A Causal Consistency Long-Chain Inference Method Based on Structured Semantic Parsing
By decomposing task instructions into structured semantic representations and constructing causal reasoning graphs, the problem of unstable causal logic in large models for complex tasks is solved, interpretable pipelined execution is achieved, and the accuracy and reliability of long-chain reasoning are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENYANG INST OF COMPUTING TECH CO LTD THE CHINESE ACAD OF SCI
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-30
AI Technical Summary
Existing large models lack explicit structure in complex task execution, making intermediate processes difficult to interpret, subtask boundaries ambiguous, and problems such as causal inversion, information jumps, or path breaks easily occur in long-chain reasoning.
By using a structured semantic parsing method, task instructions are decomposed into structured semantic representations with causal relationships, a multi-stage causal reasoning graph is constructed, and confidence assessment and logical verification are performed on intermediate states. A hierarchical backtracking strategy is then used for correction.
An explicit causal reasoning framework was implemented, which improved the logical accuracy and path stability of long-chain reasoning and enhanced the robustness and reliability of the system.
Smart Images
Figure CN122311482A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of semantic reasoning, specifically a causal consistency long-chain reasoning method based on structured semantic parsing. Background Technology
[0002] When applying large models to complex task execution, the inference process often lacks explicit structure and logical control. Existing methods mostly employ end-to-end implicit prediction, which, while generating the final result, leaves intermediate processes difficult to interpret, subtask boundaries ambiguous, and prone to problems such as causal inversion, information jumps, or path breaks in long-chain inference. Traditional task decomposition methods focus on how to break down tasks but neglect causal consistency between subtasks, resulting in a lack of logical stability in the generated execution plan. Therefore, how to construct an inference mechanism similar to pipeline execution, making each step interpretable, checkable, and error-correctable, has become a crucial scientific problem that urgently needs to be solved. Summary of the Invention
[0003] Based on the aforementioned technical problems, this invention proposes a causal consistency long-chain reasoning method based on structured semantic parsing to address the issues of lack of explicit structure, difficulty in interpreting intermediate states, and instability of causal logic in existing large-scale model reasoning processes.
[0004] The technical solution adopted by the present invention to achieve the above objectives is as follows:
[0005] The causal consistency long-chain reasoning method based on structured semantic parsing includes the following steps:
[0006] 1) Decompose task instructions into structured semantic representations with causal relationships using a language model;
[0007] 2) Construct a multi-stage causal reasoning graph based on structured semantic representation and generate reasoning paths that conform to the task logic;
[0008] 3) Conduct confidence assessment and logical verification of intermediate states in the reasoning process.
[0009] Step 1) includes the following steps:
[0010] 1.1) By using a pre-trained language model and combining it with causal prompts, the task instructions are decomposed into multiple sub-task sequences with causal relationships;
[0011] 1.2) The attention-based relation extraction algorithm identifies the causal dependencies between subtask sequences, uses them as edges in the graph neural network, and uses subtasks as nodes to construct a task causal relationship graph.
[0012] Step 2) includes the following steps:
[0013] 2.1) The task causal relationship graph is reasoned through a multi-stage causal reasoning engine to construct a causal consistency graph, an execution stage is assigned to each sub-task node to obtain the reasoning path to complete the task, and reasoning paths that do not conform to causal logic are eliminated through counterfactual reasoning.
[0014] 2.2) By using a hierarchical planning strategy, long-chain tasks are divided into macro stages and micro steps. First, the macro stages, including preparation, execution, and verification, are planned, and then the micro sub-tasks within each stage are refined to generate executable reasoning paths layer by layer.
[0015] Step 3) includes the following steps:
[0016] 3.1) Calculate the confidence score of the current intermediate state at the end of each inference phase;
[0017] 3.2) Perform logical verification on the intermediate state to check whether it meets the preset causal constraints, which include the integrity of the preconditions, the correctness of the dependency relationship, and the detection of circular dependencies.
[0018] 3.3) If an anomaly is detected during confidence testing or logic verification, a backtracking correction mechanism or replanning is triggered.
