Event relation extraction method and device, electronic equipment and storage medium

Abstract

The application provides an event relation extraction method and device, electronic equipment and storage medium, and relates to the technical field of information processing. The event relation extraction method comprises the following steps: obtaining a to-be-analyzed document and a to-be-analyzed event pair; inputting the to-be-analyzed document and the to-be-analyzed event pair into a relation acquisition model to obtain a target event relation, wherein the target event relation is used to represent the event relation between the to-be-analyzed event pair, the target event relation is obtained by integrating the event relations predicted by multiple agents in the relation acquisition model, and the multiple agents respectively predict the event relations by simulating the interaction of a debate competition. The method can effectively improve the performance of a large model in an event relation extraction task and guarantee the accuracy of the event relation extraction result.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to an event relationship extraction method, apparatus, electronic device, and storage medium. Background Technology

[0002] Event relationship extraction refers to extracting causal, temporal, and sub-event relationships between events from a given document. A document contains multiple events; by extracting the relationships between these events, the development of the events can be effectively understood, enhancing the comprehension of the document content. In recent years, Large Language Models (LLMs) have achieved remarkable performance on various tasks in natural language processing, demonstrating powerful understanding and generation capabilities. However, LLMs still suffer from the illusion problem, generating seemingly plausible but ultimately flawed results, limiting their application. Event knowledge, which can clarify the relationships between document content, can enhance the reasoning ability of LLMs; therefore, in the era of LLMs, event relationship extraction has become even more crucial.

[0003] Current methods for extracting event relationships based on large models primarily employ a single large model, treating it as an agent that progressively extracts event relationships. However, validation has shown that a single large model is not only overconfident in its output but also extremely stubborn, which in turn affects model performance and the accuracy of the output event relationship results. Summary of the Invention

[0004] This invention provides an event relationship extraction method, apparatus, electronic device, and storage medium to solve the problem of poor performance of a single large model in event relationship extraction in the prior art, effectively improve the performance of large models in event relationship extraction tasks, and ensure the accuracy of event relationship extraction results.

[0005] This invention provides a method for extracting event relationships, comprising the following steps: Obtain the document to be analyzed and the event pair to be analyzed, wherein the event pair to be analyzed is a pair of events in the document to be analyzed; The document to be analyzed and the event pair to be analyzed are input into the relationship acquisition model to obtain the target event relationship. The target event relationship is used to characterize the event relationship between the event pairs to be analyzed. The target event relationship is obtained by predicting and integrating the event relationships by multiple agents in the relationship acquisition model. The prediction of the event relationships by the multiple agents is obtained by simulating a debate interaction by the multiple agents. Each of the multiple agents is used to represent an instance of a large model.

[0006] According to the event relationship extraction method provided by the present invention, the plurality of agents include a host agent, a positive agent, a negative agent, and an audience agent. The step of inputting the document to be analyzed and the event pairs to be analyzed into a relationship acquisition model to obtain the target event relationship includes: Based on the document to be analyzed and the event pair to be analyzed, obtain the positive debate viewpoints corresponding to the positive agent and the negative debate viewpoints corresponding to the negative agent, wherein the positive debate viewpoints and the negative debate viewpoints are different; Based on the affirmative and negative argument points, the interaction process between different participants in the debate is simulated by the host agent, the affirmative agent, the negative agent, and the audience agent to obtain the event relationships predicted by each of the multiple agents. The target event relationship is obtained by integrating the event relationships predicted by the multiple agents.

[0007] According to an event relationship extraction method provided by the present invention, the step of obtaining the affirmative debate viewpoints corresponding to the affirmative agent and the negative debate viewpoints corresponding to the negative agent based on the document to be analyzed and the event pair to be analyzed includes: A summary is generated for the event pairs to be analyzed, and the event relationships are predicted using the summary to obtain the affirmative side's argument. Sentences related to the event to be analyzed are extracted from the document to be analyzed to obtain a sentence set, and the event relationship is predicted using the sentence set to obtain the opposing argument.

