Complex problem solving method based on dynamic theme graph and adaptive re-planning

Abstract

The application belongs to the technical field of artificial intelligence, and relates to a complex problem solving method based on a dynamic topic graph and adaptive re-planning, which comprises the following steps: S1, extracting a semantic element set from a standardized problem and constructing an initial dynamic topic graph; S2, retrieving a candidate entity set from an external knowledge source, performing constraint propagation and candidate contraction on the candidate entity set; S3, performing retrieval based on the current candidate entity set and the corresponding current dynamic topic graph, updating the current dynamic topic graph, and performing reasoning through a large language model to obtain a candidate answer; and S4, performing consistency verification on the candidate answer, and when the candidate answer fails to pass the consistency verification, the current candidate entity set fails to be effectively contracted, the comprehensive evidence set is insufficient, or the updated current dynamic topic graph has conflicts, performing adaptive re-planning and re-searching and reasoning. The method effectively improves the stability, controllability and overall solving performance of the complex problem solving process.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology and relates to a method for solving complex problems, particularly a method for solving complex problems based on dynamic topic graphs and adaptive replanning. Background Technology

[0002] With the rapid development of artificial intelligence, large language modeling, and knowledge-enhanced reasoning technologies, complex problem solving based on large language models has been widely applied in fields such as complex knowledge question answering, intelligent search, scientific research auxiliary analysis, enterprise knowledge management, legal text analysis, technical solution derivation, task planning, and decision support. Complex problem solving typically requires models to output accurate answers, structured conclusions, or reasoning paths to the target problem based on a problem description inputted in natural language, combined with external knowledge sources, historical context, or multi-step intermediate reasoning results. In practical applications, to improve the model's ability to model the structure of complex problems and multi-step reasoning processes, existing technologies typically introduce chain reasoning mechanisms, problem decomposition mechanisms, retrieval enhancement generation mechanisms, graph structure representation mechanisms, or multi-round reflection and verification mechanisms to further model and optimize complex problems.

[0003] In ideal complex problem-solving scenarios, a model should possess strong natural language understanding capabilities while maintaining precise representation of the problem's subject matter, constraints, target attributes, and intermediate dependencies. It should also effectively model the joint constraints between different conditions, the dependency propagation between different steps, and the dynamic correction of the solution path by external knowledge evidence. However, in current complex problem-solving practices, model optimization remains relatively crude, lacking refined modeling of the interactions between shared subject matter, joint constraints, dynamic intermediate state maintenance, and inference path backtracking correction. This leads to problems in complex problem scenarios such as inaccurate subject matter localization, insufficient constraint utilization, unstable intermediate entity binding, disconnect between retrieval and inference processes, and the gradual accumulation and amplification of erroneous steps.

[0004] Especially given the ever-expanding scale and increasingly complex scenarios of problem-solving tasks, models not only need to understand the natural language problem itself, but also accurately identify the core subject matter, multiple constraints, and relationships between target attributes. Furthermore, they must integrate entities, relationships, attributes, and evidence paths from external knowledge sources to progressively solve the problem. For such tasks, the subject matter, constraints, external evidence, and target attributes have different emphases and are easily fragmented during the solution process: if the solution process overemphasizes the linear generation capability of natural language, the model may overlook explicit constraints and implicit intermediate variables, leading to deviations in the reasoning path; if the solution process relies too heavily on a fixed decomposition structure, the model's responsiveness to subsequent new evidence, conflicting information, and path corrections may be insufficient; if the solution process only handles the relationships between the subject matter and other entities through implicit attention, it is difficult to fully express the structural connections of "multiple conditions jointly constraining the same object" or "multiple intermediate steps sharing the same subject." Therefore, balancing the ability to locate the subject object, the ability to jointly model multiple constraints, the ability to utilize external knowledge, and the ability to dynamically correct the reasoning path during the solution of complex problems has become a key technical issue in the optimization of solving complex problems in large language models. Existing literature also points out that the core difficulty in solving complex problems lies in simultaneously capturing the synergistic relationship between problem decomposition, knowledge retrieval, multi-step reasoning, and dynamic correction, while existing methods often suffer from shortcomings such as fixed structural representation, insufficient utilization of shared variables, limited path adjustment capabilities, and insufficient state maintenance.

[0005] In recent years, research and engineering practices related to problem structure modeling, knowledge utilization, and dynamic reasoning in solving complex problems can be mainly summarized into the following technical paths.

[0006] 1. A solution for complex problems based on single-chain reasoning.

[0007] These approaches typically rely directly on the generative capabilities of large language models on natural language sequences. They generate a series of intermediate reasoning steps based on the original question text through chained reasoning hints, step-by-step explanations, or multi-turn text expansion, ultimately outputting the answer. A representative idea behind this approach is to capture the intermediate logical relationships and multi-step reasoning processes within the problem by progressively unfolding the problem-solving path in the natural language space.

[0008] While such approaches can effectively utilize the language reasoning capabilities of large language models and achieve good results in some simple multi-step question-answering tasks, they essentially still rely primarily on linear text generation. Their ability to express structural relationships in complex problems, such as shared subject objects, simultaneous effects of multiple constraints, and cross-dependencies of multiple intermediate variables, is relatively limited. When a problem involves multiple joint constraints, reuse of multiple intermediate entities, and target attributes depending on multiple evidence paths, a single-chain reasoning approach often fails to fully characterize the complex internal structure of the problem, resulting in insufficient learning ability for subject binding, reuse of intermediate variables, and global path control. Related research also points out that a single linear reasoning chain is insufficient to fully express the shared variables and cross-dependency structures in solving complex problems.

[0009] 2. A solution scheme for complex problems based on static problem decomposition.

[0010] To enhance the model's ability to handle complex problems, some existing solutions introduce problem decomposition mechanisms. This involves breaking down the original problem into several sub-problems, sub-tasks, or intermediate reasoning units, and then solving them step-by-step based on the decomposition results. The core idea of ​​these solutions is to explicitly decompose the logical structure of complex problems, transforming complex tasks into multiple simpler solution units, thereby improving the model's ability to handle complex problems.

[0011] Compared to single-chain reasoning, this type of approach can enhance the model's ability to express multi-step reasoning and hierarchical logical structures to some extent. However, its problem decomposition is usually based on a fixed tree structure or static hierarchical structure, making it difficult to simultaneously consider shared subject objects, cross-branch constraint propagation, and subsequent evidence-driven structural adjustments. Especially when multiple constraints act on the same subject object in the problem, or when subsequent evidence retrieval can negatively affect the results of earlier decompositions, static problem decomposition methods are often insufficient to fully reflect the dynamic dependencies of complex problems at different solution stages. Existing literature also points out that existing problem decomposition methods typically use a fixed structure and lack the ability to dynamically adjust the structure based on the solution status and evidence feedback.

[0012] 3. A solution for complex problems based on enhanced single knowledge retrieval.

[0013] To further improve the performance of solving complex problems, some existing solutions have begun to introduce retrieval-enhanced generation mechanisms. These mechanisms retrieve relevant external knowledge by means of vector retrieval, document retrieval, knowledge graph retrieval, or external tool calls. The main idea behind these solutions is to obtain relevant factual evidence, entity attributes, or intermediate relationship paths from external knowledge sources before the large language model generates the answer or during the reasoning process, thereby enhancing the model's ability to utilize real-world knowledge.

