Method and system for large model agent interactive question and answer task decision-making

Abstract

The application provides a large model intelligent agent interactive question and answer task decision method and system, relates to the natural language processing technical field, and the method comprises the following steps: step 1, receiving user input through multi-round dialogue, extracting key information to form an initial memory node set, while analyzing the fuzziness of the query intention and the information integrity, and generating an uncertainty evaluation result; step 2, based on the uncertainty evaluation result, the timeliness and importance degree of each node are evaluated, an enhanced memory node set is generated, a dynamic memory structure is constructed according to the semantic and time sequence relationship, and a structured memory network is formed. The application improves the understanding accuracy of the intelligent agent on the multi-round dialogue context, the logical coherence of the decision response and the continuous adaptability of the task processing capability.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a decision-making method and system for large-scale intelligent agent interactive question-answering tasks. Background Technology

[0002] In large-scale intelligent agent interactive question-answering tasks, traditional techniques, when matching dialogue context, mostly focus on semantic similarity matching with the current query, rarely constructing a demand evolution path according to the dialogue sequence, and lacking the traceability and verification of the logical connection between historical information and current demand. This may cause the agent to ignore the progressive relationship of user demand, resulting in a weak connection between the output response and the logic of the historical dialogue, or even a one-sided problem. For example, in the dialogue between the student and the agent, the student first expresses that they want to practice triangle-related proof problems, then says that it would be best to have problems related to graph congruence, and finally adds that the problem stem should preferably include the condition that the sides are equal. When the traditional agent receives the last round of query, it only performs semantic matching for triangle proof problems and the condition that the sides are equal, recommending proof problems containing the relevant conditions, but does not trace the core premise of graph congruence in the previous dialogue. Some of the recommended problems are proof problems about the relationship between the sides of ordinary triangles, which is disconnected from the logical demand of the student to practice congruence-related proofs, resulting in the recommended content not meeting expectations. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a decision-making method and system for interactive question-answering tasks of large-scale intelligent agents, so as to improve the accuracy of the agent's understanding of multi-turn dialogue context, the logical coherence of decision response, and the continuous adaptability of task processing capabilities.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0005] Firstly, a decision-making method for large-scale intelligent agent interactive question-answering tasks, the method comprising:

[0006] Step 1: Receive user input through multi-turn dialogue, extract key information to form an initial memory node set, and analyze the ambiguity of the query intent and the completeness of information to generate an uncertainty assessment result;

[0007] Step 2: Based on the uncertainty assessment results, evaluate the timeliness and importance of each node, generate a set of memory nodes with corresponding metadata identifiers, and construct a dynamic memory structure according to semantic and temporal relationships to form a structured memory network.

[0008] Step 3: Based on the structured memory network, identify the core memory nodes related to the current dialogue context, and trace the associated nodes along the forward and backward dialogue sequence to define the local memory range;

[0009] Step 4: Within the local memory range, select key node sequences to construct a temporal evolution path, analyze the changes in node states in the temporal evolution path, calculate the logical consistency metric, and generate path correction values.

[0010] Step 5: Integrate the path correction value with the structured memory network. When a new round of user queries is received, adjust the retrieval strategy and similarity calculation, and combine the uncertainty assessment results to obtain the target memory node set from the structured memory network.

[0011] Step 6: Combine the target memory node set with the current user query input large language model to generate a task decision response. Update the importance and correlation of relevant nodes in the structured memory network based on the interaction results to achieve continuous evolution of task decision-making capabilities.

[0012] Secondly, a large-scale intelligent agent interactive question-answering task decision-making system includes:

[0013] The information extraction module is used to receive user input through multi-turn dialogue, extract key information to form an initial memory node set, and analyze the ambiguity of the query intent and the completeness of the information to generate uncertainty assessment results.

[0014] The memory construction module is used to evaluate the timeliness and importance of each node based on the uncertainty assessment results, generate a set of memory nodes with corresponding metadata identifiers, and construct a dynamic memory structure according to semantic and temporal relationships to form a structured memory network.

[0015] The context association module is used to identify core memory nodes related to the current dialogue context based on a structured memory network, and to trace associated nodes in the forward and backward directions of the dialogue sequence to define the local memory range.

[0016] The path verification module is used to select key node sequences to construct a time-series evolution path within the local memory range, analyze the changes in the node states in the time-series evolution path, calculate the logical consistency metric, and generate path correction values.

[0017] The retrieval optimization module integrates path correction values ​​with the structured memory network. When a new round of user queries is received, the retrieval strategy and similarity calculation are adjusted, and the target memory node set is obtained from the structured memory network by combining the uncertainty assessment results.

[0018] The response update module is used to combine the target memory node set with the current user query input large language model to generate a task decision response. Based on the interaction results, it updates the importance and correlation of relevant nodes in the structured memory network to achieve continuous evolution of task decision-making capabilities.

[0019] Thirdly, a computing device includes:

[0020] One or more processors;

[0021] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0022] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0023] The above-described solution of the present invention has at least the following beneficial effects:

[0024] By assessing uncertainty in the initial stage, we can identify ambiguities and information gaps in user queries in advance, thus anchoring the decision-making direction. The structured memory network organizes memory nodes according to semantics and time sequence. With the delineation of local memory range, it reduces interference from irrelevant information and avoids the resource consumption of traversing the entire memory, thereby improving retrieval and response efficiency. At the same time, the logical consistency measurement and path correction mechanism of the time sequence evolution path avoids logical breaks in node state transitions, ensuring that the decision-making process conforms to the context logic and reduces conflicts between responses.

[0025] By dynamically updating the importance and relationships of memory nodes, the memory network can adaptively adjust with the interaction process, avoiding the solidification of decision-making patterns. In multi-turn dialogues, structured memory and temporal path analysis can accurately capture long-term contextual relationships. This closed-loop design of accurate retrieval, logical verification and dynamic iteration allows the agent to maintain the relevance of the current decision in complex question-and-answer tasks, and continuously enhance its ability to handle more complex scenarios through continuous learning. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the decision-making method for large-scale intelligent agent interactive question-answering tasks provided in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of a large-scale intelligent agent interactive question-answering task decision system provided in an embodiment of the present invention. Detailed Implementation

[0028] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0029] like Figure 1 As shown, embodiments of the present invention propose a decision-making method for large-scale intelligent agent interactive question-answering tasks, the method comprising the following steps:

[0030] Step 1: Receive user input through multi-turn dialogue, extract key information to form an initial memory node set, and analyze the ambiguity of the query intent and the completeness of information to generate an uncertainty assessment result;

[0031] Step 2: Based on the uncertainty assessment results, evaluate the timeliness and importance of each node, generate a set of memory nodes with corresponding metadata identifiers, and construct a dynamic memory structure according to semantic and temporal relationships to form a structured memory network.

[0032] Step 3: Based on the structured memory network, identify the core memory nodes related to the current dialogue context, and trace the associated nodes along the forward and backward dialogue sequence to define the local memory range;

[0033] Step 4: Within the local memory range, select key node sequences to construct a temporal evolution path, analyze the changes in node states in the temporal evolution path, calculate the logical consistency metric, and generate path correction values.

[0034] Step 5: Integrate the path correction value with the structured memory network. When a new round of user queries is received, adjust the retrieval strategy and similarity calculation, and combine the uncertainty assessment results to obtain the target memory node set from the structured memory network.

[0035] Step 6: Combine the target memory node set with the current user query input large language model to generate a task decision response. Update the importance and correlation of relevant nodes in the structured memory network based on the interaction results to achieve continuous evolution of task decision-making capabilities.

[0036] In this embodiment of the invention, uncertainty assessment in the initial stage identifies ambiguities and information gaps in user queries in advance, providing a foundation for decision-making. The structured memory network organizes memory nodes semantically and temporally, and with the delineation of local memory ranges, it reduces interference from irrelevant information and avoids the resource consumption of traversing the entire memory, thereby improving retrieval and response efficiency. At the same time, the logical consistency measurement and path correction mechanism of the temporal evolution path avoids logical breaks in node state transitions, ensuring that the decision-making process conforms to the context logic and reduces conflicts between responses.