[0008] According to the event relationship extraction method provided by the present invention, the interaction process between different participants in the debate includes an opening stage, a statement stage, a rebuttal stage, a free discussion stage, and a summary stage. Based on the affirmative and negative debate viewpoints, the interaction process between different participants in the debate is simulated by the moderator agent, the affirmative agent, the negative agent, and the audience agent to obtain the event relationships predicted by each of the multiple agents, including: In the opening phase, the host AI introduces the debate rules, process, theme, roles of the affirmative and negative AI, documents to be analyzed, event pairs to be analyzed, the affirmative debate viewpoints corresponding to the affirmative AI, and the negative debate viewpoints corresponding to the negative AI. During the opening statement phase, the first debater of the affirmative team and the first debater of the negative team speak in turn, stating their respective debate viewpoints and problem-solving strategies. The problem-solving strategies are used to characterize the strategies adopted to predict the event relationships between the event pairs to be analyzed. During the rebuttal phase, the affirmative and negative intelligent agents take turns questioning each other's arguments and problem-solving approaches. During the free discussion phase, the affirmative and negative intelligent agents clarified their respective argumentative viewpoints. In the summary phase, the first speaker of the affirmative team and the first speaker of the negative team predict the first event relationship and the second event relationship based on all the information of the interaction process, respectively. The audience agent predicts the third event relationship based on all the information of the interaction process, the first event relationship, and the second event relationship. The host agent predicts the fourth event relationship based on the first event relationship, the second event relationship, and the third event relationship.

[0009] According to an event relationship extraction method provided by the present invention, in the rebuttal phase, the affirmative and negative intelligent agents take turns questioning each other's argumentative viewpoints and problem-solving approaches, including: The second debater of the affirmative team raises questions about the shortcomings of the debate viewpoints and problem-solving approaches put forward by the first debater of the negative team. The second debater of the opposing team responds to the questions raised by the second debater of the affirmative team, while the second debater of the opposing team also raises questions about the shortcomings of the second debater of the affirmative team's argumentative viewpoints and problem-solving approaches. The third debater of the opposing team responds to the questions raised by the second debater of the affirmative team, while the third debater of the opposing team also raises questions about the shortcomings of the third debater of the affirmative team's argumentation and problem-solving approaches. The affirmative and negative agents engage in question-and-answer sessions in this manner until all agents in both the affirmative and negative agents participate in the question-and-answer process.

[0010] According to an event relationship extraction method provided by the present invention, during the free discussion phase, the affirmative and negative agents respectively clarify their respective argumentative viewpoints, including: Based on the previous interaction information, the affirmative agent regenerates the summary of the event pair to be analyzed and reiterates its own argument and problem-solving approach. The negative agent re-extracts the set of sentences related to the event pair to be analyzed from the document to be analyzed and reiterates its own argument and problem-solving approach. The affirmative and negative AI agents refute each other's arguments, with a maximum of four rounds of rebuttal.

[0011] According to an event relationship extraction method provided by the present invention, before obtaining the document to be analyzed and the event pair to be analyzed, the method further includes: The relationship acquisition model is configured with multiple agents, each with a different stance. These agents include a pro-side agent, a con-side agent, a host agent, and an audience agent. The rules for the debate process in the relationship acquisition model are set, and the debate includes an opening stage, a debate stage, and a summary stage. The debate stage includes an opening statement stage, a rebuttal stage, and a free discussion stage. The relationship acquisition model sets up debate rules, which are used to instruct the affirmative and negative agents to predict event relationships, adopt debate methods, and take debate steps based on the document to be analyzed, the event pairs to be analyzed, and the debate methods to be adopted before the debate.

[0012] The present invention also provides an event relationship extraction device, comprising the following modules: The information acquisition module is used to acquire the document to be analyzed and the event pair to be analyzed, wherein the event pair to be analyzed is a pair of events in the document to be analyzed. The relation extraction module is used to input the document to be analyzed and the event pair to be analyzed into the relation acquisition model to obtain the target event relation. The target event relation is used to characterize the event relationship between the event pairs to be analyzed. The target event relation is obtained by predicting and integrating the event relations by multiple agents in the relation acquisition model. The prediction of the event relations by the multiple agents is obtained by simulating the interaction of a debate competition by the multiple agents. Each agent in the multiple agents is used to represent an instance of a large model.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the event relationship extraction method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the event relationship extraction method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the event relationship extraction method as described above.