[0014] Compared to solutions that rely solely on model parameter knowledge, this type of approach can improve the model's ability to utilize external knowledge to some extent. However, its focus is usually on the retrieval process itself, while its contribution to topic object localization, joint propagation of multiple constraints, and continuous maintenance of intermediate states is relatively limited. For problem scenarios with both complex constraint structures and significantly shared topic objects, relying solely on single-round or coarse-grained retrieval enhancements is still insufficient to comprehensively address the issues of topic binding, path stability, and dynamic correction in solving complex problems.

[0015] 4. Solution schemes for complex problems based on static graph representation or graph reasoning.

[0016] To alleviate the limitations of single-chain reasoning or single-retrieval enhancements, a more advanced approach has been proposed: graph representation and graph reasoning. This involves representing problems, entities, relationships, or intermediate steps as graph structures within the same solution framework, and enhancing the model's ability to model structural information through graph retrieval, graph traversal, graph-constrained reasoning, or graph neural network propagation mechanisms. This approach can, to some extent, achieve the synergistic utilization of both the "problem's logical structure" and the "external knowledge structure," thus achieving better performance in complex knowledge-answering and multi-hop reasoning tasks compared to single-chain or single-retrieval models.

[0017] This type of solution is also one of the existing implementation paths that is relatively close to this invention. Its similar implementation can generally be summarized as follows: constructing a hierarchical or graphical intermediate representation on the input problem, structurally encoding several sub-problems, intermediate entities, or knowledge nodes in the problem; then performing local retrieval, path selection, step-by-step reasoning, or multi-round verification based on this structure to finally generate the answer. Some existing solutions further introduce verification mechanisms, reflection mechanisms, or multi-model collaboration mechanisms on the basis of problem decomposition and graphical representation, jointly utilizing the problem structure and external knowledge to improve the solving effect of complex problems. Its core idea is to simultaneously utilize the problem's logical structure, external knowledge evidence, and multi-step intermediate results, and adjust the solving process through intermediate verification or path selection mechanisms.

[0018] Although this type of scheme is superior to single chain reasoning or single knowledge retrieval enhancement schemes, it still has two obvious limitations: (1) The problem structure representation usually tends to be static hierarchical expansion or fixed intermediate graph representation, making it difficult to explicitly construct a unified topic graph that can be dynamically updated around the shared topic object. It is not adaptable to situations such as multiple constraints jointly limiting the same object and multiple intermediate nodes reusing the same topic object; (2) The solution control granularity usually tends to be adjusted at the overall problem level or overall path level. In particular, although the existing intermediate verification mechanism can judge whether the current path is reasonable as a whole, it is still difficult to perform more fine-grained dynamic updates for different topic candidates, different constraint nodes and different evidence edges in the same problem. Therefore, it is difficult to fully adapt to the solution requirements of multiple constraints, multiple candidates and multiple evidence states coexisting in complex problem scenarios.

[0019] 5. The existing implementation that is most similar to the present invention.

[0020] From the perspective of existing technical approaches, the most similar implementation to this invention is a complex problem-solving scheme based on problem structure representation, external knowledge retrieval, and intermediate verification control. Its representative implementation can be summarized as follows: First, a structured problem representation is constructed for the input complex problem, for example, by breaking the problem down into several sub-problems or intermediate reasoning units, forming a tree-like or graph-like intermediate solution structure. Second, knowledge retrieval is performed around this structure to obtain local and global evidence. Third, a large language model is used to progressively reason about the retrieval results and intermediate structure, and the solution results are verified through a validation mechanism at intermediate steps or the final answer. Finally, the answer to the problem is output. Compared to pure chain-based reasoning schemes, this type of scheme already possesses a certain degree of structured expression capability and retrieval enhancement capability, and demonstrates better solution performance in complex problem solving.

[0021] The representative implementation path of this type of closest solution usually has the following characteristics: (1) The input problem is logically decomposed from top to bottom to form a tree-like or hierarchical problem representation; (2) Local retrieval and global retrieval are performed based on the intermediate problems and key entities obtained from the decomposition to obtain external knowledge that supports the solution; (3) The retrieved evidence is combined with the intermediate problem representation to complete the intermediate node solution and the final answer generation in sequence; (4) The intermediate answer or the final result is checked through the verification model or consistency verification module, and local correction is performed when necessary.

[0022] However, the closest solutions in this category are still mainly complex problem-solving methods based on static problem structure and overall reasoning control, and still have the following shortcomings: (1) The problem structure is usually based on a hierarchical structure or fixed graph structure built at one time, lacking a mechanism for continuous updating and explicit unified modeling around shared subject objects, making it difficult to adapt to the state changes of subject objects at different solution stages; (2) Constraints are usually indirectly reflected in the form of sub-problems or intermediate steps, lacking an explicit constraint propagation and gradual shrinking mechanism for subject candidate sets, making it difficult to adapt to the solution requirements of joint constraint limitation and gradual selection of candidate subjects; (3) Although external knowledge can be retrieved and used for intermediate reasoning, there is a lack of a mechanism to continuously write new evidence back to the unified subject graph and maintain intermediate states, so the system's ability to respond to dynamic state changes is limited in the scenarios of new evidence, conflicting evidence, and local path correction; (4) Intermediate verification or final verification focuses more on the correctness of the overall steps or overall answer, while the unified adaptive replanning mechanism between node states, relation edge states, candidate entity states, and path states in the subject graph is still insufficient. 6. Existing solutions lack common technologies.

[0023] In summary, while existing solutions for complex problems have improved the model’s ability to utilize problem structure and external knowledge to some extent, they still generally suffer from the following key technical deficiencies: (1) lack of a unified graphical representation mechanism that can simultaneously take into account shared topic objects, joint constraints, and target attribute solutions; (2) difficulty in providing fine-grained constraint propagation and dynamic contraction capabilities around the topic candidate set for the filtering effect of different constraints on different topic candidates; (3) lack of a mechanism for continuous maintenance and dynamic updating of topic state, constraint state, evidence state, and target state during the solution process, making it difficult to make full use of new evidence obtained from subsequent retrieval; (4) difficulty in adaptively and collaboratively utilizing topic graph structure, external knowledge evidence, and large language model reasoning according to the specific solution state; (5) existing verification and control mechanisms are mostly concentrated on coarse-grained adjustment at the overall problem level, which is insufficient for adapting to scenarios with multiple candidates, multiple constraints, and multiple pieces of evidence in complex problem solving; (6) difficulty in providing unified and refined structural support for topic location, constraint propagation, evidence completion, and path correction in complex problem solving. Based on the above analysis, there is an urgent need for a new solution to complex problems. Summary of the Invention

[0024] To address the problems existing in the prior art, this invention proposes a method for solving complex problems based on dynamic topic graphs and adaptive replanning. This method can simultaneously optimize topic object localization, constraint propagation, external evidence utilization, and reasoning path correction during the complex problem-solving process. By introducing a dynamic topic graph modeling mechanism, a constraint propagation and candidate shrinking mechanism, and an adaptive replanning mechanism, it adaptively coordinates the relationships between topic states, constraint states, evidence states, and target states based on the topic objects in the input problem, different constraints, and subsequent external knowledge feedback. This provides a reliable technical foundation for stable optimization and comprehensive performance improvement in complex problem solving.

