A multi-agent based event co-reference resolution method

Abstract

The present application belongs to the field of information technology, and particularly relates to an event co-reference resolution method based on multi-agent. The specific process is as follows: the cue words of event attribute recognition agent are spliced with the chapter text to be analyzed for event co-reference, and then the chapter text is input into the event attribute recognition agent; the event attribute recognition agent outputs event type and role according to the input content or outputs nothing; when the output of the event attribute recognition agent is nothing, the external source information integration agent is queried; after the return result of the external source information integration agent is added, the event attribute recognition agent generates the corresponding event type and role again; the event type and role are spliced with the chapter text, and then the chapter text is input into the event co-reference prediction agent; the event co-reference prediction agent realizes event co-reference resolution for each event mention according to the sequence.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a method for resolving event coreference based on multiple agents. Background Technology

[0002] Event coreference resolution is a task that clusters event mentions in text according to whether they refer to the same event. Existing event coreference resolution methods mainly focus on the consistency between event type and event role. These methods use machine learning or deep learning to extract features such as the type and role of the event mentioned, and then use classification and clustering algorithms to determine event coreference and complete the clustering.

[0003] In large language model-based methods, existing work leverages the advantages of large language models in general language understanding and generation tasks, attempting to guide them to generate text containing event attributes such as event roles, which serve as text features to assist smaller language models in subsequent tasks. Min et al. employed a two-step prompting approach, instructing GPT-4 to generate a summary description of events covering detailed information such as event roles and event time attributes. This summary text, along with the original text, was then input into RoBERTa. LARGE The model is fine-tuned to complete subsequent predictions. Wei et al. directly marked the specific locations of events and their semantic roles in the text, and used cue engineering to directly instruct the GPT-3.5 and LLaMA-2 models to generate predictions of event coreference. Zeng et al. (application publication number CN119577139 A) used large language models to provide linguistic features, and combined the reasoning results of the large language model with the classification results of the pre-trained model to complete the task of event association analysis (this task is not the event coreference resolution task).

[0004] In the task of resolving event coreference, existing work has not utilized the influence of external attributes of events on event coreference resolution.

[0005] In the task of resolving event coreference, existing technologies do not make full use of the reasoning ability of large language models. Instead, they directly prompt the model to output the result of event coreference or directly instruct the model to generate linguistic features related to the task of resolving event coreference (such as paraphrasing of events) for use in downstream classifiers.

[0006] Existing methods either directly prompt the model to output clustering results or pair events together and predict coreferences sequentially to form event coreference chains. Both approaches have the following problems: the former requires the model to directly output the final result, which is difficult and struggles to generate accurate results; the latter has high time complexity, and combining it with the thinking chain capabilities of large models will incur significant time overhead. Furthermore, in methods that mention "pairs," the model performs independent reasoning for each pair of mentions, rather than considering the global perspective of event coreferences. Summary of the Invention

[0007] In view of this, a multi-agent event coreference resolution method is invented, which can make full use of the agent's thought chain reasoning ability to improve coreference prediction performance.

[0008] The technical solution for implementing the present invention is as follows:

[0009] A multi-agent-based event coreference resolution method, the specific process of which is as follows:

[0010] The event attribute recognition agent's prompts are concatenated with the text of the passage to be analyzed for event coreference, and then input into the event attribute recognition agent; the event attribute recognition agent outputs the event type and role according to the input content, or outputs nothing;

[0011] When the output of the event attribute recognition agent is empty, it continues to output query text and asks the external information integration agent for a query. The external information integration agent organizes the query statement, combines it with the query text from the Internet, and returns the corresponding query text to the event attribute recognition agent. After appending the return result from the external information integration agent, the event attribute recognition agent generates the corresponding event type and role again, or repeats this step if the output is empty.

[0012] When the event attribute recognition agent outputs the event type and role, the event type, role, and chapter text are concatenated and input into the event coreference prediction agent. The event coreference prediction agent analyzes and searches for the coreference set to which the event mention belongs in the previously formed event coreference set for each event mention in sequence, until the last event mention, thereby achieving event coreference resolution.

