relationship graph is updated, and the structure and content of the knowledge graph can be continuously improved, so that it can better reflect the semantic association between knowledge and support deeper knowledge reasoning and application.
[0128] In an embodiment of the present invention, the system uses a pre-trained language model or knowledge graph embedding technology to map each knowledge element to a vector representation in a high-dimensional semantic space, and obtains high-dimensional semantic space vector representation data, so that semantically similar knowledge elements are closer in spatial distance. Then, the system calculates the distance (such as cosine distance, Euclidean distance, etc.) between any two knowledge element vector representations to obtain semantic distance measurement data, which is used to characterize the semantic similarity between knowledge elements. According to the semantic distance measurement data, a threshold is set or a clustering algorithm is used to identify knowledge element pairs with a close semantic distance, and they are regarded as knowledge elements with potential relationships, and knowledge element potential relationship data is obtained. Then, the system classifies the potential relationships of knowledge elements according to the semantic information of the knowledge elements and a predefined set of relationship types (such as causal relationships, inclusion relationships, time sequence relationships, etc.), infers specific relationship types, and obtains relationship type data. According to the knowledge element potential relationship data, combined with multimodal semantic association data, event co-occurrence frequency and other information, each group of potential relationships is evaluated for strength, such as calculating indicators such as relationship support and confidence, to obtain relationship strength data. Then, based on the relationship type data and relationship strength data, the system comprehensively considers the prior knowledge and confidence distribution of different relationship types, calculates the confidence of each group of potential relationships, and obtains the relationship confidence data. Based on the knowledge element potential relationship data and relationship confidence data, the existing dynamic relationship graph is updated. For example, potential relationships with confidence exceeding the threshold are added as new relationship edges, the weights of relationship edges are updated, or existing relationship edges with low confidence are deleted. Finally, an updated dynamic relationship graph is obtained to represent the semantic association and evolution process between knowledge elements.
[0129] Preferably, step S4 comprises the following steps:
[0130] Step S41: sorting the knowledge element nodes in the dynamic relationship graph by node importance to obtain importance sorting data;
[0131] Step S42: performing semantic feature clustering according to the dynamic relationship graph to obtain node semantic feature clustering data; performing relationship type division according to the dynamic relationship graph and importance ranking data to obtain a node relationship subgraph set;
[0132] Step S43: constructing a multi-dimensional hierarchical structure of the node relationship subgraph set according to the node semantic feature clustering data to obtain multi-level knowledge organization structure data;
[0133] Step S44: encapsulating knowledge units according to the multi-level knowledge organization structure data to obtain knowledge unit data; establishing multi-level indexes for the multi-level knowledge organization structure data to obtain a multi-level index table;
[0134] Step S45: performing memory fragment storage on the knowledge unit data to obtain memory fragment storage data; performing memory fragment association on the memory fragment storage data to obtain memory network data;
[0135] Step S46: Integrate the multi-level memory base according to the memory network data and the multi-level index table to obtain a multi-level memory base.
[0136] The present invention can distinguish the importance of different knowledge elements by ranking the importance of knowledge element nodes in the dynamic relationship map, helping the system to give priority to knowledge elements with high importance in the subsequent knowledge retrieval and reasoning process, thereby improving the efficiency and accuracy of information processing. Clustering is performed according to semantic features, knowledge element nodes with similar semantics are divided together, and node relationship sub-graphs are divided according to relationship types, so that complex knowledge graphs can be decomposed into multiple semantically and structurally relatively independent sub-parts, which is convenient for hierarchical management and efficient retrieval. According to the semantic feature clustering results and the relationship type division results, the node relationship sub-graphs are organized into a multi-level knowledge organization structure, which can build a knowledge system with clear hierarchy and structure, and improve the comprehensibility and interpretability of knowledge. The knowledge units in the multi-level knowledge organization structure are encapsulated, and a multi-level index table is established, so that the knowledge information can be stored in a more structured manner, and it is supported to quickly retrieve the target knowledge unit according to multiple dimensions such as semantic categories and relationship types, thereby improving the efficiency of knowledge access. Storing knowledge unit data in the form of memory fragments and establishing associations between memory fragments can simulate the fragmentation and association characteristics of human brain memory, build a knowledge storage method that is closer to the human brain's cognitive model, and support more flexible knowledge association and reasoning. Integrating memory network data and multi-level index tables to build a multi-level memory library can combine the advantages of the two storage methods to achieve effective organization, storage and access to massive knowledge information, support rapid retrieval, semantic association, knowledge reasoning and other application needs, and provide a basis for building intelligent question-answering systems and knowledge service systems.