[0037] By dynamically updating the importance and relationships of memory nodes, the memory network can adaptively adjust with the interaction process, avoiding the solidification of decision-making patterns. In multi-turn dialogues, structured memory and temporal path analysis can accurately capture long-term contextual relationships. This closed-loop design of accurate retrieval, logical verification and dynamic iteration allows the agent to maintain the relevance of the current decision in complex question-and-answer tasks, and continuously enhance its ability to handle more complex scenarios through continuous learning.

[0038] In a preferred embodiment of the present invention, step 1 above, which involves receiving user input through multi-turn dialogue, extracting key information to form an initial memory node set, and simultaneously analyzing the ambiguity and information completeness of the query intent to generate an uncertainty assessment result, may include:

[0039] In this embodiment of the invention, step 110 involves receiving user input through continuous multi-round dialogue, identifying core elements, behavioral instructions, and contextual associations as key information, and integrating the key information into an initial memory node set. Specifically, this includes: First, the agent continuously receives each input from the user during the multi-round dialogue. For example, in a dialogue between the user and the agent about practicing geometric proofs, the agent sequentially receives multiple rounds of input, such as the user saying they want to practice proofs related to triangles, preferably related to geometric congruence, and the question stem preferably containing the condition that sides are equal. Next, key information is identified for each round of user input. Core elements refer to the core objects and knowledge points involved in the user's needs. For example, from "I want to practice proofs related to triangles," the core elements are identified as triangles and proofs; from "preferably related to geometric congruence," the core element is geometric congruence; and from "preferably containing the condition that sides are equal," the core element is equal sides. Behavioral instructions refer to the actions or directions the user wants to perform, such as from above... In each round of input, the behavioral instruction is identified as "practice." Contextual association refers to the relationship between different rounds of input and the implied needs of the input content. For example, identifying congruent figures further restricts the proof of triangles, and equal sides provide supplementary conditions for proofs related to congruent figures. The overall contextual association indicates that the user's needs are progressively refined geometric proof practice requirements. Finally, each set of identified key information is integrated into an independent initial memory node. Each initial memory node contains the core elements, behavioral instructions, and contextual associations corresponding to that round of input. All independent initial memory nodes are then summarized to form an initial memory node set. For example, the first initial memory node is the core elements "triangle," "proof," "practice," and the user's initial request to practice triangle proofs. The second initial memory node is the user's supplementary condition for congruent figures to the triangle proof requirement, and so on, until a complete initial memory node set is formed.

[0040] Step 111: Based on the initial set of memory nodes, analyze the user's query intent. This involves calculating the coverage between identified keywords and the preset task domain, and assessing the ambiguity of the intent by combining the contextual relevance between different memory nodes. Specifically, this includes: First, based on the initial set of memory nodes, analyze the user's true intent for each round of queries. For example, the initial intent might be to practice geometry proofs, but the specific type of proof and the core knowledge points tested need further clarification. Next, calculate the coverage between identified keywords and the preset task domain. The identified keywords are the core elements extracted from the initial set of memory nodes. Combined with behavioral instructions, such as triangle, proof, and practice, the preset task domain is a pre-defined range of tasks that the agent can process. Taking geometry learning as an example, the preset task domain includes sub-domains such as geometry proof practice, geometry concept learning, and geometry drawing. The geometry proof practice sub-domain further covers different shape types such as triangles, quadrilaterals, and circles, as well as proof practice on different knowledge points such as congruence, similarity, and area calculation. When calculating the coverage, the identified keywords "triangle," "proof," and "practice" are compared with the sub-domains in the preset task domain to determine which part of the preset task domain the identified keywords can cover. For example, how many sub-domains can "practice" and "proof" cover? The subdomain for proof practice is defined, but triangles can only cover some graphic types within this subdomain, not quadrilaterals, circles, or other graphic types. This determines the coverage of the identified keywords and the preset task domain. Then, the contextual association between different memory nodes is calculated, analyzing the closeness of the connections between the key information contained in each node of the initial memory node set. For example, if the key information of the first initial memory node is triangle, proof, and practice, and the user initially requests practice of triangle proofs, if a new initial memory node is generated, the connection between the new node and the key information in the first node needs to be analyzed, even if there are only one or a few nodes in the initial stage. It is also necessary to analyze the degree of correlation between the core elements within the node, behavioral instructions, and contextual relationships, such as the close correlation between triangles and proof problems, and the close correlation between exercises and triangle proof problems. Finally, by combining the coverage of the identified keywords and the preset task domain, as well as the contextual correlation between different memory nodes, the degree of ambiguity of the user's intent is comprehensively evaluated. If the coverage is small and the contextual correlation is low, it indicates that there is a lot of unclear content in the user's intent, and the degree of intent ambiguity is high. For example, based solely on the key information of triangles, proof problems, and exercises, it is impossible to determine whether the user wants to practice triangle congruence proofs, similarity proofs, or other types of proof problems. In this case, the degree of intent ambiguity is high.

[0041] Step 112: Based on the degree of ambiguity of the intent, analyze the current information completeness. This involves comparing existing key information with the information items required for the preset task to identify the proportion of missing necessary information and quantify information completeness. Specifically, this includes: First, based on the degree of intent ambiguity assessed in Step 111, determine the focus of the analysis of the current information completeness. If the degree of intent ambiguity is high, it indicates that there may be a lot of unclear content in the current information, requiring a focus on analyzing which necessary information has not yet been acquired. Next, clarify the information items required for the preset task. These are pre-defined categories of key information needed to complete the task, based on the type of task being processed by the agent. Taking a geometry proof exercise task as an example, the required information items include the type of graph corresponding to the proof, the core knowledge points being tested, the key conditions to be included in the question stem, and the difficulty level of the exercise. Then, compare the key information already identified in the initial memory node set with the required information items for the preset task one by one to check whether the existing key information covers all types of content in the required information items for the preset task. For example, if the identified key information is not fully covered, the information will be compared with the key information in the preset task. The information items required for the preset task were compared with those for triangles, proof problems, and exercises. It was found that triangles in the figure type and proof problems in the task type were covered, but core knowledge points such as congruence and similarity, key conditions in the question stem such as equal sides and angles, and difficulty level were not covered. Then, the proportion of missing necessary information was identified. Necessary information refers to information categories that are indispensable to completing the preset task. For example, in the geometry proof problem exercise task, figure type, core knowledge points, and key conditions in the question stem are necessary information. Difficulty level can be selected as necessary information based on the actual situation. First, the total number of necessary information items required for the preset task was counted. Then, the number of necessary information items that were not covered by the existing key information was counted. The proportion of missing necessary information was obtained by dividing the number of uncovered necessary information items by the total number of necessary information items. Finally, the current information completeness was quantified based on this proportion. The higher the missing proportion, the lower the information completeness. For example, if the preset necessary information items are three categories: figure type, core knowledge points, and key conditions in the question stem, and one category is covered while two categories are not covered, the missing proportion is high and the information completeness is low.

