A unified information extraction method based on thought chain reasoning enhancement

Abstract

The application discloses a unified information extraction method based on thought chain reasoning enhancement, and belongs to the technical field of large model information processing. The method comprises the following steps: S1, thought chain data set formation: based on an existing information extraction data set, a training sample is constructed to form a thought chain training data set; S2, thought chain supervised fine-tuning: a large language model is supervised and fine-tuned to obtain an initial reasoning strategy model; S3, reasoning enhancement optimization: the initial reasoning strategy model is optimized by using a GRPO algorithm to obtain a final information extraction model; and S4, reasoning generation: the final information extraction model generates a thought chain reasoning path and a structured extraction result in sequence, and the final information extraction content is obtained by analyzing the output content. By using the method, a unified reasoning extraction framework is built, reasoning training samples are automatically generated, the cross-domain extraction precision is greatly improved through double model optimization, the reasoning process is traceable, the output format is easy to analyze and is normative, and the landing cost is lower.

Description

Technical Field

[0001] This invention relates to the field of large-scale model information processing technology, and in particular to a unified information extraction method based on enhanced thinking chain reasoning. Background Technology

[0002] Unified Information Extraction (UIE) aims to perform multiple extraction tasks simultaneously, such as named entity recognition, relation extraction, event detection, and event argument extraction, using a single model. It is an important research direction in the field of natural language processing. This technology can transform unstructured text into queryable and computable structured data in batches, and is widely used in knowledge graph construction, intelligent document understanding, low-resource domain information processing, and intelligent question answering. With the development of large language models and instruction fine-tuning techniques, unified information extraction is gradually shifting from traditional discriminative multi-task models to a unified text-to-structure paradigm based on generative large models.

[0003] In the prior art, the solutions closest to this invention mainly fall into two categories: The first category is unified information extraction methods based on direct generation, typically represented by UIE and Instruct UIE. These methods transform different extraction tasks into text generation problems, guiding large language models to directly output structured extraction results by constructing instructions containing task definitions, extraction schemas, and input text. While this approach achieves unified modeling for multiple tasks, it suffers from several key flaws: First, the model's prediction process is an end-to-end black-box mapping, lacking explicit intermediate inference steps. The model primarily relies on surface semantic patterns in the training data for prediction, failing to explain the contextual basis of the extraction results, resulting in poor interpretability. Second, when the distribution of test data differs significantly from training data (e.g., cross-domain, low-resource, zero-sample scenarios), or when sentence semantic structures are complex, the model is prone to errors such as entity boundary recognition errors, relation type judgment biases, and omissions of event arguments, leading to insufficient generalization ability and robustness. Third, even with richer schema descriptions and examples, the model's intermediate decision logic remains implicit in the parameters, making it impossible to verify and correct through explicit inference chains.

[0004] The second category is reasoning enhancement methods oriented towards a single subtask, typically represented by PromptNER for named entity recognition and CoT-ER and ERA-CoT for relation extraction. These methods attempt to introduce thought chain reasoning into information extraction, guiding the model to generate judgment criteria or reasoning processes before outputting results through prompts. However, these solutions have significant limitations: First, most are designed only for single subtasks such as named entity recognition and relation extraction, failing to form a unified reasoning framework that simultaneously covers entities, relations, event trigger words, and event arguments, thus unable to support joint information extraction; second, most solutions rely solely on prompt engineering for reasoning, lacking dedicated supervised fine-tuning and optimization mechanisms for the reasoning process, resulting in unstable reasoning quality and difficulty in fully leveraging the reasoning capabilities of large models; third, existing solutions pay little attention to the format validity of the output structured results, leading to format errors in the generated structured content, affecting subsequent parsing and use. Summary of the Invention

[0005] The purpose of this invention is to provide a unified information extraction method based on enhanced thinking chain reasoning, thereby solving the aforementioned technical problems.