[0025] To achieve the above objectives, the present invention provides the following technical solution: A method for solving complex problems based on dynamic topic graphs and adaptive replanning, characterized by the following steps: S1: Extract a set of semantic elements from the standardized problem and construct an initial dynamic topic graph based on the set of semantic elements. The node set of the initial dynamic topic graph includes topic nodes, constraint node set and target node. S2: Based on the topic node, retrieve the candidate entity set from the external knowledge source, and perform constraint propagation and candidate shrinking on the candidate entity set based on the constraint node set to obtain the current candidate entity set; S3: Based on the current candidate entity set and its corresponding current dynamic topic graph, a comprehensive evidence set is obtained through retrieval. The current dynamic topic graph is updated according to the comprehensive evidence set. Based on the standard question, the updated current dynamic topic graph, the current candidate entity set, and the comprehensive evidence set, reasoning is performed through a large language model to obtain candidate answers. S4: Perform consistency check on the candidate answer. If the candidate answer fails the consistency check, the current candidate entity set fails to shrink effectively, the comprehensive evidence set is insufficient, or there is a conflict in the updated current dynamic topic graph, perform adaptive replanning to obtain the adaptively replanned dynamic topic graph. Then, return to step S2 for re-retrieval and reasoning based on the adaptively replanned dynamic topic graph.

[0026] Preferably, step S2 specifically includes: S21: Around the topic node From external knowledge sources The search yields a set of candidate entities. And calculate the initial candidate score for each candidate entity. :

[0027] In the formula, Represents the set of candidate entities; Represents the first in the candidate entity set One candidate entity; This represents the total number of candidate entities in the candidate entity set; S22: Apply the set of constraint nodes sequentially to the set of candidate entities to update the initial candidate score of each candidate entity, thereby obtaining the updated candidate score. ; S23: Based on the updated candidate scores The candidate entity set is filtered to obtain the current candidate entity set:

[0028] In the formula, Indicates the use of the first The constraint conditions corresponding to each constraint node The updated set of candidate entities; Indicates the use of the first The constraint conditions corresponding to each constraint node The updated set of candidate entities; Indicates the use of the first The constraint conditions corresponding to each constraint node Updated filtering thresholds.

[0029] Preferably, in step S21, an initial candidate score is calculated for each candidate entity. Specifically:

[0030] In the formula, Indicates the first Candidate entities The initial candidate scores; This represents the balance coefficient between similarity and prior score; Indicates the first Candidate entities With topic nodes Semantic similarity between them; Indicates the first Candidate entities The prior credibility.

[0031] Preferably, step S22 specifically includes:

[0032] In the formula, Indicates the first Candidate entities Using the The constraint conditions corresponding to each constraint node Updated candidate scores; Indicates the first Candidate entities Using the The constraint conditions corresponding to each constraint node Updated candidate scores; Indicates the first The constraint conditions corresponding to each constraint node; Indicates the first Candidate entities With the The constraint conditions corresponding to each constraint node Relationship matching score; Indicates the first Candidate entities With the The constraint conditions corresponding to each constraint node Time consistency score; Indicates the first Candidate entities Violation of Article The constraint conditions corresponding to each constraint node Penalties; This represents the corresponding weight parameter. Preferably, step S3 specifically includes: S31: Based on the current candidate entity set and its corresponding current dynamic theme image Perform a local search:

[0033] In the formula, This represents the set of evidence obtained from a local search; Indicates based on the first Candidate entities Current dynamic theme image The set of constraint nodes in and target node From external knowledge sources Searching for the first Candidate entities Local retrieval functions for directly relevant attribute facts, relational facts, or evidence of target values; S32: Based on the current dynamic topic graph Current candidate entity set and external knowledge sources Perform a global search:

[0034] In the formula, This represents the set of evidence obtained from a global search. This indicates that the current dynamic topic graph is based on... From external knowledge sources Searching for the current candidate entity set A global retrieval function for relevant evidence; S33: Based on the results of the local and global searches, a comprehensive evidence set is formed. :

[0035] S34: Based on the aforementioned comprehensive set of evidence Update the current dynamic theme graph. :

[0036] In the formula, Represents the graph update function; This represents the updated current dynamic theme image; S35: Based on the aforementioned standard issue Updated current dynamic theme image Current candidate entity set and comprehensive evidence set Reasoning is performed using a large language model to obtain candidate answers. :

[0037] Preferably, the consistency check performed on the candidate answer in step S4 specifically includes: S41: Regarding the candidate answers Calculate the consistency score:

[0038] In the formula, Indicates candidate answers Consistency score; Indicates the selection of answers With the updated current dynamic theme image Consistency score; Indicates candidate answers With comprehensive evidence set Consistency score; Indicates candidate answers With the aforementioned standard problem The consistency score of the constraint node set; This represents the weight parameter for the corresponding item; S42: Based on the aforementioned consistency score and consistency threshold Determine whether the candidate answer passes the consistency check:

[0039] In the formula, This indicates that the consistency check has been passed; This indicates that the consistency check failed.

[0040] Preferably, step S4, which involves performing adaptive replanning to obtain a replanned dynamic topic graph when the candidate answer fails the consistency check, the current candidate entity set fails to shrink effectively, the comprehensive evidence set is insufficient, or the updated current dynamic topic graph contains conflicts, specifically includes: S43: Calculate replanning trigger indicators :

[0041] In the formula, Indicates the replanning trigger indicator; Indicates an indicator function; Indicates the maximum number of candidate entities allowed; This indicates the lower limit of the number of valid pieces of evidence in the comprehensive evidence set; This indicates the number of conflicts in the updated current dynamic theme graph; S44: When satisfied At that time, perform adaptive replanning to obtain the dynamic topic graph after adaptive replanning:

[0042] In the formula, This represents the dynamic topic graph after adaptive replanning. This indicates a problem based on the stated standard. Comprehensive evidence collection and candidate answers Updated current dynamic theme image The function that performs replanning.

[0043] Furthermore, this invention also provides a system for solving complex problems based on dynamic topic graphs and adaptive replanning, characterized in that it includes: The problem analysis and initial dynamic topic graph construction module is used to extract a set of semantic elements from a standardized problem and construct an initial dynamic topic graph based on the set of semantic elements. The node set of the initial dynamic topic graph includes topic nodes, constraint node set and target nodes. The candidate entity recall, constraint propagation and candidate shrinking module is used to retrieve a set of candidate entities from external knowledge sources based on the topic node, and to perform constraint propagation and candidate shrinking on the set of candidate entities based on the set of constraint nodes to obtain the current set of candidate entities. The evidence retrieval and dynamic topic graph update and large language model reasoning module is used to retrieve a comprehensive evidence set based on the current candidate entity set and its corresponding current dynamic topic graph, update the current dynamic topic graph according to the comprehensive evidence set, and reason through a large language model based on the standard question, the updated current dynamic topic graph, the current candidate entity set and the comprehensive evidence set to obtain candidate answers. The consistency verification and adaptive replanning module is used to perform consistency verification on the candidate answer and, when the candidate answer fails the consistency verification, the current candidate entity set fails to shrink effectively, the comprehensive evidence set is insufficient, or there is a conflict in the updated current dynamic topic graph, perform adaptive replanning to obtain the adaptively replanned dynamic topic graph, and perform re-retrieval and reasoning based on the adaptively replanned dynamic topic graph.

[0044] Furthermore, the present invention also provides a complex problem-solving device based on dynamic topic graphs and adaptive replanning, characterized in that it includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the complex problem-solving method based on dynamic topic graphs and adaptive replanning as described above. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements the steps of the complex problem-solving method based on dynamic topic graphs and adaptive replanning as described above.