[0042] Step 113 integrates the ambiguity of intent and the completeness of information to generate a quantified uncertainty assessment result. Specifically, this includes: First, determining the quantified values ​​for the ambiguity of intent and the completeness of information. The quantified value for the ambiguity of intent is derived from the comprehensive analysis of the coverage area and contextual relevance in Step 111; the smaller the coverage area and the lower the contextual relevance, the higher the quantified value for the ambiguity of intent. The quantified value for the completeness of information is derived from the proportion of missing necessary information in Step 112; the higher the proportion of missing information, the lower the quantified value for the completeness of information. Next, the two quantified values ​​are integrated by comprehensively considering their impact on the uncertainty of user needs. The degree of influence of intent ambiguity reflects the clarity of the user's demand direction, while information completeness reflects the completeness of the user's demand details. Both affect uncertainty. Therefore, the quantitative values ​​of intent ambiguity and information completeness are directly added together to obtain a comprehensive quantitative value. Finally, this comprehensive quantitative value is used as the uncertainty assessment result of the user's query. This result can comprehensively reflect the uncertainty of the user's current demand. For example, if the quantitative value of intent ambiguity is high and the quantitative value of information completeness is low, the comprehensive quantitative value after adding the two is high, indicating that the uncertainty of the user's demand is high. That is, the user's current query has many unclear aspects in terms of demand direction and demand details.

[0043] By comprehensively capturing key information in multi-round dialogues, accurately assessing the ambiguity of user intent and the completeness of information, and forming a comprehensive uncertainty assessment result, the intelligent agent's response is ensured to conform to the evolution of user needs, avoiding problems such as a break from historical dialogue logic or deviation from actual needs.

[0044] In a preferred embodiment of the present invention, step 2 above, based on the uncertainty assessment results, evaluates the timeliness and importance of each node, generates a set of memory nodes with corresponding metadata identifiers, and constructs a dynamic memory structure according to semantic and temporal relationships to form a structured memory network, which may include:

[0045] In this embodiment of the invention, step 220 involves filtering the initial memory node set based on the uncertainty assessment results. A timeliness score is calculated based on the creation time and last access time of each node, and an importance score is calculated by combining the frequency of node appearance and the strength of association in the dialogue. Specifically, this includes: First, a preliminary screening of the initial memory node set is performed based on the uncertainty assessment results. If the uncertainty assessment results show a high degree of uncertainty regarding user needs, it indicates that more nodes containing potentially key information need to be retained. In this case, the retention conditions are appropriately relaxed during the screening, retaining most nodes in the initial memory node set. If the uncertainty assessment results show a low degree of uncertainty regarding user needs, it indicates that the existing information is relatively clear, and obviously irrelevant or duplicate nodes can be filtered out. Next, the timeliness score of each node is calculated. First, the current dialogue time is determined. For each initial memory node, its corresponding creation time and last access time are extracted. The creation time of the node is subtracted from the current dialogue time to obtain the time interval from the node's creation to the present. Then, the last access time of the node is subtracted from the current dialogue time to obtain the time interval from the node's last access to the present. The two time intervals are weighted according to a preset time weight. The shorter the time interval, the higher the weighted score. The two weighted scores are then added together to obtain the timeliness score of the node. For example, if a node's creation time is close to the present and its last access time is also close to the present, both weighted scores are high, resulting in a high timeliness score for the node. Next, the importance score of each node is calculated. First, the frequency of each node's appearance in multiple rounds of dialogue is counted, i.e., the number of times the core elements, behavioral instructions, and other key information contained in the node appear in all dialogue rounds. Then, the association strength of the node is counted, i.e., the number of times the node has contextual associations with other initial memory nodes. The more associations, the higher the association strength. The frequency of the node's appearance and the association strength are added together to obtain the importance score of the node. For example, if a node related to graph congruence appears multiple times in the dialogue and is associated with nodes such as triangle proofs and edge congruence, its frequency of appearance and association strength scores are both high, resulting in a high importance score. Finally, the initial screening of the initial memory node set and the calculation of the timeliness and importance scores of each node are completed.

[0046] Step 221 involves comprehensively calculating the timeliness and importance scores to obtain a comprehensive evaluation value for each node. Based on this comprehensive evaluation value, the initial memory nodes are screened and adjusted to generate a set of memory nodes with corresponding metadata identifiers. Specifically, this includes: First, for each initial memory node initially screened in Step 220, its timeliness and importance scores are comprehensively calculated by adding the two scores together. This method balances the node's timeliness and content importance, ensuring the comprehensive evaluation value fully reflects the node's value. Next, a screening threshold for the comprehensive evaluation value is set. This threshold is determined based on the needs of the current dialogue task. For tasks with progressively more detailed requirements, such as geometry proof exercises, the screening threshold is set to retain nodes with high timeliness and importance. Then… The comprehensive evaluation value of each node is compared with the screening threshold. Nodes with a comprehensive evaluation value higher than the screening threshold are retained, while those with a comprehensive evaluation value lower than the screening threshold are removed. These removed nodes are mostly those that were created a long time ago, have been accessed infrequently, and have low relevance to the current needs. For example, nodes related to quadrilateral proof problems that were only mentioned at the beginning of the dialogue but not subsequently addressed. Then, the retained nodes are adjusted. If there are nodes with duplicate content or highly similar semantics, they are merged into one node, and the key information in each node is supplemented and integrated to ensure that the merged node can fully cover the relevant needs. If some nodes have incomplete key information, their content is supplemented and improved in combination with the dialogue context to make the node information more accurate. Finally, all nodes after screening and adjustment are summarized to generate a set of memory nodes with corresponding metadata identifiers.

[0047] Step 222 involves analyzing the nodes in the memory node set with corresponding metadata identifiers to establish semantic similarity relationships and temporal adjacency relationships between nodes. Specifically, this includes: First, analyzing the key information of each node in the enhanced memory node set, focusing on extracting the core elements, behavioral instructions, and contextual associations contained in the nodes. For example, extracting core elements and behavioral instructions such as triangles, proof problems, geometric congruence, side equality, and exercises from each node. Next, establishing semantic similarity relationships between nodes involves comparing the key information of any two nodes to determine the degree of overlap in their core elements and the similarity in their contextual associations. If the core elements of two nodes largely overlap and their contextual associations revolve around the same demand direction, it indicates semantic similarity, and a semantic similarity relationship can be established between these two nodes. For example, a node containing triangles, proof problems, and geometric congruence is similar to a node containing triangles, proof problems, and geometric congruence. The nodes related to triangles, proofs, and congruence proofs share highly overlapping core elements, and their contexts all revolve around practicing triangle congruence proofs. A semantic similarity relationship is established between these nodes. Then, a temporal adjacency relationship is established between the nodes. Based on the chronological order of multiple dialogues, the dialogue round corresponding to each node in the memory-enhancing node set is determined. Nodes are arranged according to the order of these dialogue rounds. If two nodes correspond to adjacent dialogue rounds (i.e., the preceding node corresponds to a certain round of dialogue, and the following node corresponds to the next round), a temporal adjacency relationship is established between these two nodes. For example, the node corresponding to "I want to practice triangle-related proofs" and the node corresponding to "I prefer questions related to geometric congruence" correspond to two adjacent rounds of dialogue, establishing a temporal adjacency relationship between them. Finally, the analysis of all memory-enhancing nodes is completed, clarifying the semantic similarity and temporal adjacency relationships between the nodes.

[0048] Step 223: Based on semantic similarity relationships and dialogue temporal adjacency relationships, construct a dynamic memory structure to form an updated structured memory network. Specifically, this includes: First, based on enhanced memory nodes and combined with established dialogue temporal adjacency relationships, arrange nodes sequentially according to the order of dialogue rounds to form a temporal node chain. This chain can intuitively present the changing process of user needs as the dialogue progresses. For example, nodes corresponding to triangle proof exercises, triangle congruence proof exercises, and triangle congruence proof exercises with equal sides are arranged according to the dialogue order, forming a temporal node chain with progressively refined needs. Next, based on the semantic similarity relationships between nodes, perform semantic clustering on the nodes in the temporal node chain, grouping semantically similar nodes into the same semantic group. Each semantic group revolves around a core need theme; for example, all nodes involving triangle congruence proofs are grouped into a single semantic group. The system first groups nodes related to the difficulty of triangle proofs into a separate semantic group. Connections between different semantic groups are established through node associations. Then, a dynamic memory structure is constructed. This structure uses a chronological node chain as its vertical framework to reflect the temporal evolution of the requirements, and semantic groups formed by semantic clustering as its horizontal associations to reflect the semantic connections within the requirements. The vertical framework and horizontal associations intertwine, ensuring that each node clearly understands its position in the dialogue sequence and is clearly associated with other semantically similar nodes. Finally, based on this dynamic memory structure, an updated structured memory network is formed. In this network, each step of the user's requirements refinement corresponds to a specific node, and the chronological progression and semantic connections between nodes are clearly visible. For example, the evolutionary path from triangle proofs to proofs of triangle congruence with equal sides can be clearly seen, as well as the connection between each node and the core semantics of congruence proofs.