[0013] Optionally, the external information integration agent of the present invention organizes the query statement, combines it with the query text from the Internet, and returns the corresponding text to the event attribute recognition agent. The specific process is as follows:

[0014] The preset prompt words of the external information integration intelligent agent system are concatenated with the query text of the event attribute recognition intelligent agent and then input into the external information integration intelligent agent.

[0015] The external information integration agent outputs query terms for initiating searches on the Internet based on the input.

[0016] Use query terms to retrieve web pages from the Internet, process the retrieved web page files, and concatenate the processed text.

[0017] The webpage text obtained by splicing above is appended to the history of input and output of the external information integration agent, and then re-entered into the external information integration agent.

[0018] The external information integration agent combines external information from web pages to answer queries from the event attribute recognition agent.

[0019] Optionally, the processing of the retrieved web page files described in this invention specifically includes: parsing the web page using the BeautifalSoup library to extract the web page text; organizing the text format and cleaning up garbled characters and special symbols; and using SimHash to remove duplicate text.

[0020] Optionally, the training process of the event attribute recognition agent of the present invention is as follows:

[0021] Training set construction: On the MAVEN-arg training corpus, the input and expected output of the event attribute recognition agent are organized using pre-set input and output format templates;

[0022] Training tasks include supervised fine-tuning of the agent's model-generated content and question optimization. Supervised fine-tuning of the agent's model-generated content guides the model to complete the correct output according to the instructions. Question optimization uses a contrastive learning method to optimize the questions asked by the agent in interaction with the agent integrating external information, thereby increasing the probability of returning valid results. The loss values ​​of the two training tasks are jointly optimized.

[0023] Optionally, in the training task, the supervised fine-tuning loss value of the agent's model-generated content is as follows: the constructed training set is input into the agent, the logits of each word are obtained at the output of the agent, and the loss value between the agent's output distribution and the real text is calculated using cross-entropy loss.

[0024] Optionally, in the training task, the question optimization loss value of this invention is: the question for which the external information integration agent successfully returns a result is denoted as y. w The question that returns [UNK] is denoted as y. l Using normalized logarithmic probabilities:

[0025]

[0026] The loss value for query term optimization is:

[0027]

[0028] Optionally, if the event coreference prediction agent of the present invention finds the coreference set to which the event mentions in the previously formed event coreference set, it adds the event to the coreference chain of this event. If it does not find the coreference set to which the event mentions belongs, it establishes a single event chain for the event and adds it to the coreference set.

[0029] Optionally, the training process of the event coreference prediction agent of the present invention is as follows: during training, step-by-step rewards and global rewards are set;

[0030] Step-by-step reward: Let the event mention in the current step be M. The model's inference conclusion is that it is co-referenced with the set of antecedents {M1, M2, ..., Mi, Mi+1, Mj} that have been assigned to the same coreference chain. Among them, M1 to Mi actually have a coreference relationship with M, while Mi+1 to Mj do not actually have a coreference relationship with M. Then the reward score for the current state is:

[0031] Global Result Reward: The agent divides event mentions into several coreference chains and scores the division as the overall score of the model's inference result. Suppose that in a text, there are actually i event mentions coreferenced with event mention m, and the agent predicts that there are j event mentions coreferenced with event mention m. Among these j event mentions, k are coreferenced with event mention m, where k ≤ i, k ≤ j. Then the score of event mention m is denoted as: That is, the global reward for the agent is the average score of all entity mentions: i m j m k m This represents the number of events mentioned in the actual, predicted, and predicted coreferences of the entity mention m that are actually coreferences of the entity mention m, while M represents the total number of mentions in the text.

[0032] Optionally, the reward function of the present invention further includes: format conformance reward, using regular expression matching to check whether the model's inference output conforms to the system prompt word requirements, if so, it is 1 point, otherwise it is 0 points.

[0033] Optionally, the external information integration agent, event attribute recognition agent, and event coreference prediction agent described in this invention adopt the open-source language model GLM-4-9B-0414.

[0034] Beneficial effects:

[0035] First, the event coreference resolution task involves multiple sub-tasks with different task types. Therefore, this method designs a multi-agent event coreference resolution system, where each agent performs its own function and collaborates to complete the task. This work designs an optimized algorithm for agents to complete their own tasks and for the collaboration between agents to adapt to each other.