[0045] Compared with existing methods for solving complex problems that mainly rely on single-chain reasoning, static problem decomposition, single-round retrieval enhancement, or overall result verification, the complex problem-solving method based on dynamic topic graphs and adaptive replanning proposed in this invention has the following significant advantages in terms of shared topic object representation ability, multi-constraint joint modeling ability, knowledge utilization integrity, dynamic correction ability of reasoning paths, and stability of complex problem solving: (1) Based on the unified modeling mechanism of dynamic topic graph, the ability to express complex problem structures is fundamentally improved.

[0046] While some existing methods for solving complex problems incorporate problem decomposition, hierarchical structure representation, or graphical intermediate representation to enhance the model's ability to model complex logical relationships and multi-step reasoning processes, their problem representation methods typically rely on linear reasoning chains, tree-like decomposition structures, or fixed intermediate graph structures. These methods primarily emphasize the unfolding of problem-solving steps and struggle to simultaneously address shared subject objects, multiple constraints, and the unified expression of multiple intermediate states. When the input problem involves multiple conditions acting on the same subject object, multiple intermediate variables revolving around the same target object, or multiple evidence paths jointly determining the target attribute, fixed hierarchical or linear structures often fail to fully reflect the complex dependencies within the problem. Existing closer solutions also typically focus on statically structured representation of the problem, combined with retrieval and reasoning to execute the solution. The emphasis is on structure guiding the reasoning process, while the ability to organize the problem into a unified graph around the same subject object remains limited.

[0047] This invention constructs a dynamic topic graph centered on the subject object, unifying constraints, target attributes, and external evidence into a single graph structure. This allows the system to not only express the overall structure of the problem during the solution process but also continuously maintain the relationships between subject nodes and various constraint and evidence nodes, thereby effectively enhancing its ability to express the internal dependency structures of complex problems. Through this approach, the system's overall ability to express shared subject objects, multiple conditional constraints, and target attribute solution relationships is significantly improved, thereby enhancing the accuracy, interpretability, and structural integrity of complex problem-solving.

[0048] (2) Based on the dynamic shrinkage mechanism of topic candidate and constraint propagation, the pertinence and precision of topic positioning process are significantly improved.

[0049] In existing technologies, the localization of subject objects during the solution of complex problems typically relies on implicit semantic matching, intermediate inference after static decomposition, or posterior filtering of the overall retrieval results. Even if some methods introduce intermediate entity recognition or step-level constraint utilization, they mostly perform coarse-grained filtering only at the overall problem level or intermediate step level, making it difficult to reflect the true filtering effect of different constraints on subject candidate objects within the same problem. Existing problem decomposition or retrieval enhancement schemes are mainly used to improve the overall solution capability of the model, but their ability to progressively propagate multiple constraints around the same subject candidate set and dynamically adjust candidate priorities remains limited.

[0050] This invention introduces a topic candidate entity recall and constraint propagation mechanism based on the topic graph. It can assess the matching degree of each candidate topic object in the input problem to different constraints, and progressively update the credibility and ranking of the candidate objects accordingly. Through this approach, the topic localization process is transformed from overall coarse-grained screening to a gradual propagation and contraction around the topic candidate set, significantly improving the refinement of the topic localization process and the efficiency of constraint utilization. For complex problems involving multiple identity conditions, relationship conditions, attribute conditions, or time conditions, this invention can more effectively achieve multi-constraint joint screening and topic object determination.

[0051] (3) Enhance the system’s ability to fully utilize external knowledge and intermediate states through dynamic topic graph updates and evidence write-back mechanisms.

[0052] While some existing solutions to complex problems can introduce external knowledge through retrieval enhancement mechanisms or retain some historical states through intermediate reasoning steps, their modeling of external evidence and intermediate solution states is often insufficiently continuous and explicit. This makes it difficult to fully express the dynamic relationships between newly added evidence, intermediate attribute values, changes in relational edges, and changes in topic states during the solution process. Existing closer solutions focus more on introducing relevant knowledge before or during reasoning, while their ability to model how newly retrieved evidence is continuously written back into the unified structure, how it drives intermediate state updates, and how it influences subsequent reasoning paths remains inadequate.

[0053] This invention constructs a dynamic topic graph and continuously writes new evidence back into it after knowledge retrieval. This enables the system not only to learn the attributes and relationships of the topic objects themselves, but also to continuously maintain the correspondence between constraint satisfaction states, target attribute states, and evidence node states. In this way, the system can more accurately express the impact of external evidence on the current solution state, enhancing its ability to fully utilize intermediate states and new evidence in complex problems. Especially in tasks such as scientific research analysis, complex interpersonal question answering, corporate knowledge question answering, and multi-step task decision-making, new knowledge evidence often directly influences the subsequent solution direction; this invention can more fully utilize this type of dynamic knowledge information.

[0054] (4) Effectively avoids the inadequacy of fixed reasoning paths to adapt to complex problem scenarios, and significantly improves the stability of the solution process and the path correctability.

[0055] In existing technologies, while a single-chain reasoning approach can effectively facilitate natural language reasoning, it is prone to continuing down the wrong path when deviations occur in intermediate steps. A static problem decomposition approach, while improving the structure of the solution, may lack the ability to backtrack and correct when subsequent evidence deviates from the initial decomposition. Single-round verification, on the other hand, can easily lead to errors in intermediate states being discovered only in the final stage. Existing literature also points out that current complex problem-solving frameworks generally suffer from fixed reasoning paths, difficulty in backtracking from erroneous steps, and a lack of flexible adaptability to subsequent knowledge feedback.

[0056] This invention introduces a dynamic topic graph modeling mechanism, a constraint propagation mechanism, and an adaptive replanning mechanism. This enables the system to simultaneously absorb topic state information, constraint state information, external evidence state information, and verification feedback information in complex problem-solving scenarios. Furthermore, it dynamically adjusts the graph structure, candidate objects, and retrieval paths based on the current solution state, effectively avoiding the problem of fixed inference paths being insufficiently adaptable to complex scenarios. This scheme maintains more stable solution performance under different problem complexities, knowledge completeness levels, and evidence consistency conditions, giving the system better robustness and path correction capabilities during training and inference processes.

[0057] (5) Improve the integrity and scalability of the overall complex problem-solving framework through a unified collaborative mechanism of dynamic topic graph, constraint propagation and adaptive replanning.

[0058] In existing technologies, different studies often improve upon one aspect of problem decomposition, knowledge retrieval, graph representation, or reasoning verification, lacking a unified and collaborative solution mechanism that integrates shared topic object modeling, constraint propagation, knowledge evidence updating, and dynamic path correction. Even though some close solutions have introduced structural representation, retrieval enhancement, and verification mechanisms to improve the performance of solving complex problems, their focus remains more on static problem representation and stepwise solution control, rather than forming a unified and collaborative framework centered on dynamic topic graphs, topic candidate propagation, and adaptive replanning.

[0059] This invention integrates dynamic topic graph modeling, topic candidate entity recall and constraint propagation, knowledge retrieval and evidence write-back, result consistency verification, and adaptive reprogramming into a single complex problem-solving framework. This achieves joint modeling of topic object localization, constraint joint utilization, external knowledge completion, and reasoning path correction, resulting in a system with higher structural integrity and scalability. This approach not only improves the accuracy of answers and the stability of reasoning in current complex problem-solving tasks but also facilitates further expansion to different types of complex task scenarios, such as complex knowledge question answering, scientific research analysis, technical solution derivation, enterprise knowledge base retrieval, and decision support.