[0019] The backtracking correction mechanism adopts a hierarchical backtracking strategy, namely:
[0020] The confidence score is compared with a preset first threshold and a second threshold, where the first threshold is higher than the second threshold. If the confidence score is higher than the first threshold, it is considered a pass, and the next stage continues. If the confidence score is between the first and second thresholds, it is considered a slight deviation, triggering a local backtracking correction. The process backtracks to the micro-level of the current stage for local adjustments and then re-deduces the results. If the confidence score is lower than the second threshold, or if the logic verification detects a causal conflict, it is considered a serious violation, triggering a global replanning. The process backtracks to the structured semantic parsing level and re-parses the task and constructs the causal graph.
[0021] The local backtracking correction specifically refers to:
[0022] Locate the abnormal node and its stage, analyze the cause of the abnormality, adjust only the micro steps within the stage, and retain the planning results of other stages; after adjustment, re-perform path fusion and consistency verification.
[0023] The global replanning specifically refers to:
[0024] Record the type and location information of serious violations, use the violation information as a feedback signal, re-execute step 1), adjust the weight distribution of causal prompts, re-segment subtasks and identify dependencies, and completely reconstruct the causal graph and inference path.
[0025] A causal consistency long-chain reasoning system based on structured semantic parsing includes:
[0026] The semantic structured representation module is used to decompose task instructions into structured semantic representations with causal relationships through a language model;
[0027] The causal consistency reasoning module is used to construct a multi-stage causal reasoning graph based on structured semantic representation and generate reasoning paths that conform to task logic.
[0028] The path evaluation and error correction module is used to evaluate the confidence level and verify the logic of intermediate states in the reasoning process.
[0029] The present invention has the following beneficial effects and advantages:
[0030] 1. This invention constructs an explicit causal reasoning framework, transforming implicit black-box reasoning into interpretable pipelined execution.
[0031] 2. By introducing causal consistency graphs and counterfactual constraints, this invention significantly improves the logical accuracy and path stability of long-chain reasoning.
[0032] 3. The evaluation and error correction mechanism designed in this invention can monitor the reasoning state in real time and intervene, thereby enhancing the robustness and reliability of the system. Attached Figure Description
[0033] Figure 1 This is an overall flowchart of the causal consistency long-chain reasoning method based on structured semantic parsing of the present invention;
[0034] Figure 2 This is a diagram showing the overall network structure of the causal consistency long-chain reasoning method based on structured semantic parsing of this invention. Detailed Implementation
[0035] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0036] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.
[0037] A causal consistency long-chain reasoning method based on structured semantic parsing includes the following steps:
[0038] A semantic structured representation module is constructed to obtain complex task instructions and perform structured semantic parsing to obtain a sequence of subtasks containing multiple causal dependencies. The sequence of subtasks constitutes a structured semantic representation.
[0039] Construct a causal consistency reasoning module to enable multi-stage causal reasoning for complex tasks;
[0040] A path evaluation and error correction module is constructed to evaluate the confidence and verify the logic of intermediate states in the reasoning process, thereby ensuring the causal reliability of the reasoning process.
[0041] The semantic structured representation module performs the following steps:
[0042] The causal semantic parsing module decomposes the input natural language task into semantic units with causal relationships;
[0043] The structured representation building module organizes the semantic modules and their dependencies into a directed acyclic graph or tree structure.
[0044] The causal semantic parsing module performs the following steps:
[0045] By combining a pre-trained language model with an attention-based relation extraction algorithm, causal dependencies between semantic units are identified. Graph neural networks are then used to model and optimize the dependencies between task nodes, generating a structured task causal relationship graph. This allows a continuous task flow to be divided into sub-task nodes with clear sequential dependencies.
[0046] The causal consistency reasoning module generates a reasoning path that conforms to the task logic. Specifically, it inputs the structured semantic representation into a multi-stage causal reasoning engine. This engine constructs a causal consistency graph, assigns an execution stage to each sub-task node, and generates a global reasoning path based on the dependencies between stages. The causal consistency graph introduces a counterfactual constraint mechanism to ensure that the execution order between sub-tasks conforms to causal logic.