[0016] The event relationship extraction method, apparatus, electronic device, and storage medium provided by this invention, for the event relationship extraction task, enhance the understanding of event relationships of a large model by organizing multi-agent debates based on the relationship acquisition model of a large model, thereby improving the accuracy of the large model in event relationship extraction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the event relationship extraction method provided by the present invention.

[0019] Figure 2 This is a flowchart illustrating the method for obtaining target event relationships provided by the present invention.

[0020] Figure 3 This is a flowchart illustrating the method for obtaining argumentative viewpoints provided by the present invention.

[0021] Figure 4 This is a flowchart illustrating the method for obtaining the event relationships predicted by multiple agents, as provided by the present invention.

[0022] Figure 5 This is a flowchart illustrating the method for constructing a relationship acquisition model provided by the present invention.

[0023] Figure 6 This is a schematic diagram of the overall architecture of the event relationship extraction system provided by the present invention.

[0024] Figure 7 This is a schematic diagram of the event relationship extraction device provided by the present invention.

[0025] Figure 8 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] Event relationship extraction refers to extracting causal, temporal, and sub-event relationships between events from a given document. A document contains multiple events; by extracting the relationships between these events, the development of the events can be effectively understood, enhancing the comprehension of the document content. In recent years, Large Language Models (LLMs) have achieved remarkable performance on various natural language processing tasks, demonstrating powerful understanding and generation capabilities. However, LLMs still suffer from the illusion problem, generating seemingly plausible but ultimately flawed results, limiting their application. Event knowledge, which can clarify the relationships between document content, can enhance the reasoning ability of LLMs; therefore, in the era of LLMs, event relationship extraction has become even more crucial.

[0028] Current research has begun to explore the performance of large models in event relationship extraction tasks, mainly employing thought chain and self-reflection methods. The thought chain-based event relationship extraction method allows the large model to output intermediate processes step by step until it outputs the relationship category. The self-reflection-based method, based on the thought chain, allows the large model to iteratively check its output results until the maximum number of iterations is reached, or until the large model accepts the output result.

[0029] Current methods for extracting event relationships based on large models primarily employ a single large model, treating it as an agent that progressively extracts event relationships. However, validation has shown that a single large model not only becomes overconfident in its output but also extremely stubborn, which in turn affects model performance and the accuracy of the output event relationship results.

[0030] In view of this, embodiments of the present invention provide an event relationship extraction method, which involves acquiring a document to be analyzed and a pair of events to be analyzed; inputting the document to be analyzed and the pair of events to be analyzed into a relationship acquisition model to obtain a target event relationship, which is used to characterize the event relationship between the pair of events to be analyzed. The target event relationship is obtained by integrating the event relationships predicted by multiple agents in the relationship acquisition model, and the predictions of the event relationships by the multiple agents are obtained by simulating debate interactions. This method can effectively improve the performance of large models in event relationship extraction tasks and ensure the accuracy of the event relationship extraction results.

[0031] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0032] Figure 1This is a flowchart illustrating the event relationship extraction method provided by the present invention. The event relationship extraction method can be applied to electronic devices, which can be various types of devices with information processing capabilities. For example, the electronic device may include a personal computer, laptop computer, PDA, or server; the electronic device may also be a mobile terminal, such as a mobile phone, in-vehicle computer, tablet computer, or projector. Figure 1 As shown, the method may include the following steps 101 to 102: Step 101: Obtain the document to be analyzed and the event pair to be analyzed, wherein the event pair to be analyzed is a pair of events in the document to be analyzed.

[0033] It should be noted that the document to be analyzed and the event pair to be analyzed can be obtained through user input or through transmission from other devices. This invention does not limit the method of obtaining the document to be analyzed and the event pair to be analyzed.

[0034] Step 102: Input the document to be analyzed and the event pair to be analyzed into the relationship acquisition model to obtain the target event relationship. The target event relationship is used to characterize the event relationship between the event pairs to be analyzed. The target event relationship is obtained by predicting and integrating the event relationships by multiple agents in the relationship acquisition model. The prediction of the event relationships by the multiple agents is obtained by simulating a debate interaction by the multiple agents. Each agent in the multiple agents is used to represent an instance of a large model.