[0137] In an embodiment of the present invention, the system uses a node importance ranking algorithm based on the topological structure and node attribute information of the dynamic relationship graph to score the importance of each knowledge element node in the graph, and obtains importance ranking data for distinguishing the importance of different knowledge elements. According to the semantic feature vectors (such as word embedding, topic model representation, etc.) of the knowledge element in the dynamic relationship graph, a clustering algorithm (such as K-means, DBSCAN, etc.) is used to divide the semantically similar knowledge element nodes into the same category to obtain node semantic feature clustering data. At the same time, the system divides the knowledge element nodes of the same relationship type and the edges connected to them into a subgraph based on the dynamic relationship graph and the importance ranking data, and obtains a node relationship subgraph set, such as a subgraph containing a causal relationship, a subgraph containing a time sequence relationship, etc. Combined with the node semantic feature clustering data and the node relationship subgraph set, the dynamic relationship graph is organized into a multi-level knowledge organization structure. For example, the semantic category can be used as a high-level structure, the relationship type can be used as a middle-level structure, and the importance ranking can be used as a bottom-level structure to construct a multi-level tree or graph knowledge organization structure data for the storage and retrieval of knowledge. According to the multi-level knowledge organization structure data, each knowledge element node and its related information (such as semantic features, relationship types, importance scores, etc.) are encapsulated into an independent knowledge unit to obtain knowledge unit data. At the same time, the system establishes a multi-level index table based on the multi-level knowledge organization structure, such as establishing a primary index based on semantic categories and a secondary index based on relationship types, so as to quickly locate and access the target knowledge unit. Each knowledge unit data is stored as an independent memory fragment to obtain memory fragment storage data, simulating the fragmentation characteristics of human brain memory. Then, the system establishes the association relationship between memory fragments based on the semantic association, relationship type, co-occurrence frequency and other information between memory fragments to form memory network data, simulating the association and reasoning process of human brain memory. The memory network data and the multi-level index table are integrated to construct a multi-level memory library. The multi-level index table is used to quickly locate the memory fragment where the target knowledge unit is located, while the memory network data is used to support the association, reasoning and evolution of knowledge. The resulting memory library can effectively organize, store and access massive knowledge information and support flexible knowledge application.
[0138] Preferably, step S5 comprises the following steps:
[0139] Step S51: Tracking the interaction flow through the sensor array to obtain interaction flow data; timestamping the interaction flow data and performing short-term memory caching to obtain timestamp data and short-term memory cache data;
[0140] Step S52: performing subject focus identification on the short-term memory cache data to obtain short-term memory data; performing time decay factor calculation according to the timestamp data to obtain the time decay factor;
[0141] Step S53: performing long-term memory retrieval on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory retrieval data;
[0142] Step S54: performing long-term memory association on the long-term memory retrieval data and the short-term memory data to obtain a candidate association data set; performing time weighting on the candidate association data set according to the time decay factor to obtain time weight data; performing semantic relationship reasoning on the candidate association data set according to the time weight data to obtain semantic relationship reasoning data; performing long-term memory screening on the candidate association data set according to the semantic relationship reasoning data to obtain long-term memory association data;
[0143] Step S55: Perform dynamic scenario update modeling based on the long-term memory association data and the semantic relationship reasoning data to obtain a dynamic scenario model; perform scenario model abstraction processing on the dynamic scenario model to obtain structured scenario data.
[0144] The present invention can more accurately record the time when an event occurs by introducing a timestamp mark when tracking the interaction flow, and provide support for subsequent operations such as time decay calculation and time weighting. By performing short-term memory caching on the interaction flow data, the user's interaction information can be temporarily stored within a certain time range to avoid the loss of important information. By adding a time decay factor during long-term memory retrieval, the relevance of the memory data can be reasonably adjusted according to the time of the interaction, so that the memory data closer to the current interaction time obtains a higher weight and better reflects the user's immediate intention. Time-weighted processing is performed on the candidate association data set, so that more time-closed data can be considered during semantic relationship reasoning, so that the reasoning result is more in line with the current scenario. By assigning time weights, the influence of data at different time points on the current semantic relationship reasoning can be quantified. By using semantic relationship reasoning data during long-term memory screening, long-term memories closely related to the current scenario can be selected more accurately to form high-quality long-term memory association data.
[0145] As an example of the present invention, refer to Figure 2 As shown, in this example, step S5 includes:
[0146] Step S51: Tracking the interaction flow through the sensor array to obtain interaction flow data; timestamping the interaction flow data and performing short-term memory caching to obtain timestamp data and short-term memory cache data;
[0147] In the embodiment of the present invention, the sensor array continuously tracks the interaction process between the user and the system, records the user's voice, text, behavior and other information, and forms interaction stream data. The system stores the latest interaction stream data in the short-term memory cache, and sets the cache capacity and time window, for example, retaining the interaction information within the last 10 minutes, and marking the timestamp to obtain short-term memory cache data with time stamp.
[0148] Step S52: performing subject focus identification on the short-term memory cache data to obtain short-term memory data; performing time decay factor calculation according to the timestamp data to obtain the time decay factor;
[0149] In the embodiment of the present invention, the short-term memory cache data is analyzed to identify the subject focus of the user's current interaction, such as the user is asking for information about a product, performing a task, etc., and the information related to the subject focus is extracted to form short-term memory data. At the same time, the time decay factor is calculated based on the timestamp, and the more recent the information, the smaller the time decay factor, such as using an exponential decay function.