[0060] In summary, by introducing a dynamic topic graph modeling mechanism, a topic candidate and constraint propagation mechanism, and an adaptive replanning mechanism, this invention achieves refined control over the interaction between topic objects, constraints, target attributes, and external evidence in complex problems without relying on static expansion of a single path. This effectively improves the stability, controllability, and overall performance of the problem-solving process. Attached Figure Description

[0061] Figure 1 This is a flowchart of the complex problem-solving method based on dynamic topic graphs and adaptive replanning of the present invention.

[0062] Figure 2 This is a schematic diagram of the complex problem-solving system based on dynamic topic graphs and adaptive replanning of the present invention.

[0063] Figure 3 This is a structural block diagram of the complex problem-solving device based on dynamic topic graphs and adaptive replanning of the present invention. Detailed Implementation

[0064] Before detailing any embodiment of the invention, it should be understood that the invention, in its application, is not limited to the details of the construction and arrangement of the components set forth in the following description or illustrated in the following figures. The invention can have other embodiments and can be practiced or carried out in various ways. Furthermore, it should be understood that the wording and terminology used herein are for descriptive purposes and should not be considered limiting. The use of “comprising” or “having” and variations thereof in this invention is intended to cover the items set forth below and their equivalents, as well as any additional items. Unless otherwise specified or limited, the terms “installation,” “connection,” “support,” and “linkage,” and variations thereof are used broadly and cover both direct and indirect installation, connection, support, and linking. Moreover, “connection” and “linkage” are not limited to physical or mechanical connections or links. Furthermore, firstly, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention. Secondly, the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple. The term "a" should not be construed as a limitation on the quantity.

[0065] To address the shortcomings of existing complex problem-solving schemes, this invention proposes a complex problem-solving method based on dynamic topic graphs and adaptive replanning. In the complex problem-solving process, a dynamic topic graph modeling mechanism, constraint propagation and candidate shrinking mechanism, and adaptive replanning mechanism are introduced. Based on the topic objects, constraints, target attributes, and subsequent retrieval feedback in the input problem, topic nodes, constraint nodes, target nodes, and evidence relationships are uniformly modeled and dynamically updated. This constructs a unified solution mechanism that takes into account topic location, joint utilization of constraints, continuous evidence completion, and reasoning path correction. This mechanism is used for subsequent optimization of the complex problem-solving model, thereby improving the overall performance, adaptability, and solution stability of the complex problem-solving system.

[0066] Specifically, this invention improves existing complex problem-solving mechanisms in the following ways: (1) When multiple conditions in the input problem jointly constrain the same subject object, a dynamic subject graph is constructed with the subject object as the center, and multiple constraint nodes are uniformly associated with the subject node to enhance the system's ability to express shared subject entities and multiple constraint dependencies; (2) When different constraints have different filtering effects on candidate subject objects, candidate subject objects are progressively filtered and dynamically sorted through constraint propagation and candidate shrinkage mechanisms to improve the system's accuracy in locating complex subjects; (3) When external knowledge retrieval results supplement new relational information or intermediate evidence, dynamic subject graphs are used to refine the dynamic subject graph. The node status, relation edges and target attributes in the graph are continuously updated to enhance the system’s ability to maintain the state of complex solution processes; (4) When the solution result of the target problem is affected by multiple intermediate relations and external evidence, the topic graph information and evidence set are jointly utilized through graph-driven retrieval and large language model collaborative reasoning mechanism to enhance the system’s ability to solve complex multi-hop problems; (5) When the sample encounters unstable topic candidates, insufficient evidence, constraint conflicts or decreased answer consistency during the solution process, the topic graph structure, candidate status and retrieval path are dynamically repaired and reorganized through adaptive replanning mechanism to coordinate the system’s comprehensive utilization of multiple key information. The main technological breakthroughs achieved by this invention are reflected in the following three aspects: (1) Introduce a dynamic topic graph modeling mechanism for solving complex problems.

[0067] In the process of solving complex problems, instead of using only linear reasoning chains, static problem decomposition structures, or implicit intermediate state representations, a dynamic topic graph is constructed for the subject objects, constraints, and target attributes in the input problem. This enables unified modeling of shared subject entities, multiple constraint relationships, and intermediate state changes, thereby enhancing the system's ability to express complex problem structures.

[0068] (2) Establish a constraint propagation and candidate shrinkage mechanism around the topic candidate objects.

[0069] By performing multi-constraint progressive propagation and candidate narrowing on topic candidates, we can not only screen relevant objects at the overall problem level, but also dynamically adjust the credibility and priority of different candidate objects for different constraint types. This improves the system's adaptability to complex topic localization tasks and avoids the topic binding bias problem caused by coarse-grained reasoning strategies.

[0070] (3) Implement an adaptive replanning mechanism that integrates retrieval feedback and inference verification.

[0071] By introducing retrieval feedback, graph state updates, and result consistency checks during the solution of complex problems, and by adaptively replanning the dynamic topic graph and retrieval path when necessary, the system can not only learn the attributes and relationship information of the topic objects themselves, but also promptly correct the solution path according to subsequent changes in evidence, thereby enhancing the accuracy, robustness, and generalization ability of solving complex problems.

[0072] Figure 1 A flowchart of the complex problem-solving method based on dynamic topic graphs and adaptive replanning of the present invention is shown. Figure 1 As shown, the method for solving complex problems based on dynamic topic graphs and adaptive replanning of the present invention includes the following steps: S1: Problem analysis and initial dynamic topic graph construction.

[0073] First, let the original complex problem be: , This represents a natural language question input by the user. Let the question preprocessing function be... The standardization problem is then expressed as:

[0074] in, This represents the problem after text cleaning, entity phrase normalization, constraint phrase extraction, target attribute identification, and ambiguity resolution; in other words, it is a standardized problem.

[0075] Then, regarding the standardization issue Extracting a set of semantic elements:

[0076] in: Indicates a topic description item; Represents a set of constraints; This represents the target attribute item.

[0077] For example, a standardized question could be: "In what year did the man who served as President of the United States, was one of the authors of the Declaration of Independence, and later founded the University of Virginia, die?" but,

[0078]

[0079]

[0080]

[0081] By extracting a set of semantic elements, the original natural language problem can be transformed into a unified representation of the topic, constraints, and objectives, which facilitates subsequent dynamic topic graph structure modeling and computational processing.

[0082] Finally, based on the semantic element set Construct the initial dynamic theme graph:

[0083] in: This represents the initial dynamic theme graph; Represents the initial set of nodes; This represents the initial set of edges.

[0084] In this invention, the node set is represented as:

[0085] in: Indicates the topic node; Represents the set of constraint nodes; Indicates the target node.

[0086] The edge set is represented as:

[0087] in: This represents the set of connecting edges between topic nodes and constraint nodes. This represents the set of connecting edges between the topic node and the target node.

[0088] Therefore, the topic, constraint, and objective terms can be organized into a unified graph structure, so that subsequent calculations can be carried out around the same topic node.

[0089] S2: Candidate entity recall and constraint propagation and candidate shrinkage.

[0090] Based on the topic nodes, a set of candidate entities is retrieved from external knowledge sources. Constraint propagation and candidate shrinking are then performed on the candidate entity set based on the set of constraint nodes to obtain the current candidate entity set, which specifically includes: 1. Recall candidate entities.

[0091] Around the theme node From external knowledge sources The set of candidate entities to be recalled:

[0092] in: Represents the set of candidate entities; Represents the first in the candidate entity set One candidate entity; This represents the total number of candidate entities in the candidate entity set.

[0093] Let external knowledge sources be :

[0094] in: Representing a knowledge graph; Represents a document knowledge base; Represents a vector retrieval library; This refers to a table library or structured database.