[0047] The generation process of reasoning paths adopts a hierarchical planning strategy, that is, first planning the macro stage, including preparation, execution, verification and other steps, and then refining the micro sub-tasks within each stage to ensure the semantic coherence of long-chain reasoning.
[0048] The path evaluation and error correction module performs the following steps:
[0049] The confidence assessment module calculates the confidence score of the current intermediate state at the end of each inference phase. The confidence score is calculated based on the satisfaction of preconditions, resource availability, and historical execution success rate.
[0050] The logic verification module detects whether the intermediate state meets the preset causal constraints, which include the integrity of the preconditions, the correctness of the dependency relationship, and the detection of circular dependencies.
[0051] The error correction module triggers a backtracking correction mechanism when it detects an anomaly.
[0052] The error correction mechanism employs a tiered backtracking strategy: the confidence score is compared with a preset first threshold and a second threshold, where the first threshold is higher than the second threshold. If the confidence score is higher than the first threshold, the process is considered successful, and the next stage continues. If the confidence score is between the first and second thresholds, it is considered a minor deviation, triggering a local backtracking correction. The process retraces to the micro-level of the current stage for local adjustments and then re-deduces the results. If the confidence score is lower than the second threshold, or the logic verification unit detects a serious causal conflict, the process is considered a serious violation, triggering a global replanning. The process retraces to the structured semantic parsing level, and the task parsing and causal graph construction are performed again.
[0053] Local backtracking correction includes: locating abnormal nodes and their respective stages; analyzing the causes of the abnormality, including missing preconditions, incorrect execution order, or resource allocation conflicts; adjusting only the micro-steps within that stage while retaining the planning results of other stages; and re-performing path fusion and consistency verification after adjustment.
[0054] The global replanning process includes: recording the type and location information of serious violations; inputting violation information as a feedback signal into the structured semantic parsing module; adjusting the weight distribution of causal prompts, re-segmenting subtasks and identifying dependencies; and completely reconstructing the causal graph and inference path.
[0055] A causal consistency long-chain reasoning system based on structured semantic parsing includes the following modules:
[0056] The semantic structured representation module is used to decompose complex tasks into structured semantic units with causal relationships;
[0057] The causal consistency reasoning module is used to construct multi-stage causal reasoning graphs and generate reasoning paths that conform to the task logic.
[0058] The path evaluation and error correction module is used to evaluate the confidence level and verify the logic of intermediate states, ensuring the causal reliability of the reasoning process.
[0059] The semantic structured representation module includes:
[0060] The causal semantic parsing module is used to parse the input task into a sequence of subtasks containing causal relationships by combining a large model with causal prompts;
[0061] A structured representation building module is used to organize the subtask sequence and its dependencies into a graph structure.
[0062] The causal consistency reasoning module includes:
[0063] The causal consistency graph construction module is used to construct causal graphs based on the causal dependencies of subtasks and eliminate paths that do not conform to causal logic through counterfactual reasoning;
[0064] The hierarchical planning module is used to divide long-chain tasks into macro-stages and micro-steps, and refine them layer by layer to generate executable reasoning paths.
[0065] The path evaluation and error correction module includes:
[0066] The confidence assessment module is used to calculate the confidence level of intermediate results at each inference stage;
[0067] The logic verification module is used to detect whether the intermediate state meets the preset causal constraints and whether the preconditions are met.
[0068] The error correction execution module is used to trigger backtracking correction or replanning when an anomaly is detected.
[0069] Example
[0070] This invention relates to a causal consistency long-chain reasoning method based on structured semantic parsing, comprising a semantic structured representation module, a causal consistency reasoning module, and a path evaluation and error correction module. The semantic structured representation module parses complex tasks described in natural language into structured semantic units with causal relationships; the causal consistency reasoning module constructs a multi-stage causal reasoning graph based on this structured representation, generating explicit and interpretable reasoning paths; the path evaluation and error correction module monitors and verifies intermediate states during the reasoning process, ensuring the reliability and causal consistency of each step. This invention aims to make the reasoning process like an execution pipeline, with each step possessing interpretability, checkability, and error correctability.