[0035] It should be noted that the relationship acquisition model includes multiple agents. The event relationships predicted by these agents are obtained through simulated debate interactions. These predicted event relationships are then integrated to obtain the target event relationship, i.e., the event relationship between the event pairs to be analyzed. This event relationship extraction method based on a large model and multiple agents can effectively improve the performance of large models in event relationship extraction tasks and enhance their reasoning capabilities.

[0036] It is understandable that this invention uses a given article or event. and In this debate, different agents analyze and discuss to obtain the final event relationships. For the event relationship extraction task, a relationship acquisition model based on a large model is constructed. A multi-agent debate method is employed, where a large model agent organizes a debate to enhance the large model's understanding of event relationships, thereby improving the accuracy of the large model in event relationship extraction.

[0037] For example, the present invention obtains a given document ,in Document The first in One word. ,in and They represent the first and The event trigger word, and . This is a predefined set of event relationship types. Event trigger words are words in a document that can trigger an event. In event-related tasks, event trigger words are generally used to represent events, so the term "event" will be used directly in the following description. Assume... For a large-scale multi-agent event prediction system, i.e., a relation acquisition model, given an event... and Event prediction system It needs to be based on the document content. To predict events and The relationship between them.

[0038] Figure 2 This is a flowchart illustrating the method for obtaining target event relationships provided by the present invention. For example... Figure 2 As shown, the multiple intelligent agents include a host intelligent agent, a positive intelligent agent, a negative intelligent agent, and an audience intelligent agent. The step of obtaining the target event relationship by inputting the document to be analyzed and the event pair to be analyzed into an input relationship acquisition model may include: Step 201: Based on the document to be analyzed and the event pair to be analyzed, obtain the positive debate viewpoints corresponding to the positive agent and the negative debate viewpoints corresponding to the negative agent, wherein the positive debate viewpoints and the negative debate viewpoints are different.

[0039] It should be noted that there are many ways to obtain the affirmative debate viewpoints corresponding to the affirmative agent and the negative debate viewpoints corresponding to the negative agent. For example, it can be obtained by analyzing the event pair to be analyzed and then combining it with the document to be analyzed, or by compressing the document to be analyzed and then combining it with the event pair to be analyzed. This invention does not limit the method of obtaining the affirmative debate viewpoints corresponding to the affirmative agent and the negative debate viewpoints corresponding to the negative agent based on the document to be analyzed and the event pair to be analyzed.

[0040] It is clear that the viewpoints of the affirmative and negative sides differ. By establishing affirmative and negative viewpoints with different stances, the accuracy of the obtained target event relationships can be further improved.

[0041] Step 202: Based on the affirmative and negative debate viewpoints, the interaction process between different participants in the debate is simulated by the host agent, the affirmative agent, the negative agent, and the audience agent to obtain the event relationships predicted by each of the multiple agents.

[0042] It should be noted that, based on the affirmative and negative argumentative viewpoints, the interaction process between different participants in the simulated debate, through the host AI, the affirmative AI, the negative AI, and the audience AI, can follow the process of a real courtroom debate or the process of a regular debate. This invention does not limit the method by which the interaction process between different participants in the simulated debate, based on the affirmative and negative argumentative viewpoints, is conducted through the host AI, the affirmative AI, the negative AI, and the audience AI.

[0043] In the debate process, the affirmative and negative agents can engage in multiple rounds of interaction based on their respective viewpoints, gradually refining and generating their own answers through discussion. The debate process aims to enable participants in different types of debates to present their respective results regarding the relationship between events through discussions between the affirmative and negative agents.

[0044] Step 203: Integrate the event relationships predicted by the multiple agents to obtain the target event relationship.

[0045] It should be noted that after obtaining the event relationships predicted by multiple agents, the event relationship extraction results of the agents are integrated. The goal of this part is to fuse the prediction results of different types of agents to generate a more comprehensive and accurate final prediction result. After the interaction process ends, each type of agent, including the pro-party agent, the con-party agent, the facilitator agent, and the audience agent, outputs their respective event relationship prediction results. Next, these four prediction results are used as a reference to generate the final event relationships using a large model.