[0150] Step S53: performing long-term memory retrieval on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory retrieval data;
[0151] In an embodiment of the present invention, a multi-level memory library is searched based on keywords, subject information and time decay factors in short-term memory data, and priority is given to searching long-term memory that is close to the time decay factor of short-term memory. For example, long-term memory is divided into regions with different time periods according to the time decay factor, and regions close to the short-term memory time period are searched preferentially to find long-term memory information related to the current interaction and obtain long-term memory retrieval data.
[0152] Step S54: performing long-term memory association on the long-term memory retrieval data and the short-term memory data to obtain a candidate association data set; performing time weighting on the candidate association data set according to the time decay factor to obtain time weight data; performing semantic relationship reasoning on the candidate association data set according to the time weight data to obtain semantic relationship reasoning data; performing long-term memory screening on the candidate association data set according to the semantic relationship reasoning data to obtain long-term memory association data;
[0153] In the embodiment of the present invention, the long-term memory retrieval data and the short-term memory data are subjected to association analysis, and the semantic connection between them is found through time series analysis, causal reasoning, similarity calculation and other methods to form a candidate association data set. Next, the candidate data set is weighted by the time decay factor, and the data closer to the current moment obtains a larger weight coefficient to generate time weight data. Then, based on the time weight data, the candidate association data set is subjected to semantic relationship reasoning to mine the implicit logical relationship and new semantic knowledge between the data. Finally, the candidate association data set is screened according to the semantic relationship reasoning data, and the noise data with weak semantic relevance is filtered out to obtain high-quality long-term memory association data.
[0154] Step S55: performing dynamic scenario update modeling according to the long-term memory association data and the semantic relationship reasoning data to obtain a dynamic scenario model; performing scenario model abstraction processing on the dynamic scenario model to obtain structured scenario data;
[0155] In an embodiment of the present invention, long-term memory association data and semantic relationship reasoning data are integrated into a dynamic scenario model. The scenario model is represented by a graph structure, including event nodes and relationship edges. The fusion process is to update the graph structure, such as inserting new event nodes, adjusting node attributes, adding and deleting relationship edges, etc., thereby realizing the dynamic evolution of the scenario model, so that it can continuously reflect changes in user intentions and interaction contexts. On this basis, the dynamic scenario model is abstracted, key events and relationships are extracted, redundant information is removed, and finally concise and structured scenario data is formed, which is convenient for subsequent storage, retrieval and utilization.
[0156] Preferably, step S6 comprises the following steps:
[0157] Step S61: decoding the structured scenario data to obtain scenario information decoding data;
[0158] Step S62: performing multi-level index positioning on the scene information decoding data according to the multi-level memory library to obtain memory fragment positioning data; performing semantic similarity sorting on the memory fragment positioning data according to the time decay factor to obtain answer semantic similarity sorting data;
[0159] Step S63: performing relationship path matching on the multi-level memory library according to the scenario information decoding data to obtain relationship path matching data;
[0160] Step S64: using the memory fragment link relationship data in the multi-level memory library to perform memory network traversal according to the answer semantic similarity sorting data and the relationship path matching data and perform time threshold control according to the time decay factor to obtain an initial candidate answer list;
[0161] Step S65: performing problem scenario analysis according to the scenario information decoding data to obtain problem scenario parameters;
[0162] Step S66: performing question answer matching calculation on the question scenario parameters according to the initial candidate answer list to obtain a question answer matching value; when the question answer matching value is greater than or equal to a preset matching threshold, the initial candidate answer list is used as a candidate answer list and step S69 is executed; when the question answer matching value is less than the preset matching threshold, step S67 is executed;
[0163] Step S67: Expand external knowledge according to the question scenario parameters to obtain external knowledge data; filter external information on the external knowledge data to obtain a list of external candidate answers;
[0164] Step S68: perform answer fusion sorting on the initial candidate answer list and the external candidate answer list to obtain a candidate answer list;
[0165] Step S69: Generate answers based on the candidate answer list to obtain accurate answer data.
[0166] The present invention can convert abstract scenario information into specific user intentions, question backgrounds, etc. by decoding structured scenario data, providing a clearer target for subsequent answer retrieval and matching. By using multi-level indexes to quickly locate memory fragments containing answers and sorting them according to semantic similarity, the answer search range can be narrowed, the retrieval efficiency can be improved, and answers related to the current scenario semantics can be given priority. Through relationship path matching, knowledge paths related to the current scenario can be found, such as the causal chain of events, the time development sequence, etc., to provide richer context information for answer retrieval and improve the accuracy and relevance of answers. Combining the results of semantic similarity sorting and relationship path matching, traversing the association relationship between memory fragments can more comprehensively consider various relevant information and find the candidate answer that best matches the current scenario. In-depth analysis of the problem scenario and extraction of key parameters such as time, place, user intention, emotion, etc. can more accurately characterize the characteristics of the problem and provide more refined evaluation indicators for subsequent answer matching. According to the problem scenario parameters, the matching degree of the candidate answers is calculated and the matching threshold is set, which can effectively filter out answers that do not meet the requirements, improve the reliability of the answers, and provide a judgment basis for whether external knowledge expansion is needed. When the appropriate answer cannot be found in the memory bank, external knowledge expansion can be used to obtain more extensive knowledge information, make up for the shortcomings of the memory bank, and improve the answer coverage of the system. By integrating and sorting the answers retrieved from the internal memory bank and the answers obtained from external knowledge expansion, the reliability and relevance of information from different sources can be comprehensively considered to obtain a more comprehensive and accurate list of candidate answers. Based on the final list of candidate answers, answers can be generated. The appropriate answer presentation method can be selected according to the user's needs and preferences, and the answers can be integrated and optimized to ultimately generate more accurate and easier-to-understand answer data.