[0049] By selecting nodes with timeliness and importance to form an enhanced memory set, and then constructing a structured memory network based on semantic and temporal relationships, it can accurately retain key information about user needs and clearly present the evolution of needs as the dialogue progresses, avoiding the problem of intelligent agents omitting key requirements or experiencing breaks in historical dialogue logic when responding.

[0050] In a preferred embodiment of the present invention, step 3 above, based on a structured memory network, identifies core memory nodes related to the current dialogue context and traces associated nodes forward and backward along the dialogue timeline to define the local memory range, may include:

[0051] In this embodiment of the invention, step 330 involves performing similarity matching in a structured memory network based on the semantic features of the current dialogue context to identify the core memory node most relevant to the current query semantics. Specifically, this includes: first, extracting the semantic features of the current dialogue context, which is the user's latest input content and the immediately preceding dialogue content. Core elements, behavioral instructions, and contextual associations are extracted as semantic features. For example, if the user's last input question stem ideally includes the condition of equal sides, the semantic features of the current dialogue context are: core elements triangle, proof question, geometric congruence, equal sides; behavioral instruction practice; and contextual association requiring the addition of the condition of equal sides to the proof question stem. Next, all nodes in the structured memory network are traversed, and the core elements, behavioral instructions, and contextual associations contained in each node are extracted one by one. Then, the semantic features of the current dialogue context are compared with the key information of each node for similarity matching calculation. First, the similarity of the core elements of both is statistically analyzed. The score is determined by the number of overlapping elements. A higher overlap rate results in a higher score. Next, the consistency of behavioral instructions is assessed: completely identical instructions receive the highest score, followed by partially related instructions, and completely unrelated instructions receive the lowest score. Then, the similarity of contextual associations is analyzed. Contextual associations revolving around the same need and with matching levels of detail receive the highest score, those with similar directions but significantly different levels of detail receive the next highest score, and those with different directions receive the lowest score. The scores for core element overlap, behavioral instruction consistency, and contextual association similarity are summed to obtain the total similarity score between the node and the semantic features of the current dialogue context. Finally, the total similarity scores of all nodes are compared, and the node with the highest total similarity score is selected as the core memory node most relevant to the current query semantics. For example, in a structured memory network, nodes containing core elements such as triangles, proof problems, geometric congruence, and equal sides; behavioral instruction practice; and contextual associations such as user-supplemented triangle congruence proof problems requiring equal sides have the highest total similarity score with the semantic features of the current dialogue context, and this node is the core memory node.

[0052] Step 331: Based on the core memory node, retrieve nodes generated after the initial dialogue sequence in a forward direction, while simultaneously tracing back nodes generated before, extracting all directly and indirectly related nodes to form an extended set of related nodes. Specifically, this includes: First, locating the core memory node in the temporal node chain of the structured memory network, identifying the dialogue round corresponding to that node (e.g., the core memory node corresponds to the 3rd round of dialogue). The temporal node chain is arranged sequentially from the 1st to the 3rd round of dialogue. Next, retrieval is performed in a forward direction, starting from the dialogue round corresponding to the core memory node, searching backward for all nodes generated after that round. If the core memory node corresponds to the last round of dialogue and no new nodes are generated subsequently, the forward retrieval yields no results. If new dialogue rounds and corresponding nodes exist, all these nodes are extracted. Then, tracing back in the dialogue sequence, starting from the dialogue round corresponding to the core memory node, searching forward for all nodes generated before that round (e.g., the core memory node corresponds to the 3rd round of dialogue), tracing back to extract nodes corresponding to the 2nd and 1st rounds of dialogue. These nodes record the user's needs from the initial request... The process involves gradual refinement. For example, the core elements corresponding to the second round of dialogue are triangles, proof problems, and graph congruence; behavioral instruction practice; and context-related nodes where users limit triangle proof problems to congruence-related nodes. The core elements corresponding to the first round of dialogue are triangles, proof problems, and behavioral instruction practice; and the nodes where users propose triangle proof practice. Then, all directly and indirectly related nodes are extracted. Directly related nodes are those with semantic similarity or temporal adjacency to the core memory node, such as the second round of dialogue node which is temporally adjacent to the core memory node, and the core elements triangles, congruence proofs, and edge equality that are semantically similar to the core memory node; and nodes for behavioral instruction practice. Indirectly related nodes are those with semantic similarity or temporal adjacency to directly related nodes, such as the first round of dialogue node which is temporally adjacent to the second round of dialogue node, thus the first round of dialogue node is an indirectly related node to the core memory node. Finally, the nodes obtained from forward retrieval, the nodes obtained from reverse tracing, and all directly and indirectly related nodes are summarized, and duplicate nodes are removed to form an extended set of related nodes.

[0053] Step 332: Based on the extended set of related nodes, analyze the distribution density and semantic relevance of the nodes, and define the local memory range according to the preset relevance threshold. First, analyze the distribution density of nodes in the extended set of related nodes. Distribution density refers to the concentration of nodes in the dialogue sequence. First, determine the time interval corresponding to the current dialogue, count the number of nodes in the extended set of related nodes that are within this time interval, and then count the total number of nodes in the extended set of related nodes. Divide the number of nodes within the time interval by the total number of nodes to obtain the distribution density value of the nodes. The higher the distribution density value, the more concentrated the nodes are in the time range related to the current dialogue, and the stronger the relevance to the current need. For example, if there are 5 nodes in the extended set of related nodes, and 4 of them are located in the time interval of the last 3 rounds of dialogue, the distribution density value is high. Next, analyze the semantic relevance of each node in the extended set of related nodes. Calculate the similarity between the key information of each node and the semantic features of the current dialogue context again. The calculation method is the same as in step 330, that is, the core element overlap score plus the behavioral instruction consistency score. The system calculates the semantic relevance score for each node by adding contextual relevance scores. A higher score indicates a stronger semantic relevance between the node and the current query. Then, a preset relevance threshold is set, comprising a distribution density threshold and a semantic relevance score threshold. This threshold is pre-defined based on the tightness of the demand correlation in the dialogue task. For tasks with strong demand progression, such as geometric proof exercises, the threshold is set to retain nodes that are sequentially relevant to the current demand and semantically closely related. Next, the node's distribution density value is compared to the distribution density threshold, and the semantic relevance score of each node is compared to the semantic relevance score threshold. Nodes with distribution density values ​​higher than the distribution density threshold and semantic relevance scores higher than the semantic relevance score threshold are retained. Finally, the area covered by these retained nodes is defined as the local memory range. This range focuses on historical demand information highly relevant to the current user query. For example, the defined local memory range includes nodes related to triangle congruence proof exercises from rounds 1 to 3 of the dialogue, excluding other graphical proof nodes mentioned earlier but irrelevant to the current demand.

[0054] By accurately locating the core memory nodes related to the current query, comprehensively tracing the historical nodes that are related before and after, and then delineating the local memory range by analyzing the node distribution and semantic association, it can not only focus on the core related information of the current need, but also fully cover the evolution link of user needs, generate a logically consistent response, and provide focused and comprehensive memory support, avoiding the problem of the intelligent agent relying solely on the current query keywords, resulting in a one-sided response or being out of touch with historical needs.