[0095] Meanwhile, the initial candidate score for each candidate entity is defined as:

[0096] in: Indicates the first Candidate entities The initial candidate scores; This represents the balance coefficient between similarity and prior score; Indicates the first Candidate entities With topic nodes Semantic similarity between them; Indicates the first Candidate entities The prior credibility.

[0097] In this invention, By making the topic node Corresponding topic description item and candidate entities The corresponding name or description text is input into the pre-trained language representation model, and after obtaining the corresponding vector representation, cosine similarity is used to calculate it. Can be candidate entities It generates at least one of the following in knowledge graphs, document knowledge bases, and vector retrieval databases: recall ranking, occurrence frequency, entity popularity, historical click statistics, and knowledge link confidence, and can be further normalized to a preset interval.

[0098] Therefore, a set of possible candidate entities can be obtained around the theme node, providing a starting point for gradually narrowing down the theme scope through constraints.

[0099] 2. Implement constraint propagation.

[0100] The set of constraint nodes Each constraint node in the process is applied sequentially to the candidate entity set. Update the score for each candidate entity:

[0101] in: Indicates the first Candidate entities Using the The constraint conditions corresponding to each constraint node Updated candidate scores; Indicates the first Candidate entities Using the The constraint conditions corresponding to each constraint node Updated candidate scores; Indicates the first The constraint conditions corresponding to each constraint node; Indicates the first Candidate entities With the The constraint conditions corresponding to each constraint node Relationship matching score; Indicates the first Candidate entities With the The constraint conditions corresponding to each constraint node Time consistency score; Indicates the first Candidate entities Violation of Article The constraint conditions corresponding to each constraint node Penalties; This represents the corresponding weight parameter.

[0102] In this invention, the relationship matching score Used to characterize the Candidate entities With the The constraint conditions corresponding to each constraint node The degree of matching at the relational level can be determined based on whether the candidate entity has a corresponding relational path, relational edge, or relational description in the external knowledge source. When a direct correspondence or high-confidence relational evidence exists, the relational matching score is higher; otherwise, a lower value is used. Time consistency score. Used to characterize the Candidate entities Time-related information and the first The constraint conditions corresponding to each constraint node The degree of consistency can be determined based on the matching result between the candidate entity's corresponding event, tenure, occurrence time, or duration and the time condition in the constraint node: when the constraint node does not contain a time condition, this item can be set to zero. Constraint violation penalty item. Used to characterize candidate entities For the The constraint conditions corresponding to each constraint node The degree of violation can be determined based on the existence of attributes, relationships, temporal information, or external evidence that contradict the semantic constraints: when there is a clear conflict, the penalty term takes a higher value; otherwise, it takes a lower value or zero. The scores for the above items can be obtained through one or more of the following methods: rule matching, knowledge graph query, text semantic matching, vector similarity calculation, or large language model judgment, and can be weighted and fused after normalization.

[0103] Therefore, multiple constraints can be gradually superimposed on the same set of candidate entities, and the scores of candidate entities can be dynamically updated based on the constraint matching situation.

[0104] 3. Perform candidate shrinkage.

[0105] Based on the updated candidate scores The candidate entity set is filtered to obtain the current candidate entity set, which consists only of the updated candidate scores. Only entities that are greater than or equal to the filtering threshold are considered as updated candidate entities.

[0106] in: Indicates the use of the first The constraint conditions corresponding to each constraint node The updated set of candidate entities; Indicates the use of the first The constraint conditions corresponding to each constraint node The updated set of candidate entities; Indicates the use of the first The constraint conditions corresponding to each constraint node Updated filtering thresholds.

[0107] During candidate shrinkage, candidate shrinkage is stopped when the following condition is met:

[0108] in: Indicates the upper limit of the number of candidate entities; This represents the maximum number of constraint propagation rounds, i.e., the maximum number of constraint nodes used.

[0109] In other words, after several rounds of constraint propagation and screening, the remaining candidate entities are reduced to no more than [number missing]. Once it reaches a certain number, the contraction can stop. Or, even if it hasn't contracted to that number yet... Within one, but the number of rounds of transmission has reached its limit. We must also stop to avoid infinite iteration.

[0110] Therefore, candidate entities can be screened based on the scores after constraint propagation, gradually narrowing down the range of subject entities until a small number of highly credible candidates are obtained.

[0111] After obtaining the updated candidate entity set It can be based on the updated set of candidate entities. Get the updated current dynamic theme image .

[0112] S3: Evidence retrieval and dynamic topic graph updating with large language model reasoning.

[0113] A comprehensive evidence set is obtained by searching based on the current candidate entity set and its corresponding current dynamic topic graph. The current dynamic topic graph is then updated based on the comprehensive evidence set. Furthermore, a candidate answer is obtained by reasoning using a large language model based on the standard question, the updated current dynamic topic graph, the current candidate entity set, and the comprehensive evidence set. This process specifically includes: 1. Perform a local search.

[0114] Based on the current candidate entity set and its corresponding current dynamic theme image Perform a local search:

[0115] in: This represents the set of evidence obtained from a local search; Indicates based on the first Candidate entities Current dynamic theme image The set of constraint nodes in and target node From external knowledge sources Searching for the first Candidate entities Local retrieval functions for directly relevant attribute facts, relational facts, or evidence of target values.

[0116] In this invention, the candidate entity set and constraint node set used in the local retrieval are both current state information after constraint propagation and candidate shrinkage, rather than initial state information.

[0117] 2. Perform a global search.

[0118] Based on the current dynamic topic graph Current candidate entity set and external knowledge sources Perform a global search:

[0119] In the formula, This represents the set of evidence obtained from a global search. This indicates that the current dynamic topic graph is based on... From external knowledge sources Searching for the current candidate entity set A global function for retrieving relevant evidence.

[0120] 3. Obtain a comprehensive set of evidence.

[0121] A comprehensive evidence set is formed based on the results of the local and global searches. :

[0122] Therefore, evidence related to the current candidate entities, constraints, and target attributes can be collected to provide factual support for dynamic topic graph updates and final inference.

[0123] 4. Update the dynamic theme image.

[0124] Based on the aforementioned comprehensive set of evidence Update the current dynamic theme graph. :

[0125] in: Represents the graph update function; This represents the updated current dynamic theme image; In this invention, the current dynamic theme graph is updated. This means updating its nodes and edges.

[0126] Node updates are represented as:

[0127] The edge update is represented as:

[0128] in: This is the node before the update; It is the updated node; These are the edges before the update; It's the updated edge.

[0129] In this invention, This represents the newly added set of nodes, which consists of the current comprehensive evidence set. Generated after parsing, deduplication, and consistency checks, it is used to import the current dynamic topic graph. New nodes that are not yet included in the list but provide support for solving complex problems; This represents the set of edges to be deleted, which is composed of the current comprehensive evidence set. The relationships identified are determined by conflicting relationships, invalid relationships, or low-confidence relationships. This represents the newly added edge set, which is composed of the current comprehensive evidence set. The information on the relationship between newly added nodes and existing nodes is generated to represent the effective connection relationships in the updated dynamic theme graph.

[0130] Therefore, the retrieved factual evidence can be written back to the dynamic topic graph, allowing the dynamic topic graph to be continuously completed and updated as the solution process progresses.

[0131] 5. Perform inference and solution based on a large language model.