[0006] To achieve the above objectives, this invention provides a unified information extraction method based on enhanced thought chain reasoning, comprising the following steps: S1. Synthesis of the Thinking Chain Dataset: Based on the existing information extraction dataset, multi-path sampling, consistency filtering, and thinking chain quality assessment are performed in sequence to construct training samples containing input sentences, task schemas, thinking chain reasoning paths, and structured extraction results, forming the thinking chain training dataset. S2, Mind Chain Supervised Fine-Tuning: The input sentence and task schema are used as input to the large language model. The mind chain reasoning path and structured extraction results are concatenated to form the target output sequence. The large language model is then supervised and fine-tuned to obtain the initial reasoning strategy model. S3. Inference Enhancement Optimization: The GRPO algorithm is used to perform reinforcement learning optimization on the initial inference strategy model. By using a hybrid reward function to simultaneously constrain the extraction accuracy and the validity of the output format, the final information extraction model is obtained. S4. Reasoning Generation: Input the text to be extracted and the target schema into the final information extraction model. The final information extraction model generates the thought chain reasoning path and the structured extraction result in sequence. The output content is parsed to obtain the final information extraction content.

[0007] Preferably, the structured extraction results cover four types of tasks: named entity recognition, relation extraction, event detection, and event argument extraction.

[0008] Preferably, the multi-path sampling in S1 is as follows: for each input sample, a prompt template containing reasoning requirements is constructed, and multiple sets of candidate outputs are generated using the teacher's large language model. Each set of candidate outputs contains a thought chain reasoning path and the corresponding structured extraction result.

[0009] Preferably, the consistency filtering in S1 is as follows: for labeled supervised data, retain candidate samples whose structured results completely match the standard labels; if there are multiple matching candidates, select the sample with the shortest inference path; for unlabeled unsupervised data, calculate the F1 value of each candidate result as an anchor point and compare it with other candidate results, take the average value as the consistency score, and retain the candidate sample with the highest consistency score.

[0010] Preferably, the quality assessment of the thought chain in S1 is as follows: an auxiliary evaluation large language model is used as an evaluator to score the retained candidate samples. The evaluation dimensions include reasoning-result consistency and logical rationality. Samples whose scores in both dimensions reach the preset threshold are retained as the final training data.

[0011] Preferably, the training objective of the supervised fine-tuning of the thought chain in S2 is to minimize the negative log-likelihood loss, and its loss function is... for: ; in, For mathematical expectation symbol, For trainable parameters of a large language model, For training the mind chain dataset, For the input sample, Output the target sequence. For the first in the target sequence Each word element, For the first All lexical units preceding the given lexical unit, Given the input and the preceding text, the conditional probability of generating the current word.

[0012] Preferably, the specific process of using the GRPO algorithm for policy update in S3 is as follows: for each input sample, a set of candidate outputs is sampled from the old policy model, the difference between the reward value of each candidate output and the average reward within the group is calculated as the advantage value, and the current inference policy model is updated based on the advantage value.

[0013] Preferably, the total reward of the hybrid reward function is a weighted combination of the soft F1 reward and the format validity reward, as shown below: ; in, For the first Total reward for each candidate output This is a soft F1 reward, used to measure the degree of overlap and matching between the extracted results and the standard results. This is a format validity reward used to determine whether the output conforms to a preset structured format. and These are the weighting coefficients for the soft F1 reward and the format validity reward, respectively.

[0014] Preferably, the optimization objective of the GRPO algorithm includes a KL divergence regularization term to constrain the distribution difference between the current inference strategy model and the reference model, thereby preventing the current inference strategy model from deviating from the inference behavior obtained by supervised fine-tuning. The parameters of the reference model are initialized to the initial inference strategy model parameters after the thought chain fine-tuning and remain fixed during the reinforcement learning phase.

[0015] Preferably, the final information extraction model in S4 first generates a step-by-step reasoning process based on contextual evidence and schema constraints, and then generates a structured extraction result conforming to JSON format. The final information extraction content obtained after parsing includes a set of entities, a set of relations, a set of event trigger words, and a set of event arguments.

[0016] Therefore, the present invention employs the above-mentioned unified information extraction method based on enhanced thinking chain reasoning, which has the following beneficial effects: 1. The output format of the unified information extraction is improved from "directly outputting results" to a two-stage format of "first outputting the reasoning process of the thought chain, and then outputting the structured result." The reasoning process can fully cover the entire process of entity recognition, relationship judgment, event detection, and argument matching. The model output is no longer a black-box prediction, but a reasoning result with contextual basis. The extraction result can be traced and manually verified through the thought chain, which significantly improves the interpretability and error correction of information extraction.