[0055] In a preferred embodiment of the present invention, step 4 above, which involves selecting key node sequences to construct a temporal evolution path within the local memory range, analyzing the changes in node states within the temporal evolution path, calculating a logical consistency metric, and generating a path correction value, may include:

[0056] In this embodiment of the invention, step 440 involves filtering nodes within the local memory scope based on their importance and relevance, selecting key nodes in chronological order, and constructing a temporal evolution path. Specifically, this includes: first, determining all nodes within the local memory scope. These nodes are historical dialogue nodes highly relevant to the current user's query. For example, in a scenario where a user practices geometric proofs, the local memory scope contains three nodes, corresponding to three rounds of dialogue. The first node involves the core elements of a triangle proof, behavioral instruction practice, and initial contextual requirements. The second node involves the core elements of a triangle proof, graphical congruence, behavioral instruction practice, and contextual supplementation of congruence limitations. The third node... The process begins with proving triangles with equal sides as the core element, practicing behavioral instructions, and supplementing the condition of equal sides with contextual connections. Next, the importance of each node is calculated using the same method as previous steps: frequency of occurrence plus connection strength. Frequency is the number of times the core element of the node appears in all dialogue rounds, and connection strength is the number of times the node has contextual connections with other nodes. Adding these two factors gives the importance score for each node. For example, the core element of the third node, equal sides, is consistently relevant in the dialogue, resulting in a higher frequency and connection strength, thus giving it a higher importance score than the first two nodes. Finally, the tightness of connection for each node is calculated. Based on the core memory nodes identified in step 330, the key information of each node is matched with the key information of the core memory nodes using a similarity score. The matching method is a combination of core element overlap score, behavioral instruction consistency score, and contextual relevance score, resulting in a correlation score between each node and the core memory node. A higher correlation score indicates a stronger connection between the node and the current core need. For example, among the three nodes mentioned above and the core memory node, the correlation score for the third node decreases sequentially, with the third node having the highest correlation score and the first node having the lowest. Subsequently, the importance score and correlation score of each node are comprehensively calculated and added together to obtain the node's overall selection score. A comprehensive screening threshold is set based on the number of nodes in the local memory range and the strength of the demand correlation. The comprehensive screening score of each node is compared with the threshold. Nodes with scores higher than the threshold are retained, and nodes with scores lower than the threshold are removed. If all nodes in the local memory range are highly relevant to the demand, they can all be retained. Finally, the retained nodes are arranged in the order of their corresponding dialogue creation time, with the earliest created node at the top and the most recently created node at the bottom, forming a complete chronological evolution path. For example, they can be arranged in the order of the first node, the second node, and the third node. This path fully presents the user's demand refinement process from practicing triangle proof problems to practicing proof problems of triangle congruence with equal sides.

[0057] Step 441 involves analyzing the state changes between adjacent nodes in the temporal evolution path, identifying three types of changes: information addition, content modification, and information deletion, to obtain the state change analysis results. Specifically, this includes: first, extracting two adjacent nodes from the temporal evolution path, and then performing state change analysis on each group of adjacent nodes sequentially. Taking the temporal evolution path as an example, first extract the first node and the second node as the first group of adjacent nodes, then extract the second node and the third node as the second group of adjacent nodes. Next, extract the key information for each group of adjacent nodes, including core elements, behavioral instructions, and contextual relationships. The process involves comparing the key information of the previous node with that of the next node, then identifying the type of new information. If the key information of the next node contains a core element, behavioral instruction, or contextual association not present in the previous node, and this new content does not conflict with the information of the previous node, it is considered new information. For example, in the first group of adjacent nodes, the second node has the core element of graphical equality, and the contextual association supplements the equality constraint without changing the original information; therefore, it is determined to be new information. In the second group of adjacent nodes, the third node has the core element of edge equality, and the contextual association supplements the edge constraint. Under the same conditions, it is also judged as information addition. Then, the content modification type is identified. If one or more items in the key information of the subsequent node are inconsistent with the corresponding content of the previous node, it means that the original information has been adjusted or changed, which is considered content modification. For example, if a user first requests to practice difficult triangle congruence proofs and then changes to practicing simpler triangle congruence proofs, the difficulty description in the context of the subsequent node changes and conflicts with the previous node, thus it is judged as content modification. In the requirement refinement scenario, no such modification occurred, so the identification result for this type is empty. Next, the information deletion type is identified. If the key information of the subsequent node... If the information is missing core elements, behavioral instructions, or contextual connections already included in the previous node, and this omission is not a natural selection due to information refinement, then it is considered information deletion. For example, if a user first requests a practice problem on the congruence of triangles and quadrilaterals, only mentioning the congruence of triangles and deleting the content related to quadrilaterals, this is considered information deletion. If the user's needs continue to be refined without deleting the original core information, then this type of identification result is also empty. Finally, by summarizing the state change types of each group of adjacent nodes, the state change analysis result of the entire temporal evolution path is obtained. The result is two consecutive information additions, with no content modifications or information deletions.

[0058] Step 442: Based on the state change analysis results, calculate the logical consistency measure by evaluating the contradictions and conflicts and the coherence of information during the change process, and generate the corresponding path correction value. Specifically, this includes: First, evaluating the contradictions and conflicts in the state changes of nodes in the temporal evolution path, checking the state changes of each pair of adjacent nodes one by one, determining whether the key information of the later node contradicts the key information of the previous node. If contradictions exist, it is recorded as a conflict. The total number of conflicts in the entire path is counted. Simultaneously, a penalty point is assigned based on the importance of the conflict content. The core principle is... Element-based conflict scores are higher than behavioral instruction conflict scores, and behavioral instruction conflict scores are higher than context-related conflict scores. The total number of conflicts and their corresponding deductions are summed to obtain the conflict score. Fewer conflicts and milder conflicts result in a higher conflict score; for example, no conflicts result in a perfect score. Next, the coherence of information is assessed by analyzing whether the state changes of each pair of adjacent nodes conform to the logic of demand evolution. If the addition of information in the later node is a reasonable refinement of the demand in the previous node, or the modification of content is a reasonable adjustment to the demand in the previous node, and there is no logical break, then it is considered coherent. (Statistical path...) The number of adjacent nodes in the path that are seamlessly connected is used to assign a continuity score based on the degree of continuity. The more completely connected groups, the higher the continuity score. For example, if two adjacent nodes are both new information and represent progressively refined requirements, the connection is completely seamless, and the continuity score is full. Then, a logical consistency metric is calculated by directly adding the conflict score to the information continuity score to obtain the logical consistency metric value. The higher this value, the smoother the logic of user requirement changes in the time-series evolution path, the less contradictory and the tighter the connection. If there is a conflict or a break in the connection, the value will decrease. For example, if both are full scores, the logical consistency metric value is the highest. Finally, corresponding path correction values ​​are generated. The path correction values ​​are inversely related to the logical consistency metric values. The higher the logical consistency metric value, the more accurate the current temporal evolution path is, requiring less correction, and the smaller the path correction value. The lower the logical consistency metric value, the more logical contradictions or connection problems exist in the path, requiring significant correction, and the larger the path correction value is. For example, the highest logical consistency metric value results in the lowest path correction value, requiring only minor calibration of the path. If there is a conflict in the core elements of a path, the logical consistency metric value is low, and the path correction value is large, which is used to adjust the node relationships in the memory network.

[0059] By selecting key nodes to construct a chronological evolution path of user needs, the system can accurately identify the types of changes in the state of needs, evaluate the logical consistency of the path, and generate correction values. This not only clearly restores the complete process of needs from initial proposal to gradual refinement, but also ensures the logical coherence of changes in needs, providing accurate path basis for generating responses that conform to the logic of historical dialogues.