[0132] The standard problem Updated current dynamic theme image Current candidate entity set and comprehensive evidence set Common Input Large Language Model Reasoning is performed to generate candidate answers. :

[0133] Therefore, large language models can be used to integrate and reason about graph structure information and evidence sets to output candidate answers corresponding to the question.

[0134] S4: Consistency check and adaptive replanning.

[0135] In this invention, consistency verification and adaptive replanning specifically include: 1. Execution result consistency verification.

[0136] First, regarding the candidate answers Calculate the consistency score:

[0137] in: Indicates candidate answers Consistency score; This represents the weight parameter of the corresponding item.

[0138] In this invention, Indicates candidate answers With the updated current dynamic theme image The structural consistency score can be determined based on the target value corresponding to the candidate answer, the candidate entity pointing relationship, and the matching relationship between the candidate answer and the topic node, constraint node, and target node: when the candidate answer can be consistent with the node relationship and edge relationship in the current dynamic topic graph, the... Take the higher value; otherwise, take the lower value. Indicates candidate answers Combined with the current set of evidence. The evidence consistency score can be determined based on the number of pieces of evidence supporting the candidate answer in the comprehensive evidence set, the confidence level of the evidence, the degree of mutual support between the evidence, and the semantic matching degree between the candidate answer and the content of the evidence: when the candidate answer is directly supported by a large number of high-confidence pieces of evidence, the... Take the higher value; otherwise, take the lower value. Indicates candidate answers The logical consistency score with the set of constraints of the standard question can be determined based on whether the subject entity corresponding to the candidate answer simultaneously satisfies all the constraints in the question, whether there are logical conflicts between the constraints, and whether the candidate answer violates preset constraint logic rules: when the candidate answer satisfies all constraints and there are no logical contradictions, the... Take the higher value, otherwise take the lower value. The scores for each of the above items can be obtained through one or more of the following methods: rule matching, graph structure verification, statistical evidence support, semantic similarity calculation, or consistency judgment using a large language model. They can be weighted and summed after normalization to obtain the consistency score for the candidate answers. .

[0139] Then, based on the consistency score and consistency threshold Determine whether the candidate answer passes the consistency check:

[0140] in: This indicates that the consistency check has been passed; This indicates that the consistency check failed.

[0141] Therefore, it can be determined whether the candidate answer is consistent with the current dynamic topic graph, retrieval evidence, and logical constraints, ensuring the credibility of the output results.

[0142] 2. Trigger adaptive replanning.

[0143] First, when a candidate answer fails the consistency check, or the candidate entity set fails to shrink effectively, the evidence in the comprehensive evidence set is insufficient, or there are conflicts in the updated current dynamic topic graph, the replanning trigger index is calculated. :

[0144] in: Indicates the replanning trigger indicator; Indicates an indicator function; Indicates the maximum number of candidate entities allowed; This indicates the lower limit of the number of valid pieces of evidence in the comprehensive evidence set; This indicates the number of conflicts in the current dynamic theme graph after the update.

[0145] When satisfied At that time, perform adaptive replanning to obtain the dynamic topic graph after adaptive replanning:

[0146] in: This represents the dynamic topic graph after adaptive replanning. This indicates a problem based on the stated standard. Comprehensive evidence collection and candidate answers Updated current dynamic theme image The function that performs replanning.

[0147] In this invention, the adaptive replanning process may include: adjusting the retention range, ranking results, or candidate scores of candidate entities; adding intermediate nodes to connect existing nodes with new evidence; and adjusting the object scope, retrieval order, or retrieval direction of subsequent knowledge retrieval.

[0148] After obtaining the dynamic topic graph after adaptive replanning, we can return to step S2 based on the dynamic topic graph after adaptive replanning, and re-execute the retrieval and reasoning of steps S2 and S3 to obtain candidate answers again and perform consistency verification on the candidate answers obtained again until the replanning trigger index is equal to zero.

[0149] Therefore, when the current solution process fails to converge reliably, the dynamic topic graph and retrieval path can be adjusted to continue the solution task.

[0150] 3. Output the final result.

[0151] The final answer is output when a candidate answer passes the consistency check or reaches the preset termination condition, that is, when the replanning trigger metric equals zero.

[0152] Alternatively, explanatory results can be output:

[0153] in: Represents the final subject entity; This represents the final dynamic theme image.

[0154] Therefore, the final solution to complex problems can be given, and the subject entity and dynamic subject graph can be given at the same time, which facilitates the interpretation and traceability of the results.

[0155] Figure 2 A schematic diagram of the complex problem-solving system based on dynamic topic graphs and adaptive replanning of the present invention is shown. Figure 2 As shown, the complex problem-solving system based on dynamic topic graphs and adaptive replanning of the present invention includes: 1. Problem analysis and initial dynamic theme graph construction module.

[0156] The problem parsing and initial dynamic topic graph construction module is used to extract a set of semantic elements from a standardized problem and construct an initial dynamic topic graph based on the set of semantic elements. The node set of the initial dynamic topic graph includes topic nodes, constraint node set and target node.

[0157] 2. Candidate entity recall, constraint propagation, and candidate shrinking module.

[0158] The candidate entity recall, constraint propagation, and candidate shrinking module is used to retrieve a set of candidate entities from external knowledge sources based on the topic node, and to perform constraint propagation and candidate shrinking on the set of candidate entities based on the set of constraint nodes to obtain the current set of candidate entities.

[0159] 3. Evidence retrieval and dynamic topic graph updating and large language model reasoning module.

[0160] The evidence retrieval, dynamic topic graph update, and large language model reasoning module is used to retrieve a comprehensive evidence set based on the current candidate entity set and its corresponding current dynamic topic graph, update the current dynamic topic graph according to the comprehensive evidence set, and reason using a large language model based on the standard question, the updated current dynamic topic graph, the current candidate entity set, and the comprehensive evidence set to obtain candidate answers.

[0161] 4. Consistency verification and adaptive replanning module.

[0162] The consistency verification and adaptive replanning module is used to perform consistency verification on the candidate answer and, when the candidate answer fails the consistency verification, the current candidate entity set fails to shrink effectively, the comprehensive evidence set is insufficient, or there is a conflict in the updated current dynamic topic graph, perform adaptive replanning to obtain the adaptively replanned dynamic topic graph, and perform re-retrieval and reasoning based on the adaptively replanned dynamic topic graph.

[0163] Furthermore, this invention also provides a device for solving complex problems based on dynamic topic graphs and adaptive replanning. For example... Figure 3 As shown, the complex problem solving device based on dynamic topic graph and adaptive replanning of the present invention includes: a memory 11 for storing one or more programs; one or more processors 12; when the one or more programs are executed by the one or more processors 12, the one or more processors 12 implement the complex problem solving method based on dynamic topic graph and adaptive replanning of the present invention. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the complex problem-solving method based on dynamic topic graphs and adaptive replanning in the present invention.

[0164] The computer-readable storage medium includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined in this invention, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0165] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0166] The steps of the methods or algorithms described in conjunction with the embodiments disclosed in this invention can be implemented in hardware, software modules executed by a processor, or a combination of both. The software modules can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or any other form of storage medium known in the art.