[0036] Second, existing technologies fail to leverage the advantages of large-scale models' chain-of-thought reasoning capabilities in resolving event coreference tasks. In real-world examples, the coreferences between events exhibit complex semantic and logical relationships, requiring in-depth and rigorous logical reasoning. This invention introduces chain-of-thought reasoning, designs a reasoning process for the model to resolve event coreferences, and designs a reward function for step-level results. Building upon the advantages of chain-of-thought reasoning, it further improves the accuracy of model output, reduces time overhead, and guides the model to perform global clustering of event coreferences.

[0037] Third, existing work does not utilize the attribute features of external events to enhance inference for event coreference resolution tasks. As the example above illustrates, introducing external information can provide more auxiliary information for event coreference prediction, helping to improve the accuracy of event coreference prediction. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a diagram illustrating the interaction relationship between three intelligent agents. Detailed Implementation

[0040] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0042] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0043] This application provides an embodiment of a multi-agent-based event coreference resolution method, such as... Figure 1 As shown, this method includes an external information integration agent, an event attribute recognition agent, and an event coreference prediction agent. The external information integration agent receives questions related to a specified event and answers them using external information. The event attribute recognition agent identifies attributes such as event type and event role. The event coreference prediction agent uses thought chain reasoning to analyze the global event coreference relationships in the text in a step-by-step manner. The specific process of this method is as follows:

[0044] The event attribute recognition agent's prompts are concatenated with the text of the passage to be analyzed for event coreference, and then input into the event attribute recognition agent; the event attribute recognition agent outputs the event type and role according to the input content, or outputs nothing;

[0045] When the output of the event attribute recognition agent is empty, it continues to output query text and asks the external information integration agent for a query. The external information integration agent organizes the query statement, combines it with the query text from the Internet, and returns the corresponding query text to the event attribute recognition agent. After appending the return result from the external information integration agent, the event attribute recognition agent generates the corresponding event type and role again, or repeats this step if the output is empty.

[0046] When the event attribute recognition agent outputs the event type and role, the event type, role, and chapter text are concatenated and input into the event coreference prediction agent. The event coreference prediction agent analyzes and searches for the coreference set to which the event mention belongs in the previously formed event coreference set for each event mention in sequence, until the last event mention, thereby achieving event coreference resolution.

[0047] The interaction relationship between the three agents in this embodiment is as follows: Figure 1As shown: The event attribute recognition agent initiates a query to the external information integration agent. The external information integration agent organizes the query statement, combines it with query text from the internet, and returns the corresponding text to the event attribute recognition agent. The event attribute recognition agent then passes the event attribute recognition results to the event coreference prediction agent. The event coreference prediction agent combines the semantics of the discourse, the information provided by the event attribute recognition agent, its own knowledge, and reasoning ability to predict the event coreference.

[0048] The following section uses an intelligent agent as an example to explain in detail the execution process of the above method:

[0049] The external information integration agent uses the open-source language model GLM-4-9B-0414. See Appendix 1 for system prompts.

[0050] The operation flow of this intelligent agent is as follows:

[0051] (1) Concatenate the preset external information integration agent system prompt words (see Appendix 1 for system prompt words) with the query text of the event attribute recognition agent and input it into the external information integration agent;

[0052] (2) The external information integration agent outputs query terms (i.e., answers) to initiate a search on the Internet based on the input (i.e., the question);

[0053] (3) Use query terms to retrieve web pages from the Internet;

[0054] (4) Process the retrieved web page files: use the BeautifalSoup library to parse the web page and extract the main text; organize the text format and clean up garbled characters and special symbols; use SimHash to remove duplicate text; and concatenate and organize the text.

[0055] (5) Append the webpage text obtained in step 4 above to the history of input and output of the agent; re-enter the agent;

[0056] (6) The agent combines external information from the webpage to answer the query of the event attribute recognition agent. If no valid conclusion is obtained, the process is repeated 2 to 5 times until the termination. If no valid conclusion is ultimately obtained, [UNK] is returned, indicating that the result is unknown.