[0060] In a preferred embodiment of the present invention, step 5 above, which integrates the path correction value with the structured memory network, adjusts the retrieval strategy and similarity calculation when a new round of user queries is received, and obtains the target memory node set from the structured memory network based on the uncertainty assessment results, may include:

[0061] In this embodiment of the invention, step 550 involves adjusting the relationships and influence distribution among nodes in the structured memory network based on the path correction value to form an updated network structure. Specifically, this includes: First, determining the adjustment magnitude corresponding to the path correction value. A larger path correction value indicates weaker logical consistency in the previously constructed temporal evolution path, requiring more significant adjustments to the structured memory network; a smaller path correction value indicates smoother temporal evolution path logic, resulting in a smaller adjustment magnitude. Next, adjusting the relationships between nodes, including semantic similarity and temporal adjacency relationships. For adjacent nodes with weak logical connections in the temporal evolution path, if the path correction value indicates a slight deviation, the temporal adjacency strength is appropriately enhanced. For example, between a triangle proof node and a triangle congruence proof node, the detailed description of their needs is supplemented to strengthen the closeness of the relationship. If the path correction value indicates a conflict between a group of nodes, such as a user previously mentioned... After practicing difficult proofs, the exercises are changed to simpler ones, which weakens the semantic similarity between conflicting nodes and strengthens the association between the latter node and other related nodes, avoiding interference from conflicting information. Then, the influence distribution of nodes is adjusted. The influence distribution reflects the weight of a node's influence on the search results in the network. For nodes with high contribution and key logical connections in the temporal evolution path, such as nodes in triangle congruence proofs, their influence weight is increased if the path correction value is small, making them easier to be prioritized in the search. For nodes marked due to logical conflicts, their influence weight is reduced to reduce their interference with the search results. Finally, the adjusted node associations and influence distributions are integrated into the original structured memory network to form an updated network structure. In this structure, the associations between nodes are more in line with the actual evolution logic of user needs, and the distribution of node influence is more in line with the core orientation of current needs. The association strength and influence of nodes related to triangles, congruence, and edge equality are all reasonably optimized.

[0062] Step 551: Based on the updated network structure and the uncertainty assessment results, determine the memory regions that need to be prioritized for retrieval, and adjust the parameter configuration of similarity calculation to enhance the fusion of temporal and semantic features. Specifically, this includes: First, judging the clarity of user needs based on the uncertainty assessment results. If the uncertainty assessment results show a high degree of uncertainty in user needs, it indicates that there are still ambiguities in the user's needs. The retrieval scope needs to be expanded in the updated network structure, including multiple memory regions that are semantically related to the current query and temporally similar. For example, if the user only mentions practicing proof problems but does not specify the type of shape, the retrieval should focus on memory regions related to triangles and quadrilaterals. If the uncertainty assessment results show a low degree of uncertainty in user needs, it indicates that the needs are relatively clear. Only the core memory regions directly related to the current needs in the updated network structure need to be focused on. For example, if the user explicitly states their need is for proofs of triangle congruence with equal sides, then only the retrieval of memory regions containing triangles, congruence, etc., needs to be prioritized. The memory region of core elements with equal edges is selected. Next, the parameter configuration for similarity calculation is adjusted. The original similarity calculation may have focused more on semantic feature matching. At this time, it is necessary to increase the proportion of temporal features in the similarity calculation and improve the integration of temporal feature scores and semantic feature scores. The parameter configuration is adjusted so that when calculating similarity, the temporal feature scores and semantic feature scores of the node and the current query are calculated separately. The temporal feature score is determined by the interval between the dialogue round corresponding to the node and the current dialogue round. The shorter the interval, the higher the score. The semantic feature score is obtained by adding the overlap of core elements, the consistency of behavioral instructions, and the similarity of contextual relevance. Then, the temporal feature scores and semantic feature scores are added according to the adjusted ratio to obtain the comprehensive similarity score. This strengthens the influence of temporal features on similarity calculation and ensures that the time evolution of needs and content relevance are taken into account during retrieval. Finally, the adjusted retrieval strategy is clarified as focusing on the memory region of retrieval and adopting a similarity calculation method that integrates temporal and semantic features to complete the adjustment of retrieval strategy and parameter configuration.

[0063] Step 552: Using the adjusted retrieval strategy and similarity calculation parameters, perform a retrieval operation in the updated structured memory network to obtain the retrieval results. Specifically, this includes: First, based on the adjusted retrieval strategy, locate the memory regions with a focus on retrieval in the updated structured memory network, such as core memory regions related to triangles, congruence, and equal sides. Nodes within this region are highly relevant to the user's current practice needs. Next, traverse all nodes within this focused retrieval region, extracting key information for each node, including core elements, behavioral instructions, dialogue rounds, and contextual associations. Then, according to the adjusted similarity calculation parameters, calculate the similarity between each node and the current user query. First, calculate the temporal feature score, then subtract the dialogue round corresponding to the node from the current dialogue round to obtain the round interval. The shorter the interval between rounds, the higher the temporal feature score. Next, the semantic feature score is calculated by adding the score for the number of overlaps between the core elements of the node and the core elements of the current query, the score for the consistency between the behavioral instructions and the behavioral instructions of the current query, and the score for the similarity between the contextual association and the contextual association of the current query. These three scores are then added together to obtain the semantic feature score. The temporal feature score and the semantic feature score are then added together according to an adjusted ratio to obtain the comprehensive similarity score between the node and the current query. Finally, all nodes are arranged in descending order of their comprehensive similarity scores to form the search results. Nodes with higher scores have a higher matching degree with the current user query. For example, the node containing the proof of triangle congruence with equal sides has the highest comprehensive similarity score and ranks first in the search results, followed by the node containing the proof of triangle congruence, and then the node containing the proof of triangle congruence.

[0064] Step 553: Based on the results of the retrieval operation, a set of target memory nodes that best match the current user's query is obtained. Specifically, this includes: First, setting a similarity filtering threshold. This threshold is set according to the clarity of the current user's needs. If the needs are clear and the uncertainty assessment result is low, a higher threshold is set, retaining only nodes with extremely high overall similarity scores. If the needs are still ambiguous and the uncertainty assessment result is high, the threshold is appropriately lowered, retaining more potentially relevant nodes. Taking a scenario with clear needs as an example, a higher filtering threshold is set, retaining only nodes with overall similarity scores higher than this threshold. Next, nodes with overall similarity scores higher than the threshold in the retrieval operation results are further filtered to check whether these nodes can completely cover the evolution chain of the user's needs, ensuring that the filtered nodes contain both the core information of the current needs and... Key nodes in the process of progressively refining requirements should be identified to avoid only retaining the latest nodes while omitting the constraints of previous requirements. In addition to retaining nodes for proofs of triangle congruence with equal sides, nodes for proofs of triangle congruence and proofs of triangles should also be retained to fully reflect the user's process from initial requirement to progressive refinement. Then, duplicate or semantically redundant nodes discovered during the screening process should be removed. If two nodes have completely identical core information but only slightly different descriptions, the node with more complete information should be retained, and the redundant node should be deleted to ensure the simplicity and effectiveness of the target memory node set. Finally, the nodes that have undergone threshold screening, link coverage checks, and redundancy removal should be summarized to form the target memory node set that best matches the current user query. The nodes in this set are not only highly similar to the current query but also fully reflect the evolution logic of the user's requirements.

[0065] By combining path correction values ​​and uncertainty assessment results, the memory network structure, retrieval strategy, and similarity calculation are dynamically adjusted. Then, the target memory node set is obtained through precise retrieval and filtering, ensuring that the retrieval process follows the logical evolution of user needs and that the output response is highly consistent with the user's true intent.