[0167] As an example of the present invention, refer to Figure 3 As shown, in this example, step S6 includes:
[0168] Step S61: decoding the structured scenario data to obtain scenario information decoding data;
[0169] In an embodiment of the present invention, the system parses and understands structured scenario data, extracts key information, such as event type, time and place, character relationships, user intentions, etc., forms scenario information decoding data, and provides context information for subsequent answer retrieval and generation.
[0170] Step S62: performing multi-level index positioning on the scene information decoding data according to the multi-level memory library to obtain memory fragment positioning data; performing semantic similarity sorting on the memory fragment positioning data according to the time decay factor to obtain answer semantic similarity sorting data;
[0171] In the embodiment of the present invention, based on the keywords, theme information, etc. in the scenario information decoding data, the multi-level index table of the multi-level memory library is used to quickly locate the memory fragment containing the answer, and the memory fragment location data is obtained. Then, the system calculates the semantic similarity between each located memory fragment and the scenario information decoding data, for example, using indicators such as cosine similarity and Jaccard similarity, and sorts the memory fragment location data according to the time decay factor, and puts the recent related memory fragments in front, and obtains the answer semantic similarity sorting data.
[0172] Step S63: performing relationship path matching on the multi-level memory library according to the scenario information decoding data to obtain relationship path matching data;
[0173] In the embodiment of the present invention, the relationship information in the scenario information decoding data is decoded, and the relationship path matching is performed in the multi-level memory library to find the knowledge path related to the current scenario and obtain the relationship path matching data. For example, if the user asks "Why is it so cold today?", the system can find the knowledge path related to events such as "cold air moving south" and "temperature dropping" through relationship path matching.
[0174] Step S64: using the memory fragment link relationship data in the multi-level memory library to perform memory network traversal according to the answer semantic similarity sorting data and the relationship path matching data and perform time threshold control according to the time decay factor to obtain an initial candidate answer list;
[0175] In the embodiment of the present invention, the answer semantic similarity ranking data and the relationship path matching data are combined, and the memory fragment link relationship in the multi-level memory library is used to traverse the memory network, and the recent related memory fragments are given priority. A time threshold can be set, and only the memory fragments with a time decay factor less than the threshold are traversed to reduce the search scope and improve efficiency. Finally, an initial candidate answer list is obtained.
[0176] Step S65: performing problem scenario analysis according to the scenario information decoding data to obtain problem scenario parameters;
[0177] In an embodiment of the present invention, the context information decoding data is deeply analyzed to extract key parameters related to the question, such as time information, location information, user intention, emotional tendency, etc., to form question context parameters for evaluating the degree of match between the answer and the question.
[0178] Step S66: performing question answer matching calculation on the question scenario parameters according to the initial candidate answer list to obtain a question answer matching value; when the question answer matching value is greater than or equal to a preset matching threshold, the initial candidate answer list is used as a candidate answer list and step S69 is executed; when the question answer matching value is less than the preset matching threshold, step S67 is executed;
[0179] In the embodiment of the present invention, the matching degree of each answer in the initial candidate answer list is calculated according to the question scenario parameters, for example, the matching degree between the answer and the question in terms of time, location, semantics, etc. is calculated to obtain the question answer matching value. If the question answer matching value is greater than or equal to the preset matching threshold, it is considered that a suitable answer has been found, and the initial candidate answer list is used as the candidate answer list to enter the answer generation stage; otherwise, external knowledge expansion is required.
[0180] Step S67: Expand external knowledge according to the question scenario parameters to obtain external knowledge data; filter external information on the external knowledge data to obtain a list of external candidate answers;
[0181] In the embodiment of the present invention, according to the question scenario parameters, an external knowledge base or search engine is accessed to obtain external knowledge data related to the question. For example, if the user asks "What are the good restaurants nearby?", the system can access a map service or a review website to obtain information about nearby restaurants.
[0182] Step S68: perform answer fusion sorting on the initial candidate answer list and the external candidate answer list to obtain a candidate answer list;
[0183] In an embodiment of the present invention, the initial candidate answer list and the external candidate answer list are merged, for example, by using a ranking learning algorithm, a rule-based method, etc., to re-rank all candidate answers to obtain a final candidate answer list to ensure the comprehensiveness and accuracy of the answers.
[0184] Step S69: Generate answers based on the candidate answer list to obtain accurate answer data;
[0185] In an embodiment of the present invention, answers are generated based on answer information that ranks high in a candidate answer list, for example, by integrating and summarizing multiple answer information, or selecting the most appropriate answer format based on user preferences, ultimately obtaining accurate answer data and returning it to the user.