[0057] Appendix 1: Preset prompts for the external information integration intelligent agent system

[0058]

[0059] The event attribute recognition agent uses the open-source language model GLM-4-9B-0414. See Appendix 2 for system prompts.

[0060] The operation process of this intelligent agent is as follows:

[0061] (1) Concatenate the preset system prompt words (see Appendix 2) with the text of the chapter to be analyzed for event co-reference, and input them into the agent;

[0062] (2) The agent follows its input and outputs the types of events in the text.

[0063] (3) Based on the event type, determine the event role system according to the MAVEN-arg system, and output the role of the event type one by one: the model can generate a definite event role or output "[UNK]". If "[UNK]" is output, the query text will continue to be output to the external information integration agent to ask questions and query.

[0064] (4) After the historical inputs and outputs of the agent, add external information to integrate the return results of the agent, generate the corresponding event role again, or output [UNK] to indicate unknown.

[0065] Appendix 2: Prompt Words for Event Attribute Recognition Intelligent Agent System

[0066]

[0067] The training process for this agent is as follows:

[0068] Training set construction: On the MAVEN-arg training corpus, organize the agent's input and expected output according to the format templates in Appendix 2 and Appendix 3. If an event role is labeled in the training corpus, fill in the corresponding position of the event role in the template. Otherwise, the agent samples and generates multiple questions, filling in the question positions. Use each question to interact sequentially with the external information integration agent. Among these questions, randomly retain one question from which the external information integration agent returns a non-[UNK] result, and randomly retain one question from which the external information integration agent returns a [UNK] result, forming a comparison group.

[0069] Training Task: Optimize two aspects of the agent's capabilities: Supervised fine-tuning of model-generated content: Guiding the model to complete correct output according to instructions. Question Optimization: Using contrastive learning, optimize the agent's questioning interactions with the external information integration agent to increase the probability of returning valid results. Joint optimization is performed during training; the training input is a batch, and the two pairs of samples in the batch come from the same comparison group.

[0070] (1) Supervised fine-tuning of model-generated content: Input the training samples constructed above into the agent. At the output of the model, obtain the logits of each word, and use cross-entropy loss to calculate the loss value between the model output distribution and the real text. Among them, it is necessary to mask the loss part of the agent's returned result, which integrates system prompts, input text, questions, and external information, and only retain the loss of the output format template framework and the part that can be directly filled according to the original training corpus.

[0071] (2) Question optimization: The question that the external information integration agent successfully returns a result is denoted as y. w The question that returns [UNK] is denoted as y. l Using normalized logarithmic probabilities:

[0072]

[0073] The loss function for query term optimization is:

[0074]

[0075] Appendix 3: Event Attribute Recognition Agent Output Template

[0076]

[0077] Event coreference prediction agent: Based on the text of the passage and event attribute analysis information from the event attribute recognition agent, it globally predicts event coreference relationships. The input text for this agent is a concatenation of system prompts (see Appendix 4), the text of the passage, and the type and role of the event (from the output of the event attribute recognition agent). The agent sequentially analyzes and searches for the coreference set to which each event mention belongs in the previously formed coreference set, until the last event mention.

[0078] The coreference prediction inference process of this agent is as follows: Iterate through all events in the chapter. For each event, analyze whether it refers to the same event as an event (the antecedent of the event) in the previously obtained coreference chain, and give a conclusion. If the conclusion is yes, add this event to the coreference chain of this event; otherwise, the event is in a single event chain. The pseudocode description is as follows:

[0079]

[0080] The training process for this agent is described below:

[0081] This embodiment can use the PPO reinforcement learning algorithm to complete the model training. The PPO algorithm involves 3 models and 1 reward function:

[0082] Strategy Model: The GLM-4-9B-0414 model to be optimized, and strategy optimization is performed.

[0083] Value model: Use the GLM-4-9B-0414 model + value head initialization to estimate the value of the state.

[0084] Reference model: GLM-4-9B-0414 model, KL distance constraint for near-end policy optimization.

[0085] Reward Function: Provides step-level and global reward scores for the model's inference results. In the model's complete inference, the coreference prediction inference for each event mention is a sub-step. The complete inference is broken down into inference steps, and the step-by-step reward function provides a reward score at the last word output position of each step, with reward scores set to 0.5 at other positions. At the end of the model output, a weighted score is applied to the global inference result and the inference format.