[0066] In a preferred embodiment of the present invention, step 6 above, which involves generating a task decision response by combining the target memory node set with the current user query input large language model, and updating the importance and correlation of relevant nodes in the structured memory network based on the interaction results to achieve continuous evolution of task decision-making capabilities, may include:

[0067] In this embodiment of the invention, step 660 involves combining the target memory node set with the current user query to form complete contextual information, which is then input into the large language model. The large language model then generates a corresponding task decision response through inference calculations. Specifically, this includes: First, decomposing each node in the target memory node set, extracting the core elements, behavioral instructions, and contextual associations contained in each node. Taking user needs as an example, the target memory node set contains three nodes. The first node is decomposed into the core element triangle proof problem, behavioral instruction practice, and contextual association initially presenting the need. The second node is decomposed into the core element triangle proof problem with congruent figures, behavioral instruction practice, and contextual association supplementing the congruence constraint. The third node is decomposed into the core element triangle proof problem with congruent figures and equal sides, behavioral instruction practice, and contextual association supplementing the equal side condition. Next, all the decomposed node information is arranged in chronological order according to the dialogue sequence to form a demand evolution information chain. Then, the current user query content, such as the condition that the question stem should ideally include equal sides, is appended to the end of the demand evolution information chain. The overall content encompasses both the complete process of user needs evolving from initial proposal to gradual refinement, as well as the latest query requests, forming a complete contextual information. This complete contextual information is then input into a large language model. The model's reasoning and computation process involves: first, semantic parsing of the input contextual information to identify that the core user need is to practice triangle congruence proofs with conditions for equal sides; then, analyzing the logical relationship of the evolving needs, clarifying that congruence is the core constraint of triangle proofs, and equal sides are supplementary conditions in the proof stem; next, based on the parsing results, matching problems that meet all conditions from a pre-set geometric proof resource library, and generating a problem recommendation explanation that explains the fit between the recommended problems and the user's needs; finally, through the above reasoning and computation, generating a task decision response. The response includes a triangle congruence proof problem stem that meets the conditions, as well as the recommendation reason, such as the recommended proof problem stem containing the condition for equal sides of triangles, and the core examination of geometric congruence proofs, perfectly matching the user's need to supplement congruence and equal sides conditions from practicing triangle proofs.

[0068] Step 661: Obtain user feedback on the task decision response. Based on the feedback, evaluate the accuracy and effectiveness of the target memory node set in the decision-making process to obtain the node utility evaluation result. Specifically, this includes: First, obtaining user feedback through dialogue interaction. The feedback includes the user's level of acceptance of the task decision response, modification requests, or supplementary explanations. For example, if the user explicitly states that the recommended questions meet expectations, this is acceptance feedback; if the user suggests that the recommended questions are too difficult and need to be adjusted, this is modification request feedback; if the user adds that they also want to practice congruence proofs containing angle equality conditions, this is supplementary explanation feedback. Next, based on the feedback, evaluate the accuracy and effectiveness of each node in the target memory node set one by one. The accuracy evaluation determines whether the key information contained in the node is consistent with the user's actual needs, and the effectiveness evaluation determines whether the node information contributes to generating a decision response that meets expectations. If the user accepts the recommendation result, it means that the third node contains edge equality information, the second... The congruence information contained in the nodes and the triangle proof information contained in the first node are accurate and effective, playing a key supporting role in the generation of decision response. If the user proposes a question that is too difficult, and the target memory node set does not contain difficulty-related nodes, it indicates that the existing nodes are missing in the difficulty dimension, but the existing triangle, congruence, and side equality nodes are still accurate and effective. Then, an evaluation score is assigned to the accuracy and effectiveness of each node. The accuracy score is determined based on the degree of fit between the node information and the actual needs in the user feedback, with complete fit receiving the highest score, partial fit receiving the second highest score, and non-fit receiving the lowest score. The effectiveness score is determined based on the degree of contribution of the node information to the decision response, with core contribution nodes receiving the highest score, auxiliary contribution nodes receiving the second highest score, and no contribution nodes receiving the lowest score. Finally, the accuracy score and effectiveness score of each node are added together to obtain the utility evaluation score of that node. The utility evaluation scores of all nodes are summarized to form the node utility evaluation result, clarifying the value of each node in the target memory node set.

[0069] Step 662: Based on the node utility evaluation results, recalculate the importance of relevant nodes in the structured memory network and adjust the strength of the associations between nodes. Specifically, this includes: First, recalculating the importance of relevant nodes. The recalculation is based on the original importance score, which is obtained by adding the node's frequency of occurrence to its association strength. The original importance score is then directly added to the utility evaluation score in the node utility evaluation results to obtain the recalculated importance of the node. Second, the original importance score of a node is obtained by adding the frequency of occurrence of equivalence information to its association strength with other nodes. If its utility evaluation score is high, the addition of the two significantly increases the node's importance. For missing nodes mentioned in user feedback, such as difficulty-related nodes, if the user clearly specifies the required difficulty level, the node is added and assigned a base value. Based on the importance score, the strength of the relationships between nodes is adjusted. This strength includes semantic similarity and temporal adjacency. For nodes with high utility evaluation scores, their relationships with other related nodes are strengthened. For example, if the third node has a high utility evaluation score for edge-equal information, its semantic similarity with the second node for equal information is strengthened, as is their temporal adjacency, reinforcing the logic of detailed needs. For nodes with low utility evaluation scores, such as those containing irrelevant difficulty information, their relationships with core nodes are weakened to reduce their interference with subsequent decisions. Finally, the recalculated importance and adjusted relationship strengths are synchronously updated in the structured memory network to ensure that the value ranking and relationship logic of nodes in the network are consistent with the user's actual needs feedback.

[0070] Step 663 involves adjusting the importance and correlation of nodes to adapt the structured memory network to the interaction process, enabling continuous evolution of task decision-making capabilities. Specifically, this includes: First, after adjusting the importance and correlation of nodes, the priority ranking and correlation logic of the internal nodes of the structured memory network better align with the latest user feedback. Nodes related to the equality of triangles have higher importance, and the correlation between nodes is tighter. The network structure accurately reflects the user's core needs for practicing geometric proofs. Then, when a new query is received from the user, the adjusted structured memory network can more quickly locate the core memory nodes related to the new query based on the updated node importance and correlation, reducing interference from irrelevant nodes. For example, if a user requests to practice similar equality proofs, the network can directly prioritize retrieving nodes related to the equality of triangles with high importance. The network eliminates the need to repeatedly filter out invalid information. Then, when generating new task decision responses, the network structure, having adapted to previous interaction feedback, more accurately matches the user's potential needs and preferences. For example, if a user previously indicated that a recommended question was of suitable difficulty, the network will prioritize questions at that difficulty level to avoid difficulty bias. Finally, as multiple rounds of dialogue continue, the structured memory network continuously adjusts the importance and relationships of nodes based on each user feedback, gradually optimizing its understanding and response to user needs. This allows task decision responses to evolve from initially meeting basic needs to meeting personalized and refined user needs, continuously improving task decision-making capabilities. For instance, it can evolve from initially recommending triangle proofs to accurately recommending triangle congruence proofs that match the user's difficulty preferences and include specific conditions.

[0071] By generating accurate responses from complete demand context inputs and optimizing the importance and relationships of memory network nodes based on user feedback, we can ensure that each task decision response highly matches the user's real needs. At the same time, the structured memory network can continuously adapt to changes in user needs as the interaction process progresses, thereby continuously improving the understanding and response accuracy of demands and achieving a steady evolution of the task decision-making capabilities of large-scale intelligent agents.

[0072] like Figure 2 As shown, embodiments of the present invention also provide a large-scale model intelligent agent interactive question-answering task decision system, including:

[0073] The information extraction module is used to receive user input through multi-turn dialogue, extract key information to form an initial memory node set, and analyze the ambiguity of the query intent and the completeness of the information to generate uncertainty assessment results.

[0074] The memory construction module is used to evaluate the timeliness and importance of each node based on the uncertainty assessment results, generate a set of memory nodes with corresponding metadata identifiers, and construct a dynamic memory structure according to semantic and temporal relationships to form a structured memory network.

[0075] The context association module is used to identify core memory nodes related to the current dialogue context based on a structured memory network, and to trace associated nodes in the forward and backward directions of the dialogue sequence to define the local memory range.