[0017] 2. Through an automated process of "multi-path sampling - consistency filtering - quality assessment," a training dataset containing a complete inference chain is automatically constructed by extracting data from existing unlabeled information. This method eliminates the need for manual annotation of the inference process, significantly reducing data construction costs. Simultaneously, multiple filtering mechanisms ensure the consistency and logical rationality of the inference path and extraction results, providing high-quality supervision signals for training the model's inference capabilities.

[0018] 3. Supervised fine-tuning enables the model to master a two-stage output paradigm of "inference-extraction," and then GRPO (Group Relative Policy Optimization) reinforcement learning method is used to further optimize the inference behavior. A hybrid reward function balances extraction accuracy (soft F1 reward) and output format effectiveness, improving extraction accuracy in complex scenarios while ensuring the parsingability of structured output. Experiments show that this method significantly outperforms existing generative methods and general inference models in cross-domain zero-shot scenarios for entity recognition, relation extraction, and event extraction.

[0019] 4. The explicit reasoning mechanism guides the model to make logical judgments based on contextual evidence and schema constraints, rather than simply relying on surface patterns in the training data. This allows the model to better transfer to unseen domains and low-resource tasks. In small-sample settings, this method achieves higher extraction performance with less labeled data compared to the baseline model, demonstrating superior sample efficiency.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] Fig. 1 This is a flowchart of a unified information extraction method based on enhanced thinking chain reasoning according to the present invention. Fig. 2 This is an architecture diagram of a unified information extraction method based on enhanced thinking chain reasoning according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0023] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] like Figs. 1-2 As shown, this invention provides a unified information extraction method based on enhanced reasoning using thought chains. It improves the traditional direct generation method of unified information extraction into a two-stage generation model of "explicit reasoning process + structured result output." Through three core steps—thought chain data synthesis, supervised fine-tuning, and reinforcement learning enhancement—it improves the model's extraction accuracy and result interpretability in cross-domain, low-resource scenarios. This embodiment generally includes four parts: task formal definition and thought chain dataset synthesis, supervised fine-tuning of thought chains, inference enhancement optimization, and inference and output parsing, as detailed below: S1. Formal Definition of Task: Let the input sentence be a sequence of words. ,in, Indicates the first in the sentence Each word element, Indicates an index. This represents the total sentence length. Let the target schema set be... for: ; in, Represents a collection of entity types. Represents a set of relation types. Represents a set of event types. This represents the set of event argument roles. The goal of unified information extraction is to identify all instances in the input sentence that conform to this schema set. Semantic instances; structured extraction results Represented as: ; in, Represents a collection of entities, where each entity instance is represented as... , For entity reference text, For entity type tags; Represents a set of relations, where each relation instance is represented as , and These are subject entity mentions and object entity mentions, respectively. For relation type labels; This represents the set of event trigger words, where each trigger word instance is represented as... , For trigger words to mention text, For event type labels; Represents the set of event arguments, where each event argument instance is represented as , For the argument to refer to the text, This refers to the labeling of argument roles. Unlike traditional generative unified information extraction methods, this method requires the model to not only generate the final structured result. It also generates the intermediate thought chain reasoning process: ; in, Indicates the reasoning path of the thought process. Indicates the first One reasoning step, This represents the total number of reasoning steps. This thought process explains how the model obtains structured extraction results based on the input sentence, contextual evidence, and schema constraints. Therefore, the model's final target output is... Defined as: ; Mind Chain Dataset Synthesis: By automatically constructing reasoning supervision data oriented towards unified information extraction, this addresses the problem that existing information extraction data only contains inputs and labels and lacks intermediate reasoning paths, ultimately constructing the Mind Chain training dataset. Each training sample is in the following format: The structured results are represented in JSON format. Specifically, the steps include: (1) Data sampling and filtering: First, four types of samples are obtained from the existing information data resources: named entity recognition, relation extraction, event detection and event argument extraction. In order to reduce the negative impact of long-tail label distribution on model training, a type-aware sampling method is adopted to balance the amount of data of different semantic types. At the same time, a rule-based filtering method is used to remove invalid samples, including samples with too short sentence length, insufficient context information, and inability to support the complete reasoning process.