[0086] The reward function is designed as follows:

[0087] Step-by-step rewards:

[0088] Let the event mention in the current step be M. The model's inference conclusion is that it is co-referenced with the set of antecedents {M1, M2, ..., Mi, Mi+1, Mj} that have been assigned to the same coreference chain. Specifically, M1 to Mi are actually co-referenced with M, while Mi+1 to Mj are not actually co-referenced with M. Therefore, the reward score for the current state is:

[0089] Global result reward:

[0090] The agent divides event mentions into several coreference chains and scores the division as the overall score of the model's inference result. Suppose that in a text, there are actually i event mentions that are coreferenced with event mention m, and the agent predicts that there are j event mentions that are coreferenced with event mention m. Among these j event mentions, k are coreferenced with event mention m, where k ≤ i, k ≤ j. Then, the score of event mention m is denoted as: That is, the global reward for the agent is the average score of all entity mentions: i m j m k m This represents the number of events mentioned in the actual, predicted, and predicted coreferences of the entity mention m that are actually coreferences of the entity mention m, while M represents the total number of mentions in the text.

[0091] Formatting follows the reward:

[0092] The model's inference output is checked using regular expression matching to see if it follows the system's prompts. If it does, it scores 1 point; otherwise, it scores 0 points.

[0093] This embodiment can also employ optimization algorithms such as GRPO (Group Relative Policy Optimization) and DAPO (Dynamic sAmpling Policy Optimization).

[0094] Appendix 4: Event Coreference Prediction Agent System Prompt Words

[0095]

[0096] This application has the following characteristics:

[0097] First, a multi-agent system is used to complete the event coreference resolution task: This method designs a multi-agent system consisting of an external information integration agent, an event attribute recognition agent, and an event coreference prediction agent, which work together to complete the event coreference resolution task.

[0098] Secondly, based on the multi-agent system, this method optimizes each agent for its own task. For example, supervised fine-tuning is used for the event attribute recognition agent, a global reasoning process for event coreference prediction is designed for the event coreference prediction agent, and a step-level reward reinforcement learning fine-tuning method is designed specifically for the characteristics of the referential resolution task. This method also optimizes the adaptation between agents. For example, for the event attribute recognition agent, a contrastive learning method is used to enhance its ability to adapt to agents integrating external information.

[0099] Finally, considering the characteristics of the event coreference resolution task, a large-scale model-based thought chain reasoning process for event coreference resolution was designed. The model is guided to start with each event mention, perform coreference discrimination with its antecedents and previously inferred coreference chains, and gradually complete clustering. The PPO reinforcement learning algorithm was used for optimization. In the design of the reward function, rewards were given at the step-level for each sub-step of the model's reasoning, taking into account the characteristics of the event coreference resolution task.

[0100] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for resolving event coreference based on multi-agent systems, characterized in that, The specific process is as follows: The event attribute recognition agent's prompts are concatenated with the text of the passage to be analyzed for event coreference, and then input into the event attribute recognition agent; the event attribute recognition agent outputs the event type and role according to the input content, or outputs nothing; When the output of the event attribute recognition agent is empty, it continues to output query text and asks the external information integration agent for query; the external information integration agent organizes the query statement, combines it with the query text from the Internet, and returns the corresponding query text to the event attribute recognition agent. After adding the return results of the external information integration agent, the event attribute recognition agent generates the corresponding event type and role again, or if the output is empty, repeats this step. When the event attribute recognition agent outputs the event type and role, the event type, role, and chapter text are concatenated and input into the event coreference prediction agent. The event coreference prediction agent analyzes and searches for the coreference set to which the event mention belongs in the previously formed event coreference set for each event mention in sequence, until the last event mention, thereby achieving event coreference resolution.