[0046] Understandably, to address the issue of agent role diversity, this invention categorizes agents into four types: facilitator agent, affirmative agent, negative agent, and audience agent, exploring the different participants' understanding of event relationships and their impact on the final result. Furthermore, this invention generates affirmative debate viewpoints for the affirmative agent and negative debate viewpoints for the negative agent, then simulates the interaction process between different participants in a debate using multiple agents to obtain the event relationships predicted by each agent. Finally, these predictions are integrated to obtain the target event relationship. In other words, for the task of event relationship extraction, a multi-agent debate method based on a large model is constructed, enhancing the large model's ability in event relationship extraction and improving the accuracy of the extracted results.

[0047] Figure 3 This is a flowchart illustrating the method for obtaining argumentative viewpoints provided by the present invention. For example... Figure 3 As shown, obtaining the affirmative debate viewpoints of the affirmative agent and the negative debate viewpoints of the negative agent based on the document to be analyzed and the event pair to be analyzed may include: Step 301: Generate a summary for the event pair to be analyzed, and use the summary to predict the event relationships to obtain the affirmative argument; Step 302: Extract sentences related to the event to be analyzed from the document to be analyzed to obtain a sentence set, and use the sentence set to predict the event relationship to obtain the opposing argument.

[0048] It should be noted that a key challenge in event relationship extraction is long-distance dependency, meaning that two events may be distributed across different sentences in a document, and modeling the dependencies between events is a difficult problem. This invention uses document compression to reduce the distance between events.

[0049] This invention proposes two document compression strategies—generate first, then extract—as arguments for and against: generation first, then extraction; and selection first, then extraction. Generation first, then extraction refers to generating summaries for the event pairs to be analyzed and using these summaries to predict event relationships. Selection first, then extraction refers to extracting a set of sentences related to the event pairs from the document to be analyzed and using this set of sentences to predict event relationships.

[0050] Understandably, during the debate among intelligent agents, the pro and con agents hold different viewpoints. That is, for the problem of multi-agent interaction, this invention considers the adversarial and cooperative relationships between agents, thereby effectively improving the agents' understanding of the content of the document to be analyzed.

[0051] Figure 4 This is a flowchart illustrating the method for obtaining the event relationships predicted by multiple agents, as provided by this invention. Figure 4 As shown, the interaction process between different participants in the debate includes the opening stage, the opening statement stage, the rebuttal stage, the free discussion stage, and the summary stage. Based on the arguments of the affirmative and negative sides, the interaction process between different participants in the debate is simulated by the facilitator, the affirmative side's agent, the negative side's agent, and the audience's agent. The event relationships predicted by each of these agents can be obtained, and may include: Step 401: In the opening phase, the host agent introduces the debate rules, process, theme, roles of the affirmative and negative agents, documents to be analyzed, event pairs to be analyzed, the affirmative debate viewpoints corresponding to the affirmative agents, and the negative debate viewpoints corresponding to the negative agents.

[0052] It should be noted that in the opening phase, the debate is initiated by a moderator agent, which introduces the debate rules, procedures, and the roles of the two sides (the affirmative and the negative). The moderator agent is also responsible for clarifying the topic of this debate, namely the type of event relationship to be predicted and the relevant background information of the input document to be analyzed.

[0053] Step 402: During the opening statement phase, the first debater of the affirmative team and the first debater of the negative team speak in turn, stating their respective debate viewpoints and problem-solving strategies. The problem-solving strategies are used to characterize the strategies adopted to predict the event relationships between the event pairs to be analyzed.

[0054] It should be noted that during the opening statements, the first speaker's AI agents for the affirmative and negative sides spoke in turn, presenting their respective viewpoints and the core evidence and problem-solving strategies supporting their predictions of event relationships. The first speaker's AI agent for the affirmative side presented its viewpoints based on a "generate first, then extract" strategy, while the first speaker's AI agent for the negative side presented its corresponding viewpoints based on a "select first, then extract" strategy.

[0055] Step 403: In the rebuttal phase, the affirmative and negative intelligent agents take turns questioning each other's argument and problem-solving approaches.

[0056] It should be noted that during the rebuttal phase, both intelligent agents take turns questioning each other's viewpoints and evidence.