[0026] (2) Multipath sampling for a given input sentence and schema set A prompt template containing reasoning requirements is constructed, and multiple sets of candidate outputs are generated using a teacher's large language model. The teacher's large language model is a pre-trained large language model used for batch generation of multiple sets of candidate reasoning samples. The candidate set is represented as follows: ; in, Indicates the first Candidate structured extraction results corresponding to each reasoning step This indicates the number of sampling paths. By sampling along multiple paths, the random bias introduced by a single decoding path is reduced, providing diverse candidate supervision signals for subsequent screening.

[0027] (3) Consistency filtering is performed based on whether the original data has standard labels, using two strategies: For data with supervised sources, candidate results are directly filtered. The results are compared with standard structured results, and only candidate samples whose structured results perfectly match the standard annotations are retained. If multiple perfectly matching candidates exist, the sample with the shortest inference path is selected; if no matching candidate exists, the sample is discarded. For unsupervised data, a consensus-based consistency filtering strategy is adopted: each candidate result... Using each candidate result as an anchor point, calculate the F1 score with all other candidate results, and use the average F1 score as the consistency score for that candidate result. Finally, retain the candidate result with the highest consistency score.

[0028] (4) The quality assessment of the thinking chain introduces an auxiliary evaluation big language model as an evaluator. The auxiliary evaluation big language model serves as an independent big language model for judging reasoning logic and the matching degree between reasoning and results. It scores the quality of samples after consistency filtering. The evaluation dimensions include two aspects: Reasoning-outcome consistency: Determine whether the content of the thought chain can support the final structured extraction result, and whether the reasoning process corresponds to the conclusion.

[0029] Logical rationality: Determine whether the reasoning process is coherent, conforms to the semantics of the text and the schema definition, and whether there are obvious logical errors.

[0030] The sample is retained in the final training dataset of the Mind Chain only when the scores of both dimensions reach the preset threshold.

[0031] Through the above steps, a high-quality training dataset with "textual evidence-reasoning path-structured results" aligned is obtained, enabling the model to learn an interpretable reasoning-based information extraction process, rather than just learning a direct mapping from input to output.

[0032] S2. Supervised Fine-tuning of the Thinking Chain: The large language model is fine-tuned using the thinking chain training dataset under supervision, enabling the model to generate inference paths first and then structured results. After training, an initial inference strategy model is obtained. For each training sample The input sentence and the target schema are concatenated to form the model input sample. The thought process reasoning path and structured extraction results are combined to form the target output sequence. The training objective is to maximize the given input. Generate target sequence in time The conditional probability is equivalent to minimizing the negative log-likelihood loss, and the loss function is expressed as: ; in, For mathematical expectation symbol, For trainable parameters of a large language model, For the first in the target sequence Each word element, For the first All lexical units preceding the given lexical unit, Given the input and the preceding text, this loss function is used to generate the conditional probability of the current word. Through optimization of this loss function, the model learns a behavioral pattern that first outputs the stepwise reasoning process and then outputs the structured extraction result under the constraints of the input text and schema.

[0033] S3. Inference Enhancement Optimization: Supervised fine-tuning alone is insufficient to further improve inference quality, extraction accuracy, and output format stability. Therefore, this invention introduces a reinforcement learning-based inference enhancement mechanism, using the GRPO algorithm to further optimize the initial inference strategy model. The final information extraction model is obtained by training a dedicated large language model for inference-based extraction through two rounds of supervised fine-tuning and GRPO. The training data in this stage is sampled from an independent dataset that has never participated in the synthesis of thought chains, thus avoiding data leakage. Let the input sample be... Candidate output sampling for each input sample From the old strategy model The middle sampling obtained a total A set of candidate outputs, denoted as , Indicates the first There are 10 candidate outputs, each containing a thought chain reasoning path and a structured extraction result, i.e. .

[0034] A hybrid reward function is designed to quantitatively evaluate the quality of each candidate output. The total reward is a weighted average of the soft F1 reward and the format validity reward, expressed as follows: ; in, For the first Total reward for each candidate output The soft F1 reward measures the degree of overlap between the extracted results and the standard results. Unlike the strict boundary matching F1, the soft F1 reward allows partial rewards to be given to predictions that overlap with the standard annotations but whose boundaries are not completely consistent, providing a smoother optimization signal and alleviating the problem of reward sparsity. The format validity reward is used to determine whether the structured results generated by the model meet the preset JSON output format. Only outputs with valid format, complete fields, and normal parsing can receive a positive reward. and These are the weighting coefficients for the soft F1 reward and the format validity reward, respectively, used to balance the weights of the two optimization objectives.