2. The event coreference resolution method based on multi-agent systems according to claim 1, characterized in that, The external information integration agent organizes the query statement, combines it with the query text from the Internet, and returns the corresponding text to the event attribute recognition agent. The specific process is as follows: The preset prompt words of the external information integration intelligent agent system are concatenated with the query text of the event attribute recognition intelligent agent and then input into the external information integration intelligent agent. The external information integration agent outputs query terms for initiating searches on the Internet based on the input. Use query terms to retrieve web pages from the Internet, process the retrieved web page files, and concatenate the processed text. The webpage text obtained by splicing above is appended to the history of input and output of the external information integration agent, and then re-entered into the external information integration agent. The external information integration agent combines external information from web pages to answer queries from the event attribute recognition agent.

3. The event coreference resolution method based on multi-agent systems according to claim 2, characterized in that, The specific steps for processing the retrieved webpage files are as follows: using the BeautifalSoup library to parse the webpage and extract the main text; organizing the text format and cleaning up garbled characters and special symbols; and using SimHash to remove duplicate text.

4. The event coreference resolution method based on multi-agent systems according to claim 1, characterized in that, The training process of the event attribute recognition agent is as follows: Training set construction: On the MAVEN-arg training corpus, the input and expected output of the event attribute recognition agent are organized using pre-set input and output format templates; Training tasks include supervised fine-tuning of the agent's model-generated content and question optimization. Supervised fine-tuning of the agent's model-generated content guides the model to complete the correct output according to the instructions. Question optimization uses a contrastive learning method to optimize the questions asked by the agent in interaction with the agent integrating external information, thereby increasing the probability of returning valid results. The loss values ​​of the two training tasks are jointly optimized.

5. The event coreference resolution method based on multi-agent systems according to claim 4, characterized in that, In the training task, the supervised fine-tuning loss value of the agent's model-generated content is as follows: the constructed training set is input into the agent, and the logits of each word are obtained at the output of the agent. The loss value between the agent's output distribution and the real text is calculated using cross-entropy loss.

6. The event coreference resolution method based on multi-agent systems according to claim 4, characterized in that, In the training task, the question optimization loss value is: a question in which the external information integration agent successfully returns a result is denoted as: Questions that return [UNK] are denoted as Using normalized logarithmic probabilities: The loss value for query term optimization is: 。 7. The event coreference resolution method based on multi-agent systems according to claim 1, characterized in that, If the event coreference prediction agent finds the coreference set to which the event mentions in the previously formed event coreference set, it adds the event to the coreference chain of this event. If it does not find the coreference set to which the event mentions belongs, it establishes a single event chain for the event and adds it to the coreference set.

8. The event coreference resolution method based on multi-agent systems according to claim 7, characterized in that, The training process of the event co-reference prediction agent is as follows: during training, step-by-step rewards and global rewards are set; Step-by-step reward: Let the event mention in the current step be M. The model's inference conclusion is that it is co-referenced with the set of antecedents {M1, M2, …, Mi, Mi+1, Mj} that have been assigned to the same coreference chain. Among them, M1~Mi actually have a coreference relationship with M, while Mi+1~Mj do not actually have a coreference relationship with M. Then the reward score for the current state is: ; Global Result Reward: The agent divides event mentions into several coreference chains and scores the overall score of the model's inference result based on the scores of these chains. Suppose that in a text, there are actually i event mentions that are coreferenced with event mention m. The agent predicts that there are j event mentions that are coreferenced with event mention m, and among these j event mentions, k are coreferenced with event mention m. , The score for mentioning event m is then denoted as: That is, the global reward for the agent is the average score of all entity mentions: , This represents the number of event mentions that are actually co-referenced with event mention m, the number of event mentions that are predicted to be co-referenced with event mention m, and the number of event mentions that are actually co-referenced with event mention m in the predicted co-reference. Indicates the total number mentioned in the text; Reward function: Provides step-level and global result rewards for the model's inference results.

9. The event coreference resolution method based on multi-agent systems according to claim 8, characterized in that, The reward function also includes: format compliance reward, using regular expression matching to check whether the model's inference output follows the system prompt word requirements; if so, it is 1 point, otherwise it is 0 points.

10. The event coreference resolution method based on multi-agent systems according to claim 1, characterized in that, The external information integration agent, event attribute recognition agent, and event coreference prediction agent adopt the open-source language model GLM-4-9B-0414.

Images