[0186] Preferably, step S65 includes the following steps:
[0187] Step S651: Perform time information analysis based on the scenario information decoded data to obtain problem time information data; perform subject area analysis based on the scenario information decoded data to obtain problem subject area data;
[0188] Step S652: Perform user intention analysis based on the context information decoded data to obtain problem user intention data; perform emotion analysis based on the context information decoded data to obtain problem emotion data;
[0189] Step S653: Extract question scenario parameters based on question time information data, question subject domain data, question user intention data, and question emotion data; obtain question scenario parameters.
[0190] The present invention can extract the time range and subject category of the question by analyzing the time information and subject area of the scenario information decoding data, so as to help the system locate the information related to the question more accurately when retrieving and matching the answer, eliminate the interference of irrelevant information, and improve the accuracy and efficiency of the answer. By analyzing the user's intention and emotion, the user's question can be understood more accurately, such as distinguishing whether the user is asking for information, seeking help or expressing emotion, as well as the user's emotional state, so as to select a more appropriate answer generation strategy, such as providing a more detailed explanation, a more friendly suggestion or a more empathetic response, and improve user satisfaction. By integrating information such as time information, subject area, user intention and emotional emotion into the problem scenario parameters, the characteristics of the problem can be more comprehensively characterized, and more refined evaluation indicators can be provided for subsequent answer matching, helping the system to select answers that are more matched with the question in terms of time, subject, intention, emotion, etc., and improving the accuracy of the answer and user experience.
[0191] In the embodiment of the present invention, the system uses natural language processing technology, such as named entity recognition, keyword extraction, etc., to analyze the situation information decoding data, identify the time information related to the question, such as a specific time point, time period, time relationship, etc., and form the problem time information data. At the same time, the system judges the subject area of the question, such as sports, entertainment, science and technology, etc., based on the keywords and subject words in the situation information decoding data, and forms the problem subject area data, which is used to narrow the answer search scope. According to the user's questioning method, modal particles, emoticons, etc. in the situation information decoding data, combined with the predefined user intention classification system, the user's real intention is judged, such as asking for information, seeking help, expressing emotions, etc., to form the problem user intention data. At the same time, the system uses sentiment analysis technology, such as a dictionary-based method, a machine learning-based method, etc., to identify the emotional tendency expressed by the user in the situation information decoding data, such as positive, negative, neutral, etc., to form the problem emotion data, which is used to more accurately understand the user's needs. The problem time information data, the problem subject area data, the problem user intention data and the problem emotion data are integrated to form structured problem situation parameters, which are used to characterize the key features of the problem, so as to perform more accurate answer matching and sorting in the future. For example, time information, subject area, user intention, emotional emotions, etc. can be used as different feature dimensions to construct a problem scenario parameter vector to characterize the overall characteristics of the problem.
[0192] Preferably, step S69 includes the following steps:
[0193] Step S691: parsing the candidate answer list to obtain candidate answer parsing data;
[0194] Step S692: Refining the answer content of the candidate answer analysis data to obtain refined answer data;
[0195] Step S693: performing answer multimodal enhancement on the refined answer data to obtain answer multimodal enhanced data;
[0196] Step S694: The refined answer data is refined according to the answer multimodal enhancement data to obtain accurate answer data.
[0197] The present invention can convert different forms of answer information into structured data by parsing candidate answers and extracting key information, such as text content, source, related pictures, etc., so as to facilitate the subsequent processing and analysis of the system, such as answer refinement, multimodal enhancement, etc. Refining the content of candidate answers, removing redundant information, and refining the core content can make the answers more concise and easy to understand, improve the user's reading experience, and facilitate the system to perform subsequent multimodal enhancement and answer integration. Performing multimodal enhancement on the refined answers, such as converting text answers into voice broadcasts, or generating related pictures and videos based on the answer content, can make the answers more vivid and vivid, improve user participation and interest, and enhance user experience. According to the answers after multimodal enhancement, the refined answers are finally integrated and optimized, such as selecting a suitable answer presentation method according to the user's preferences, or adjusting the expression style of the answer according to different interactive scenarios, etc., so that the final answer can be more accurate, more in line with the user's needs and expectations, and improve user satisfaction.
[0198] In an embodiment of the present invention, the system parses each answer in the candidate answer list, extracts key information, such as the text content, source, related pictures, videos, etc. of the answer, and forms structured candidate answer parsing data, which provides a basis for subsequent answer refinement and enhancement. The candidate answer parsing data is analyzed to remove redundant information and extract the core content, such as extracting the key sentences of the answer using text summarization technology, or extracting the key entities and relationships in the answer using information extraction technology, to form refined answer data, making the answer more concise and easy to understand. Based on the refined answer data, the answer is enhanced using multimodal generation technology, such as text-to-speech, text-to-image generation, etc., such as converting the text answer into a voice broadcast, or generating related pictures, videos, etc. according to the answer content, to form answer multimodal enhancement data, making the answer more vivid and vivid, and improving the user experience. Based on the answer multimodal enhancement data, the refined answer data is finally integrated and optimized, such as selecting a suitable answer presentation method according to the user's preference, or adjusting the expression style of the answer according to different interaction scenarios, etc., to finally form accurate answer data and return it to the user.