[0057] Furthermore, during the rebuttal phase, the affirmative and negative intelligent agents take turns questioning each other's arguments and problem-solving approaches. This can include: the second speaker of the affirmative agent raising questions about the shortcomings of the arguments and problem-solving approaches presented by the first speaker of the negative agent; the second speaker of the negative agent responding to the questions raised by the second speaker of the affirmative agent, while simultaneously raising questions about the shortcomings of the arguments and problem-solving approaches presented by the second speaker of the affirmative agent; the third speaker of the negative agent responding to the questions raised by the second speaker of the affirmative agent, while simultaneously raising questions about the shortcomings of the arguments and problem-solving approaches presented by the third speaker of the affirmative agent; the affirmative and negative agents continue this question-and-answer process until all agents of both sides have participated in the question-and-answer session.

[0058] For example, the second speaker's AI agent on the affirmative side raises questions about potential weaknesses in the negative side's strategy, and the second speaker's AI agent on the negative side responds to the questions raised by the second speaker's AI agent on the affirmative side. Simultaneously, the second speaker's AI agent on the negative side raises questions about potential weaknesses in the affirmative side's viewpoint. This question-and-answer process continues between the affirmative and negative sides until all AI agents on both sides have participated in the question-and-answer session.

[0059] Step 404: During the free discussion phase, the affirmative and negative AI agents clarify their respective arguments.

[0060] It should be noted that during the free discussion phase, which is the open discussion phase, both the affirmative and negative AI agents clarify their respective viewpoints.

[0061] Furthermore, during the free discussion phase, the affirmative and negative intelligent agents clarify their respective argumentative viewpoints, which may include: combining previous interaction information, the affirmative intelligent agent regenerates a summary of the event pair to be analyzed and reiterates its own argumentative viewpoints and problem-solving approach; the negative intelligent agent re-extracts the set of sentences related to the event pair to be analyzed from the document to be analyzed and reiterates its own argumentative viewpoints and problem-solving approach; the affirmative and negative intelligent agents refute each other's argumentative viewpoints, with a maximum of four rounds of rebuttal.

[0062] For example, combining the previous information, the positive agent regenerates information about the event pairs. The opposing agent extracts a summary of the text, reiterating its own viewpoints and problem-solving strategies. Similarly, the opposing agent re-extracts a set of sentences from the text that are useful for predicting the relationships between events, and reiterates its own viewpoints and problem-solving strategies. Both the affirmative and opposing agents can refute each other's viewpoints. This stage involves a maximum of four rounds of discussion.

[0063] Step 405: In the summary stage, the first debater of the affirmative team and the first debater of the negative team predict the first event relationship and the second event relationship based on all the information of the interaction process, respectively. The audience team predicts the third event relationship based on all the information of the interaction process, the first event relationship and the second event relationship. The host team predicts the fourth event relationship based on the first event relationship, the second event relationship and the third event relationship.

[0064] It should be noted that in the summary phase, at the end of the debate, the first speaker's AI agents for both the affirmative and negative sides summarize their final viewpoints and predicted results. Subsequently, the audience AI agent, acting as a neutral party, provides a comprehensive summary of the debate process. It combines the arguments of both sides with information observed during the debate to generate a comprehensive outcome. Furthermore, the moderator AI agent also participates in the summary phase. As the supervisor and rule-maker of the entire debate process, the moderator AI agent comprehensively considers the performance of both sides and the analysis results of the audience AI agent in its summary.

[0065] The debate process includes multiple stages to simulate the interactions between different participants in a real debate, ensuring the accuracy and comprehensiveness of the outcome regarding the relationships between events.

[0066] In some embodiments, a relationship retrieval model may be constructed before obtaining the document to be analyzed and the event pair to be analyzed.

[0067] Figure 5 This is a flowchart illustrating the method for constructing a relationship acquisition model provided by this invention. Specifically, it involves constructing the agents participating in the debate and setting the rules for the agent debate competition. This section aims to set up the agents and rules in the agent debate competition, mainly including three sub-parts: agent role setting, agent debate rules, and debate speech preparation. For example... Figure 5 As shown, before obtaining the document to be analyzed and the event pair to be analyzed, the method may further include: Step 501: Set up multiple agents in the relationship acquisition model. The multiple agents are set with different positions. The multiple agents include a positive agent, a negative agent, a host agent, and an audience agent.