[0035] The GRPO policy update, based on the GRPO algorithm, first calculates the average reward of all candidate outputs within the same group. As a baseline: ; Then calculate the dominance value for each candidate output. : ; The current policy model is updated based on the advantage value, with the optimization objective being: ; in, The weights of the KL regularization terms. This is the KL divergence regularization term, used to constrain the distribution difference between the current policy model and the reference model, preventing the model from deviating excessively from the inference behavior learned during the supervised fine-tuning stage. The KL divergence is defined as: ; in, This represents the reference model, whose parameters are initialized to the initial inference policy model parameters after fine-tuning the thought chain, and remain fixed throughout the reinforcement learning phase. Policy update term. This is implemented using a pruning mechanism, defined as follows: ; in, The pruning threshold is used to limit the magnitude of policy updates in a single step, ensuring training stability. This indicates the comparison between the current policy and the old policy for candidate outputs. The probability ratio is defined as: ; Through the reinforcement learning process described above, the model retains its original ability to generate thought chains and further learns reasoning strategies that are more conducive to information extraction, while improving the accuracy and format validity of the final structured results.

[0036] S4. Reasoning and Output Parsing: In the reasoning application stage, the text to be extracted and the target schema are combined to form the input sample. The input is the final information extraction model after fine-tuning the thought chain and enhancing the reasoning. The model is generated according to a fixed "reasoning first, result later" order, specifically: first, the thought chain reasoning process is generated. The system performs step-by-step analysis on candidate entities, candidate relations, event trigger words, and event arguments in the text, and performs matching and verification with the schema definition; then it generates structured extraction results that conform to the preset JSON format. The model output is parsed to extract the structured results, which are then transformed into entity sets, relation sets, event trigger word sets, and event argument sets for downstream use. Simultaneously, the thought chain portion can be retained as an interpretation basis for the extracted results, supporting manual review and error correction.

[0037] Experimental verification: Qwen3-8B was used as the base large language model. The evaluation dataset covered the CrossNER multi-domain entity set, the FewRel relation set, and the WikiEvents event set. The evaluation metric was Micro-F1. The average Micro-F1 score for cross-domain named entity recognition is shown in Table 1.

[0038] Table 1: Experimental Results of Cross-Domain Named Entity Recognition (Micro-F1)

[0039] The average F1 values ​​for relation extraction and event extraction are shown in Table 2.

[0040] Table 2: Average F1 of relation extraction and event extraction

[0041] As shown in Table 1, the present invention achieves an average F1 score of 71.7 in cross-domain named entity recognition, which is better than GPT-OSS-20B's 67.6, Qwen3-8B's 64.8 and KnowCoder-7B's 60.1. This indicates that the present invention can improve the model's generalization ability in unseen domains through explicit reasoning enhancement.

[0042] As shown in Table 2, this invention achieves an average F1 score of 42.7 for relation extraction, and 35.4 and 29.7 for event detection and event argument extraction, respectively, all of which outperform existing non-reasoning information extraction models and general reasoning models. This result demonstrates that general reasoning ability cannot be directly converted into stable information extraction ability. Only by combining the reasoning process with schema constraints, extraction accuracy, and the effectiveness of structured output during training can the performance of complex information extraction tasks be better improved.

[0043] This invention also improves the interpretability and structural validity of the results. Before outputting the final JSON structured result, the model first generates a thought chain reasoning path, which can provide a traceable basis for the prediction of entities, relations, and event arguments; at the same time, by constraining the output results to meet the preset structured format through format validity rewards, it reduces format errors and unparseable output, and improves the convenience of subsequent system calls and manual verification.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A unified information extraction method based on enhanced reasoning through thought chains, characterized in that, Includes the following steps: S1. Synthesis of the Thinking Chain Dataset: Based on the existing information extraction dataset, multi-path sampling, consistency filtering, and thinking chain quality assessment are performed in sequence to construct training samples containing input sentences, task schemas, thinking chain reasoning paths, and structured extraction results, forming the thinking chain training dataset. S2, Mind Chain Supervised Fine-Tuning: The input sentence and task schema are used as input to the large language model. The mind chain reasoning path and structured extraction results are concatenated to form the target output sequence. The large language model is then supervised and fine-tuned to obtain the initial reasoning strategy model. S3. Inference Enhancement Optimization: The GRPO algorithm is used to perform reinforcement learning optimization on the initial inference strategy model. By using a hybrid reward function to simultaneously constrain the extraction accuracy and the validity of the output format, the final information extraction model is obtained. S4. Reasoning Generation: Input the text to be extracted and the target schema into the final information extraction model. The final information extraction model generates the thought chain reasoning path and the structured extraction result in sequence. The output content is parsed to obtain the final information extraction content.