[0199] Preferably, the present invention further provides an automatic optimization system based on an intelligent data memory, which is used to execute the automatic optimization method based on an intelligent data memory as described above, and the automatic optimization system based on an intelligent data memory comprises:
[0200] The interaction capture module is used to collect original interaction data through the sensor array to obtain an original interaction data set; extract initial interaction events based on the original interaction data set to obtain initial interaction event data; and encapsulate the initial interaction event data into event data to obtain interaction event data;
[0201] The semantic extraction module is used to perform multimodal semantic association analysis on the interactive event data to obtain multimodal semantic association data; perform semantic extraction on the multimodal semantic association data to obtain knowledge metadata;
[0202] The relationship weaving module is used to analyze the relationship strength of knowledge metadata to obtain relationship strength data; perform relationship extraction processing based on the relationship strength data to obtain a dynamic relationship graph;
[0203] The memory precipitation module is used to sort the importance of nodes in the dynamic relationship graph to obtain importance sorting data; perform multi-dimensional hierarchical construction processing based on the importance sorting data to obtain multi-level knowledge organization structure data; perform hierarchical storage based on the multi-level knowledge organization structure data to obtain a multi-level memory library;
[0204] The situational awareness module is used to perform short-term memory tracking and time decay factor calculation through the sensor array to obtain short-term memory data and time decay factor; perform long-term memory association on the multi-level memory bank according to the short-term memory data and time decay factor to obtain long-term memory association data; perform situational awareness according to the long-term memory association data to obtain structured situational data;
[0205] The problem scenario retrieval module is used to decode the scenario information of the structured scenario data to obtain the scenario information decoding data; perform memory network retrieval on the scenario information decoding data according to the multi-level memory bank and the time attenuation factor to obtain an initial candidate answer list; perform problem scenario analysis based on the scenario information decoding data to obtain problem scenario parameters; perform question answer matching judgment on the problem scenario parameters according to the initial candidate answer list to obtain a candidate answer list; generate answers according to the candidate answer list to obtain accurate answer data.
[0206] The present invention can comprehensively record user interaction information through multimodal data collection and event extraction, and structure it into event data, providing a basis for subsequent semantic understanding and knowledge construction. Through multimodal semantic association analysis and semantic refinement, user intentions and interactive content can be deeply understood, key knowledge points can be extracted, and materials can be provided for building knowledge graphs. Through relationship strength analysis and relationship extraction, a dynamically updated knowledge graph can be constructed to reflect the semantic association and evolution process between knowledge points, providing support for knowledge reasoning and question-answering. Through node importance sorting and multi-dimensional hierarchical construction, the knowledge graph can be organized into a structured, hierarchical multi-level memory library to improve knowledge storage and retrieval efficiency. By introducing a time decay factor for short-term memory tracking and long-term memory association, the relevance weight of the information is adaptively adjusted according to the time distance of the interaction, and recent data is selectively focused on in the semantic association and reasoning process, so as to obtain long-term memory associations and semantic relationships that are closer to the current scenario. Through the steps of situation information decoding, memory network retrieval, question scenario analysis, and answer matching, the answer that best meets user needs can be retrieved and screened from the multi-level memory library, and accurate and easy-to-understand answer data can be finally generated. Therefore, the present invention provides an automatic optimization method and system based on an intelligent data memory library, which improves the ability to understand interactive data by effectively integrating and processing multimodal information such as text, voice, and images. Dynamic relationship graphs and multi-level memory libraries are used to achieve flexible organization and updating of knowledge. Through short-term memory tracking and long-term memory association, the perception of interactive scenarios is achieved and the accuracy of answers is improved. This method aims to achieve intelligent understanding of interactive data, knowledge organization, and accurate answer generation by simulating human memory and learning processes.
[0207] Applying this automatic optimization method based on intelligent data memory to intelligent assistants can help them manage their own knowledge more systematically, use experience more efficiently, understand user context more intelligently, and provide information services more comprehensively, ultimately improving the intelligence level of the system and the efficiency of problem solving.
[0208] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.
[0209] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may 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 the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.