[0068] It should be noted that the roles of the agents need to be defined first. An agent debate requires agents with different stances to participate, mainly including three types of agents: debating agents (pro and con), moderator agents, and audience agents.

[0069] Step 502: Set the rules for the debate process in the relationship acquisition model. The debate includes an opening stage, a debate stage, and a summary stage. The debate stage includes opening statements, rebuttals, and free discussion. It should be noted that the rules for the agent debate competition need to be set next. The agent debate competition mainly consists of three steps: the opening stage, the debate stage, and the summary stage. The opening stage involves the moderator agent introducing the rules, the topic, and the positions of the different debating agents (the affirmative and negative agents), and announcing the start of the debate. The debate stage involves the debating agents (the affirmative and negative agents) respectively stating their positions and viewpoints, and exchanging and discussing with other debating agents (the affirmative and negative agents) to deepen their understanding of the relationships between events. The summary stage involves the debating agents (the affirmative and negative agents), the moderator agent, and the audience agents each summarizing their own viewpoints and outputting their respective relationships between events.

[0070] Step 503: Set the debate speech rules in the relationship acquisition model. The debate speech rules are used to instruct the affirmative agent and the negative agent to predict the event relationship, the debate method and the debate steps before the debate, based on the document to be analyzed, the event pairs to be analyzed and the debate method and the debate steps.

[0071] It should be noted that the final step involves preparing debate speeches. Before the moderating agent announces the start of the debate, the opposing agents—the affirmative and negative agents—need to prepare their speeches. Specifically, given an article and the events within it… and The debate agents with different positions, namely the affirmative and negative agents, need to give their own preliminary results, which mainly include three aspects: the results of the event relationship, the methods used, and the specific steps of the analysis.

[0072] Furthermore, in step 501, the objective of this invention is to employ an agent debate format, where different types of agents deepen their understanding of the relationships between events through debate. The agents participating in the agent debate are mainly divided into three types: debating agents (positive and negative agents), moderator agents, and audience agents.

[0073] (1) Moderator Agent. This agent is responsible for interpreting and enforcing the competition rules to ensure the fairness of the competition and the smooth progress of the debate. In addition, the moderator agent also needs to guide the discussion and act as a referee when necessary. At the end of the agent debate, the moderator agent needs to give its own understanding of the issue.

[0074] (2) Debate Agents. The debate agents consist of affirmative and negative agents, and are the main participants in the debate. They promote a deeper understanding of the debate topic among all participants by presenting their own viewpoints and engaging in discussions with the opposing side. The affirmative side includes three debate agents: the first, second, and third speaker. The negative side includes three debate agents: the first, second, and third speaker. The first speaker of the affirmative (or negative) side is responsible for presenting their own viewpoint at the beginning of the debate and, in the concluding phase, providing their final conclusion based on the discussion. The second and third speakers of the affirmative (or negative) side are responsible for answering questions raised by the opposing side based on the debate results and for asking questions about the opposing side's viewpoints.

[0075] (3) Audience Agent. This agent acts as an observer of the debate, deepening its understanding of the topic by listening to the content of the debate. From the perspective of an observer, the audience agent is able to more objectively evaluate and understand the debate topic.

[0076] In some embodiments, the present invention also constructs a memory mechanism for the agent. This mechanism introduces dynamic memory units into the agent, enabling it to record key information, track the debate process, and dynamically adjust its statements and positions based on new inputs. The memory mechanism mainly includes the following three core modules: (1) Information storage module. This module is used to store all the information received by the agent during the debate, including the opponent's statements, questions and the agent's own speech record.

[0077] (2) Information Retrieval and Update Module. The agent can retrieve stored information in real time to support its rapid response to opponents' viewpoints or to propose new arguments. At the same time, when new and important information emerges during the debate, this module can dynamically update the stored content to ensure the integrity and timeliness of the stored information.

[0078] (3) Memory Optimization Module. To avoid memory overload, this module uses a large model to streamline the memory information and optimize the agent's memory capacity. It prioritizes retaining information that is crucial to the current debate topic and removes irrelevant or low-relevance information.

[0079] Understandably, the agent's memory mechanism enables the agent to efficiently store, manage, and utilize information during debate, thereby enhancing its reasoning and decision-making abilities.

[0080] The following describes an exemplary application of the embodiments of the present invention in a practical application scenario.