2. The unified information extraction method based on enhanced reasoning chain as described in claim 1, characterized in that: The structured extraction results cover four types of tasks: named entity recognition, relation extraction, event detection, and event argument extraction.

3. The unified information extraction method based on enhanced reasoning chain as described in claim 1, characterized in that, The multi-path sampling in S1 is as follows: for each input sample, a prompt template containing reasoning requirements is constructed, and multiple candidate outputs are generated using the teacher's large language model. Each candidate output contains a thought chain reasoning path and the corresponding structured extraction result.

4. The unified information extraction method based on enhanced reasoning chain as described in claim 3, characterized in that, The consistency filtering in S1 is as follows: For labeled supervised data, candidate samples whose structured results completely match the standard labels are retained. If there are multiple matching candidates, the sample with the shortest inference path is selected. For unlabeled unsupervised data, the F1 score of each candidate result is calculated with other candidate results as anchor points, and the average value is taken as the consistency score. The candidate sample with the highest consistency score is retained.

5. The unified information extraction method based on enhanced reasoning chain as described in claim 4, characterized in that, The quality assessment of the thought chain in S1 is as follows: an auxiliary evaluation large language model is used as an evaluator to score the retained candidate samples. The evaluation dimensions include reasoning-result consistency and logical rationality. Samples whose scores in both dimensions reach the preset threshold are retained as the final training data.

6. The unified information extraction method based on enhanced reasoning chain as described in claim 1, characterized in that: The training objective of supervised fine-tuning of the thought chain in S2 is to minimize the negative log-likelihood loss, and its loss function is... for: ； in, For mathematical expectation symbol, For trainable parameters of a large language model, For training the mind chain dataset, For the input sample, Output the target sequence. For the first in the target sequence Each word element, For the first All lexical units preceding the given lexical unit, Given the input and the preceding text, the conditional probability of generating the current word.

7. The unified information extraction method based on enhanced reasoning chain as described in claim 1, characterized in that, The specific process of using the GRPO algorithm for policy update in S3 is as follows: For each input sample, a set of candidate outputs is sampled from the old policy model, the difference between the reward value of each candidate output and the average reward within the group is calculated as the advantage value, and the current inference policy model is updated based on the advantage value.

8. The unified information extraction method based on enhanced reasoning chain as described in claim 7, characterized in that: The total reward of the hybrid reward function is a weighted combination of the soft F1 reward and the format validity reward, as shown below: ； in, For the first Total reward for each candidate output This is a soft F1 reward, used to measure the degree of overlap and matching between the extracted results and the standard results. This is a format validity reward used to determine whether the output conforms to a preset structured format. and These are the weighting coefficients for the soft F1 reward and the format validity reward, respectively.

9. The unified information extraction method based on enhanced reasoning chain as described in claim 7, characterized in that: The GRPO algorithm incorporates a KL divergence regularization term into its optimization objective to constrain the distribution difference between the current inference policy model and the reference model, thereby preventing the current inference policy model from deviating from the inference behavior obtained through supervised fine-tuning. The parameters of the reference model are initialized to the initial inference policy model parameters after the thought chain fine-tuning and remain fixed during the reinforcement learning phase.

10. The unified information extraction method based on enhanced reasoning chain as described in claim 1, characterized in that: The final information extraction model in S4 first generates a step-by-step reasoning process based on contextual evidence and schema constraints, and then generates a structured extraction result in JSON format. The final information extraction content obtained after parsing includes a set of entities, a set of relations, a set of event trigger words, and a set of event arguments.

Images