Claims
1. An automatic optimization method based on intelligent data memory, characterized in that: The following steps are involved: Step S1: collecting original interaction data through a sensor array to obtain an original interaction data set; extracting initial interaction events according to the original interaction data set to obtain initial interaction event data; and encapsulating the initial interaction event data into event data to obtain interaction event data; Step S2: performing multimodal semantic association analysis on the interaction event data to obtain multimodal semantic association data; Perform semantic extraction on multimodal semantically related data to obtain knowledge metadata; Step S3: performing relationship strength analysis on the knowledge metadata to obtain relationship strength data; Relationship extraction is performed based on the relationship strength data to obtain a dynamic relationship graph, wherein step S3 is specifically as follows: Step S31: performing semantic space mapping on the knowledge metadata to obtain high-dimensional semantic space vector representation data; performing semantic distance measurement on the high-dimensional semantic space vector representation data to obtain semantic distance measurement data; Step S32: mining potential relationships of knowledge metadata according to semantic distance measurement data to obtain knowledge metadata potential relationship data; Inferring the relationship type based on the potential relationship data of the knowledge element to obtain the relationship type data; Step S33: Evaluate the relationship strength of the knowledge element potential relationship data to obtain relationship strength data; Calculate the relationship confidence according to the relationship type data and the relationship strength data to obtain the relationship confidence data; Step S34: updating the dynamic graph according to the knowledge element potential relationship data and the relationship confidence data to obtain a dynamic relationship graph; Step S4: sorting the importance of nodes in the dynamic relationship graph to obtain importance sorting data; performing multi-dimensional hierarchical construction processing based on the importance sorting data to obtain multi-level knowledge organization structure data; performing hierarchical storage based on the multi-level knowledge organization structure data to obtain a multi-level memory library; Step S5: Perform short-term memory tracking and time decay factor calculation through the sensor array to obtain short-term memory data and time decay factor; perform long-term memory association on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory association data; perform situation perception according to the long-term memory association data to obtain structured situation data; Step S6: Decode the scenario information of the structured scenario data to obtain scenario information decoding data; perform memory network retrieval on the scenario information decoding data according to the multi-level memory bank and the time attenuation factor to obtain an initial candidate answer list; perform problem scenario analysis based on the scenario information decoding data to obtain problem scenario parameters; perform question answer matching judgment on the problem scenario parameters according to the initial candidate answer list to obtain a candidate answer list; generate answers based on the candidate answer list to obtain accurate answer data.
2. The automatic optimization method based on intelligent data memory according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: collecting original interaction data through a sensor array to obtain an original interaction data set; performing multimodal synchronization on the original interaction data set to obtain multimodal synchronized data; Step S12: extracting the perception signal from the multimodal synchronization data to obtain multimodal perception signal data; and enhancing the signal features of the multimodal perception signal data to obtain multimodal signal feature enhanced data; Step S13: performing multimodal semantic fusion on the multimodal signal feature enhancement data to obtain multimodal semantic representation data; performing interaction intent analysis on the multimodal semantic representation data to obtain interaction intent data; performing interaction behavior annotation on the interaction intent data to obtain interaction behavior annotation data; Step S14: constructing structured events for the multimodal semantic representation data, the interaction intention data, and the interaction behavior annotation data to obtain initial interaction event data; Step S15: adding event metadata to the initial interaction event data to obtain interaction event metadata; performing interaction event encapsulation on the initial interaction event data and the interaction event metadata to obtain interaction event data.
3. The automatic optimization method based on intelligent data memory according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: performing event semantic deconstruction on the interactive event data to obtain event semantic unit data; performing text semantic analysis on the event semantic unit data to obtain text semantic data; performing voice and intonation analysis on the event semantic unit data to obtain voice and intonation data; performing image scene understanding on the event semantic unit data to obtain image scene data; Step S22: performing multimodal semantic association on the text semantic data, the voice intonation data and the image scene data to obtain multimodal semantic association data; Step S23: performing semantic role labeling on the multimodal semantic association data to obtain semantic role labeling data; performing semantic relationship extraction based on the semantic role labeling data to obtain a semantic relationship graph; performing key knowledge condensation on the semantic relationship graph to obtain key knowledge data; Step S24: encapsulate the key knowledge data, semantic relationship graph and multimodal semantic association data into knowledge metadata to obtain knowledge metadata.
4. The automatic optimization method based on intelligent data memory according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: sorting the knowledge element nodes in the dynamic relationship graph by node importance to obtain importance sorting data; Step S42: performing semantic feature clustering according to the dynamic relationship graph to obtain node semantic feature clustering data; performing relationship type division according to the dynamic relationship graph and importance ranking data to obtain a node relationship subgraph set; Step S43: constructing a multi-dimensional hierarchical structure of the node relationship subgraph set according to the node semantic feature clustering data to obtain multi-level knowledge organization structure data; Step S44: encapsulating knowledge units according to the multi-level knowledge organization structure data to obtain knowledge unit data; establishing multi-level indexes for the multi-level knowledge organization structure data to obtain a multi-level index table; Step S45: performing memory fragment storage on the knowledge unit data to obtain memory fragment storage data; performing memory fragment association on the memory fragment storage data to obtain memory network data; Step S46: Integrate the multi-level memory base according to the memory network data and the multi-level index table to obtain a multi-level memory base.