[0167] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention based on the concept of the present invention, without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A method for solving complex problems based on dynamic topic graphs and adaptive replanning, characterized in that, Includes the following steps: S1: Extract a set of semantic elements from the standardized problem and construct an initial dynamic topic graph based on the set of semantic elements. The node set of the initial dynamic topic graph includes topic nodes, constraint node set and target node. S2: Based on the topic node, retrieve the candidate entity set from the external knowledge source, and perform constraint propagation and candidate shrinking on the candidate entity set based on the constraint node set to obtain the current candidate entity set; S3: Based on the current candidate entity set and its corresponding current dynamic topic graph, a comprehensive evidence set is obtained through retrieval. The current dynamic topic graph is updated according to the comprehensive evidence set. Based on the standard question, the updated current dynamic topic graph, the current candidate entity set, and the comprehensive evidence set, reasoning is performed through a large language model to obtain candidate answers. S4: Perform consistency check on the candidate answer. If the candidate answer fails the consistency check, the current candidate entity set fails to shrink effectively, the comprehensive evidence set is insufficient, or there is a conflict in the updated current dynamic topic graph, perform adaptive replanning to obtain the adaptively replanned dynamic topic graph. Then, return to step S2 for re-retrieval and reasoning based on the adaptively replanned dynamic topic graph.

2. The method for solving complex problems based on dynamic topic graphs and adaptive replanning according to claim 1, characterized in that, Step S2 specifically includes: S21: Around the topic node From external knowledge sources The search yields a set of candidate entities. And calculate the initial candidate score for each candidate entity. : In the formula, Represents the set of candidate entities; Represents the first in the candidate entity set One candidate entity; This represents the total number of candidate entities in the candidate entity set; S22: Apply the set of constraint nodes sequentially to the set of candidate entities to update the initial candidate score of each candidate entity, thereby obtaining the updated candidate score. ; S23: Based on the updated candidate scores The candidate entity set is filtered to obtain the current candidate entity set: In the formula, Indicates the use of the first The constraint conditions corresponding to each constraint node The updated set of candidate entities; Indicates the use of the first The constraint conditions corresponding to each constraint node The updated set of candidate entities; Indicates the use of the first The constraint conditions corresponding to each constraint node Updated filtering thresholds.

3. The method for solving complex problems based on dynamic topic graphs and adaptive replanning according to claim 2, characterized in that, In step S21, the initial candidate score for each candidate entity is calculated. Specifically: In the formula, Indicates the first Candidate entities The initial candidate scores; This represents the balance coefficient between similarity and prior score; Indicates the first Candidate entities With topic nodes Semantic similarity between them; Indicates the first Candidate entities The prior credibility.

4. The method for solving complex problems based on dynamic topic graphs and adaptive replanning according to claim 2, characterized in that, Step S22 specifically involves: In the formula, Indicates the first Candidate entities Using the The constraint conditions corresponding to each constraint node Updated candidate scores; Indicates the first Candidate entities Using the The constraint conditions corresponding to each constraint node Updated candidate scores; Indicates the first The constraint conditions corresponding to each constraint node; Indicates the first Candidate entities With the The constraint conditions corresponding to each constraint node Relationship matching score; Indicates the first Candidate entities With the The constraint conditions corresponding to each constraint node Time consistency score; Indicates the first Candidate entities Violation of Article The constraint conditions corresponding to each constraint node Penalties; This represents the corresponding weight parameter.

5. The method for solving complex problems based on dynamic topic graphs and adaptive replanning according to claim 2, characterized in that, Step S3 specifically includes: S31: Based on the current candidate entity set and its corresponding current dynamic theme image Perform a local search: In the formula, This represents the set of evidence obtained from a local search; Indicates based on the first Candidate entities Current dynamic theme image The set of constraint nodes in and target node From external knowledge sources Searching for the first Candidate entities Local retrieval functions for directly relevant attribute facts, relational facts, or evidence of target values; S32: Based on the current dynamic topic graph Current candidate entity set and external knowledge sources Perform a global search: In the formula, This represents the set of evidence obtained from a global search. This indicates that the current dynamic topic graph is based on... From external knowledge sources Searching for the current candidate entity set A global retrieval function for relevant evidence; S33: Based on the results of the local and global searches, a comprehensive evidence set is formed. : S34: Based on the aforementioned comprehensive set of evidence Update the current dynamic theme graph. : In the formula, Represents the graph update function; This represents the updated current dynamic theme image; S35: Based on the aforementioned standard issue Updated current dynamic theme image Current candidate entity set and comprehensive evidence set Reasoning is performed using a large language model to obtain candidate answers. :

6. The method for solving complex problems based on dynamic topic graphs and adaptive replanning according to claim 5, characterized in that, The consistency check performed on the candidate answer in step S4 specifically includes: S41: Regarding the candidate answers Calculate the consistency score: In the formula, Indicates candidate answers Consistency score; Indicates the selection of answers With the updated current dynamic theme image Consistency score; Indicates candidate answers With comprehensive evidence set Consistency score; Indicates candidate answers With the aforementioned standard problem The consistency score of the constraint node set; This represents the weight parameter for the corresponding item; S42: Based on the aforementioned consistency score and consistency threshold Determine whether the candidate answer passes the consistency check: In the formula, This indicates that the consistency check has been passed; This indicates that the consistency check failed.

7. The method for solving complex problems based on dynamic topic graphs and adaptive replanning according to claim 6, characterized in that, Step S4, which involves performing adaptive replanning to obtain an adaptively replanned dynamic topic graph when the candidate answer fails the consistency check, the current candidate entity set fails to shrink effectively, the comprehensive evidence set is insufficient, or the updated current dynamic topic graph contains conflicts, specifically includes: S43: Calculate replanning trigger indicators : In the formula, Indicates the replanning trigger indicator; Indicates an indicator function; Indicates the maximum number of candidate entities allowed; This indicates the lower limit of the number of valid pieces of evidence in the comprehensive evidence set; This indicates the number of conflicts in the updated current dynamic theme graph; S44: When satisfied At that time, perform adaptive replanning to obtain the dynamic topic graph after adaptive replanning: In the formula, This represents the dynamic topic graph after adaptive replanning. This indicates a problem based on the stated standard. Comprehensive evidence collection and candidate answers Updated current dynamic theme image The function that performs replanning.

8. A system for solving complex problems based on dynamic topic graphs and adaptive replanning, characterized in that, include: The problem analysis and initial dynamic topic graph construction module is used to extract a set of semantic elements from a standardized problem and construct an initial dynamic topic graph based on the set of semantic elements. The node set of the initial dynamic topic graph includes topic nodes, constraint node set and target nodes. The candidate entity recall, constraint propagation and candidate shrinking module is used to retrieve a set of candidate entities from external knowledge sources based on the topic node, and to perform constraint propagation and candidate shrinking on the set of candidate entities based on the set of constraint nodes to obtain the current set of candidate entities. The evidence retrieval and dynamic topic graph update and large language model reasoning module is used to retrieve a comprehensive evidence set based on the current candidate entity set and its corresponding current dynamic topic graph, update the current dynamic topic graph according to the comprehensive evidence set, and reason through a large language model based on the standard question, the updated current dynamic topic graph, the current candidate entity set and the comprehensive evidence set to obtain candidate answers. The consistency verification and adaptive replanning module is used to perform consistency verification on the candidate answer and, when the candidate answer fails the consistency verification, the current candidate entity set fails to shrink effectively, the comprehensive evidence set is insufficient, or there is a conflict in the updated current dynamic topic graph, perform adaptive replanning to obtain the adaptively replanned dynamic topic graph, and perform re-retrieval and reasoning based on the adaptively replanned dynamic topic graph.

9. A complex problem-solving device based on dynamic topic graphs and adaptive replanning, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the complex problem-solving method based on dynamic topic graphs and adaptive replanning as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the complex problem-solving method based on dynamic topic graphs and adaptive replanning as described in any one of claims 1-7.

Images