[0071] A causal consistency long-chain reasoning method based on structured semantic parsing includes:
[0072] A semantic structured representation module is used to enable causal analysis and structured modeling of complex tasks.
[0073] A causal consistency reasoning module is used to enable multi-stage, interpretable reasoning for tasks.
[0074] The path evaluation and error correction module is used to monitor and correct the model inference process.
[0075] The semantic structured representation module includes a causal semantic parsing module and a structured representation construction module. The causal semantic parsing module uses a pre-trained language model combined with causal prompts to decompose complex instructions into multiple sub-tasks with causal relationships. For example, the instruction "organize a meeting" can be parsed into sub-tasks such as "determine attendees," "coordinate time," "book a meeting room," and "send invitations," and identifies "determine attendees" as a precondition for "coordinate time." The structured representation construction module organizes these nodes and their dependencies into a directed acyclic graph.
[0076] The causal consistency reasoning module receives the aforementioned graph structure and constructs a causal consistency graph. This graph not only includes positive dependencies but also introduces counterfactual constraints. For example, if A is not completed, then B should not be executed, thus eliminating logically erroneous paths. Based on this, a hierarchical planning strategy is adopted: first, the macro-level stages are planned, namely the information gathering stage, the decision-making stage, and the execution stage; then, specific micro-level operational steps are refined within each stage, ultimately generating a coherent global reasoning path that conforms to causal logic.
[0077] The path evaluation and error correction module plays a role in the inference execution process. After each stage or key sub-task is completed, the confidence evaluation module calculates the confidence score of the current intermediate state. If the score is lower than a preset threshold, or the logic verification module detects a causal conflict, for example, if it is found that the time has not been coordinated when booking a meeting room, the error correction execution module intervenes, the system backtracks to the previous stage, and performs local correction or triggers global replanning based on the error type.
[0078] This invention constructs a causal consistency long-chain reasoning method based on structured semantic parsing, which has strong stability and interpretability.
[0079] like Figure 1 , Figure 2 As shown, a causal consistency long-chain reasoning method based on structured semantic parsing includes: a semantic structured representation module, a causal consistency reasoning module, and a path evaluation and error correction module. The semantic structured representation module is used to parse the input task into structured semantic units with causal relationships; the causal consistency reasoning module constructs a multi-stage causal reasoning graph based on this and generates reasoning paths; the path evaluation and error correction module performs real-time monitoring and error correction during the reasoning process, and finally outputs a stable and reliable execution plan.
[0080] The causal consistency reasoning module, for the structured semantic representation of the input, first generates a causal graph containing positive dependencies and counterfactual constraints through the causal consistency graph construction module. Then, the hierarchical planning module divides it into several macro-stages based on the complexity of the task. Within each macro-stage, a path search algorithm generates specific micro-reasoning steps, ensuring that the entire long-chain reasoning process conforms to high-level logic while also possessing detailed executableness.
[0081] Constructing a causal consistency long-chain reasoning method based on structured semantic parsing includes the following steps:
[0082] S1: Obtain task instructions and construct a structured semantic representation through the semantic structured representation module.
[0083] S2: Generate multi-stage causal reasoning paths through the causal consistency reasoning module.
[0084] S3: Performs reasoning and performs state monitoring and logic verification at each stage through the path evaluation and error correction module.
[0085] S4: If all stages pass verification, output the final reasoning result; otherwise, backtrack and correct according to the error correction mechanism, and repeat step S2 or S3.
[0086] The present invention discloses a causal consistency long-chain reasoning method based on structured semantic parsing. Through a semantic structured representation module, a causal consistency reasoning module, and a path evaluation and error correction module, it constructs an interpretable and error-correctable reasoning mechanism, which can effectively improve the logical accuracy and robustness of large models in complex task execution.