[0076] The path verification module is used to select key node sequences to construct a time-series evolution path within the local memory range, analyze the changes in the node states in the time-series evolution path, calculate the logical consistency metric, and generate path correction values.

[0077] The retrieval optimization module integrates path correction values ​​with the structured memory network. When a new round of user queries is received, the retrieval strategy and similarity calculation are adjusted, and the target memory node set is obtained from the structured memory network by combining the uncertainty assessment results.

[0078] The response update module is used to combine the target memory node set with the current user query input large language model to generate a task decision response. Based on the interaction results, it updates the importance and correlation of relevant nodes in the structured memory network to achieve continuous evolution of task decision-making capabilities.

[0079] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0080] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0081] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0082] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for decision-making in interactive question-answering tasks for large model agents, characterized by, The method comprises: Step 1, receiving user input through multiple rounds of dialogue, extracting key information to form an initial memory node set, and analyzing the ambiguity of the query intention and the information integrity to generate an uncertainty evaluation result; Step 2, based on the uncertainty evaluation result, evaluating the timeliness and importance degree of each node to generate a memory node set with corresponding metadata identifiers, and constructing a dynamic memory structure according to the semantic and time sequence relationship to form a structured memory network; including: based on the uncertainty evaluation result, screening the initial memory node set, calculating the timeliness score according to the creation time and last access time of each node, and combining the appearance frequency and correlation strength of the node in the dialogue to calculate the importance score; the timeliness score and the importance score are comprehensively calculated to obtain the comprehensive evaluation value of each node, and the initial memory node is screened and adjusted according to the comprehensive evaluation value to generate a memory node set with corresponding metadata identifiers; the nodes in the memory node set with corresponding metadata identifiers are analyzed to establish the semantic similarity relationship and the dialogue time sequence adjacent relationship between the nodes; based on the semantic similarity relationship and the dialogue time sequence adjacent relationship, the dynamic memory structure is constructed to form a structured memory network; Step 3, based on the structured memory network, identifying the core memory nodes related to the current dialogue context, and tracing the associated nodes along the dialogue time sequence in the forward and reverse directions to define the local memory range; Step 4, in the local memory range, selecting key node sequences to construct a time sequence evolution path, and analyzing the transition of node state in the time sequence evolution path to calculate the logical consistency measure and generate the path correction value; including: in the local memory range, according to the importance degree and correlation degree of the nodes, the key nodes are selected in time sequence to construct the time sequence evolution path; the state changes between adjacent nodes in the time sequence evolution path are analyzed to identify three change types of information addition, content modification and information deletion of node content to obtain the state change analysis result; based on the state change analysis result, the logical consistency measure is calculated by evaluating the contradiction conflict situation and the information connection coherence degree in the change process to generate the corresponding path correction value; Step 5, integrating the path correction value with the structured memory network, when receiving a new round of user query, adjusting the retrieval strategy and similarity calculation, and combining the uncertainty evaluation result to obtain the target memory node set from the structured memory network; Step 6, inputting the target memory node set and the current user query into the large language model to generate a task decision response, updating the importance degree and correlation relationship of the related nodes in the structured memory network according to the interaction result, and realizing the continuous evolution of the task decision ability.

2. The method of claim 1, wherein, Through multiple rounds of dialogue, user input is received, key information is extracted to form an initial memory node set, and the ambiguity of the query intention and the information integrity are analyzed to generate an uncertainty evaluation result, including: Through continuous multiple rounds of dialogue, user input is received, core elements, behavior instructions and context association are identified as key information, and the key information is integrated into an initial memory node set; Based on the initial memory node set, the user query intention is parsed, the coverage between the identified keywords and the preset task domain is calculated, and the context correlation between different memory nodes is combined to evaluate the fuzziness degree of the intention; Based on the fuzziness degree of the intention, the current information completeness is analyzed, the proportion of missing necessary information is identified by comparing the existing key information with the information items required by the preset task, and the information completeness is quantified according to the proportion; The fuzziness degree of the intention and the information completeness are integrated to generate a quantitative uncertainty evaluation result.

3. The method of claim 2, wherein, Based on the structured memory network, the core memory nodes related to the current dialogue context are identified, and the associated nodes are traced forward and backward along the dialogue time sequence to define the local memory range, including: Based on the semantic features of the current dialogue context, similarity matching is performed in the structured memory network to identify the core memory nodes most relevant to the current query semantics; Based on the core memory nodes, the nodes generated after the forward search along the dialogue time sequence are retrieved, and the nodes generated before the backward search are traced, all directly and indirectly associated nodes are extracted to form an extended associated node set; Based on the extended associated node set, the distribution density and semantic relevance of the nodes are analyzed, and the local memory range is defined according to the preset relevance threshold.

4. The method of claim 3, wherein, Integrate the path correction value with the structured memory network, when receiving a new round of user query, adjust the retrieval strategy and similarity calculation, and combine the uncertainty evaluation result to obtain the target memory node set from the structured memory network, including: According to the path correction value, adjust the association relationship and influence distribution between nodes in the structured memory network to form an updated network structure; Based on the updated network structure, combined with the uncertainty evaluation result, determine the memory area that needs to be focused on retrieval, and adjust the parameter configuration of similarity calculation to enhance the fusion degree of time sequence features and semantic features; Use the adjusted retrieval strategy and similarity calculation parameters to perform retrieval operations in the updated structured memory network to obtain the results of the retrieval operations; Through the results of the retrieval operations, filter the target memory node set that best matches the current user query.

5. The method of claim 4, wherein, Input the target memory node set and the current user query into the large language model to generate a task decision response, update the importance degree and association relationship of related nodes in the structured memory network according to the interaction result, and realize the continuous evolution of the task decision capability, including: Combine the target memory node set and the current user query to form complete context information input into the large language model, and generate a corresponding task decision response through the inference calculation of the large language model; Obtain the feedback information of the user to the task decision response, evaluate the accuracy and effectiveness of the target memory node set in the decision-making process based on the feedback information, and obtain the node utility evaluation result; According to the node utility evaluation result, recalculate the importance degree of the related nodes in the structured memory network, and adjust the association relationship between the nodes; Through the adjustment of the importance degree and the association relationship, the structured memory network adapts to the interaction process, and realizes the continuous evolution of the task decision capability.

6. A large model agent interactive question and answer task decision system, the system implements the method as claimed in any one of claims 1 to 5, characterized in that, The method comprises the following steps: An information extraction module is configured to receive user input through multi-round dialogue, extract key information to form an initial memory node set, analyze the ambiguity of query intent and the completeness of information, and generate an uncertainty evaluation result; A memory construction module is configured to evaluate the timeliness and importance of each node based on the uncertainty evaluation result, generate a memory node set with corresponding metadata identifiers, and construct a dynamic memory structure according to semantic and chronological relationships to form a structured memory network; An upper and lower correlation module is configured to identify core memory nodes related to the current dialogue context based on the structured memory network, and trace associated nodes forward and backward along the dialogue time sequence to define a local memory range; A path verification module is configured to select key node sequences to construct a time sequence evolution path within the local memory range, analyze the transition of node states in the time sequence evolution path, calculate a logical consistency measure, and generate a path correction value; A retrieval optimization module is configured to integrate the path correction value with the structured memory network, adjust the retrieval strategy and similarity calculation when receiving a new round of user query, and obtain a target memory node set from the structured memory network in combination with the uncertainty evaluation result; A response updating module is configured to input the target memory node set and the current user query into a large language model to generate a task decision response, update the importance and correlation of related nodes in the structured memory network according to the interaction result, and realize continuous evolution of the task decision capability.

7. A computing device, comprising: One or more processors; A storage device is configured to store one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method as claimed in any one of claims 1 to 5. The computer readable storage medium stores a program which is executed by the processor to implement the method as claimed in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, ​

Images