5. The automatic optimization method based on intelligent data memory according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: Tracking the interaction flow through the sensor array to obtain interaction flow data; timestamping the interaction flow data and performing short-term memory caching to obtain timestamp data and short-term memory cache data; Step S52: performing subject focus identification on the short-term memory cache data to obtain short-term memory data; performing time decay factor calculation according to the timestamp data to obtain the time decay factor; Step S53: performing long-term memory retrieval on the multi-level memory bank according to the short-term memory data and the time decay factor to obtain long-term memory retrieval data; Step S54: performing long-term memory association on the long-term memory retrieval data and the short-term memory data to obtain a candidate association data set; performing time weighting on the candidate association data set according to the time decay factor to obtain time weight data; performing semantic relationship reasoning on the candidate association data set according to the time weight data to obtain semantic relationship reasoning data; performing long-term memory screening on the candidate association data set according to the semantic relationship reasoning data to obtain long-term memory association data; Step S55: Perform dynamic scenario update modeling based on the long-term memory association data and the semantic relationship reasoning data to obtain a dynamic scenario model; perform scenario model abstraction processing on the dynamic scenario model to obtain structured scenario data.
6. The automatic optimization method based on intelligent data memory according to claim 1, characterized in that: Step S6 includes the following steps: Step S61: decoding the structured scenario data to obtain scenario information decoding data; Step S62: performing multi-level index positioning on the scene information decoding data according to the multi-level memory library to obtain memory fragment positioning data; performing semantic similarity sorting on the memory fragment positioning data according to the time decay factor to obtain answer semantic similarity sorting data; Step S63: performing relationship path matching on the multi-level memory library according to the scenario information decoding data to obtain relationship path matching data; Step S64: using the memory fragment link relationship data in the multi-level memory library to perform memory network traversal according to the answer semantic similarity sorting data and the relationship path matching data and perform time threshold control according to the time decay factor to obtain an initial candidate answer list; Step S65: performing problem scenario analysis according to the scenario information decoding data to obtain problem scenario parameters; Step S66: performing question answer matching calculation on the question scenario parameters according to the initial candidate answer list to obtain a question answer matching value; when the question answer matching value is greater than or equal to a preset matching threshold, the initial candidate answer list is used as a candidate answer list and step S69 is executed; when the question answer matching value is less than the preset matching threshold, step S67 is executed; Step S67: Expand external knowledge according to the question scenario parameters to obtain external knowledge data; filter external information on the external knowledge data to obtain a list of external candidate answers; Step S68: perform answer fusion sorting on the initial candidate answer list and the external candidate answer list to obtain a candidate answer list; Step S69: Generate answers based on the candidate answer list to obtain accurate answer data.
7. The automatic optimization method based on intelligent data memory according to claim 6, characterized in that: Step S65 includes the following steps: Step S651: Perform time information analysis based on the scenario information decoded data to obtain problem time information data; perform subject area analysis based on the scenario information decoded data to obtain problem subject area data; Step S652: Perform user intention analysis based on the context information decoded data to obtain problem user intention data; perform emotion analysis based on the context information decoded data to obtain problem emotion data; Step S653: Extract question scenario parameters based on question time information data, question subject domain data, question user intention data, and question emotion data; obtain question scenario parameters.
8. The automatic optimization method based on intelligent data memory according to claim 6, characterized in that: Step S69 includes the following steps: Step S691: parsing the candidate answer list to obtain candidate answer parsing data; Step S692: Refining the answer content of the candidate answer analysis data to obtain refined answer data; Step S693: performing answer multimodal enhancement on the refined answer data to obtain answer multimodal enhanced data; Step S694: The refined answer data is refined according to the answer multimodal enhancement data to obtain accurate answer data.
9. An automatic optimization system based on intelligent data memory, characterized in that: For executing the automatic optimization method based on the intelligent data memory as claimed in claim 1, the automatic optimization system based on the intelligent data memory comprises: The interaction capture module is used to collect original interaction data through the sensor array to obtain an original interaction data set; extract initial interaction events based on the original interaction data set to obtain initial interaction event data; and encapsulate the initial interaction event data into event data to obtain interaction event data; The semantic extraction module is used to perform multimodal semantic association analysis on the interactive event data to obtain multimodal semantic association data; perform semantic extraction on the multimodal semantic association data to obtain knowledge metadata; The relationship weaving module is used to analyze the relationship strength of knowledge metadata to obtain relationship strength data; perform relationship extraction processing based on the relationship strength data to obtain a dynamic relationship graph; The memory precipitation module is used to sort the importance of nodes in the dynamic relationship graph to obtain importance sorting data; perform multi-dimensional hierarchical construction processing based on the importance sorting data to obtain multi-level knowledge organization structure data; perform hierarchical storage based on the multi-level knowledge organization structure data to obtain a multi-level memory library; The situational awareness module is used to perform short-term memory tracking and time decay factor calculation through the sensor array to obtain short-term memory data and time decay factor; perform long-term memory association on the multi-level memory bank according to the short-term memory data and time decay factor to obtain long-term memory association data; perform situational awareness according to the long-term memory association data to obtain structured situational data; The problem scenario retrieval module is used to decode the scenario information of the structured scenario data to obtain the scenario information decoding data; perform memory network retrieval on the scenario information decoding data according to the multi-level memory bank and the time attenuation factor to obtain an initial candidate answer list; perform problem scenario analysis based on the scenario information decoding data to obtain problem scenario parameters; perform question answer matching judgment on the problem scenario parameters according to the initial candidate answer list to obtain a candidate answer list; generate answers according to the candidate answer list to obtain accurate answer data.