[0087] 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 be considered within the scope of protection of the present invention.
Claims
1. A causal consistency long-chain reasoning method based on structured semantic parsing, characterized in that, Includes the following steps: 1) Decompose task instructions into structured semantic representations with causal relationships using a language model; 2) Construct a multi-stage causal reasoning graph based on structured semantic representation and generate reasoning paths that conform to the task logic; 3) Conduct confidence assessment and logical verification of intermediate states in the reasoning process.
2. The causal consistency long-chain reasoning method based on structured semantic parsing according to claim 1, characterized in that, Step 1) includes the following steps: 1.1) By using a pre-trained language model and combining it with causal prompts, the task instructions are decomposed into multiple sub-task sequences with causal relationships; 1.2) The attention-based relation extraction algorithm identifies the causal dependencies between subtask sequences, uses them as edges in the graph neural network, and uses subtasks as nodes to construct a task causal relationship graph.
3. The causal consistency long-chain reasoning method based on structured semantic parsing according to claim 1, characterized in that, Step 2) includes the following steps: 2.1) The task causal relationship graph is reasoned through a multi-stage causal reasoning engine to construct a causal consistency graph, an execution stage is assigned to each sub-task node to obtain the reasoning path to complete the task, and reasoning paths that do not conform to causal logic are eliminated through counterfactual reasoning. 2.2) By using a hierarchical planning strategy, long-chain tasks are divided into macro stages and micro steps. First, the macro stages, including preparation, execution, and verification, are planned, and then the micro sub-tasks within each stage are refined to generate executable reasoning paths layer by layer.
4. The causal consistency long-chain reasoning method based on structured semantic parsing according to claim 1, characterized in that, Step 3) includes the following steps: 3.1) Calculate the confidence score of the current intermediate state at the end of each inference phase; 3.2) Perform logical verification on the intermediate state to check whether it meets the preset causal constraints, which include the integrity of the preconditions, the correctness of the dependency relationship, and the detection of circular dependencies. 3.3) If an anomaly is detected during confidence testing or logic verification, a backtracking correction mechanism or replanning is triggered.
5. The causal consistency long-chain reasoning method based on structured semantic parsing according to claim 4, characterized in that, The backtracking correction mechanism adopts a hierarchical backtracking strategy, namely: The confidence score is compared with a preset first threshold and a second threshold, where the first threshold is higher than the second threshold. If the confidence score is higher than the first threshold, it is considered a pass, and the next stage continues. If the confidence score is between the first and second thresholds, it is considered a slight deviation, triggering a local backtracking correction. The process backtracks to the micro-level of the current stage for local adjustments and then re-deduces the results. If the confidence score is lower than the second threshold, or if the logic verification detects a causal conflict, it is considered a serious violation, triggering a global replanning. The process backtracks to the structured semantic parsing level and re-parses the task and constructs the causal graph.
6. The causal consistency long-chain reasoning method based on structured semantic parsing according to claim 5, characterized in that, The local backtracking correction specifically refers to: Locate the abnormal node and its stage, analyze the cause of the abnormality, adjust only the micro steps within the stage, and retain the planning results of other stages; after adjustment, re-perform path fusion and consistency verification.
7. The causal consistency long-chain reasoning method based on structured semantic parsing according to claim 5, characterized in that, The global replanning specifically refers to: Record the type and location information of serious violations, use the violation information as a feedback signal, re-execute step 1), adjust the weight distribution of causal prompts, re-segment subtasks and identify dependencies, and completely reconstruct the causal graph and inference path.
8. A causal consistency long-chain reasoning system based on structured semantic parsing, characterized in that, include: The semantic structured representation module is used to decompose task instructions into structured semantic representations with causal relationships through a language model; The causal consistency reasoning module is used to construct a multi-stage causal reasoning graph based on structured semantic representation and generate reasoning paths that conform to task logic. The path evaluation and error correction module is used to evaluate the confidence level and verify the logic of intermediate states in the reasoning process.