[0081] Figure 6This is a schematic diagram of the overall architecture of the event relationship extraction system provided by this invention. The main technical problem solved by the event relationship extraction method provided by this invention is: how to effectively integrate the large-model multi-agent method with the event relationship extraction task, and how to design an effective multi-agent interaction method. The event relationship extraction system based on a large-model multi-agent proposed in this invention is mainly divided into four parts: the first part is constructing the agents participating in the debate and setting the rules for the agent debate; the second part is constructing the agent's memory mechanism; the third part is extracting the relationships between events through mutual interaction between agents; and the fourth part is integrating the event relationship extraction results of the agents. Figure 6 As shown, the event relationship extraction system includes the following functional units: The article refers to the document to be analyzed. The preprocessing unit treats each agent as an instance of a large model. Based on a multi-agent event relationship extraction method using the large model, relationships between events are obtained through an agent debate competition. The agent debate competition includes three types of agents: a moderator agent, debating agents (pro-side and anti-side agents), and audience agents. The pro-side and anti-side agents each hold a debate viewpoint and need to generate corresponding problem-solving strategies based on their viewpoints.

[0082] Memory Unit. After the start of the agent debate, the output of each round by the moderating agent, the debating agents (the affirmative and negative agents), and the audience agents is stored in the information storage module. When starting a new round of dialogue, the agents retrieve historical information from the information storage module to enhance the effectiveness and accuracy of the debate. To avoid memory overload, the memory optimization module uses a large model to streamline the stored information, optimizing the agent's memory capacity. It prioritizes retaining information crucial to the current debate topic and discards irrelevant or low-relevance information.

[0083] Debate Unit. Different types of intelligent agents use debate to understand the relationships between events. Throughout the debate, the moderator intelligent agent introduces the rules and controls the overall progress. The affirmative and negative intelligent agents respectively present their viewpoints and problem-solving approaches, raise questions about the opposing side's viewpoints and approaches, and answer those questions. Audience intelligent agents observe the entire debate. At the end of the debate, the moderator intelligent agent, the debating intelligent agents (the affirmative and negative agents), and the audience intelligent agents each provide their answers.

[0084] Prediction integration unit. This unit uses a large model to generate the final event relationship results from the four prediction outcomes.

[0085] Based on the above architecture, this invention provides a method for extracting event relationships based on a large-scale multi-agent model, the specific implementation steps of which are as follows: (1) Obtaining articles and the event To obtain the viewpoints of the debating agents, namely the affirmative and negative agents: generate first and then extract, or select first and then extract.

[0086] (2) The positivist generates corresponding problem-solving strategies based on the viewpoints extracted after initial generation. Specifically, the positivist generates information about event pairs. The first agent generates a summary of the text, using it to predict relationships between events. The second agent, based on a selection-then-extraction approach, generates a corresponding solution. Specifically, the second agent extracts a set of sentences from the text that are useful for predicting event relationships, and uses this selected set of sentences to predict the relationships between events.

[0087] (3) In the opening phase, the host agent reads out the rules, procedures, and roles of the two sides (affirmative and negative) of the agent debate. The host agent reads the article. The event And the debate viewpoints of the pro and con intelligent agents.

[0088] (4) In the opening statement stage, the first speaker of the affirmative side and the first speaker of the negative side spoke in turn, stating their respective viewpoints and problem-solving approaches. The first speaker of the affirmative side proposed a problem-solving approach based on the viewpoint of "generating first and then extracting", while the first speaker of the negative side made a corresponding statement based on the viewpoint of "selecting first and then extracting".

[0089] (5) Entering the rebuttal phase, the second speaker's AI agent on the affirmative side raises questions about potential shortcomings in the negative side's viewpoints and problem-solving approaches. The second speaker's AI agent on the negative side responds to the questions raised by the second speaker's AI agent on the affirmative side, while simultaneously raising questions about potential shortcomings in the affirmative side's viewpoints. The affirmative and negative sides engage in this question-and-answer process until all AI agents on both sides participate in the question-and-answer session.

[0090] (6) Entering the free discussion phase, both the affirmative and negative agents clarify their respective viewpoints. Based on the previous information, the affirmative agent regenerates information about the event pair. The opposing agent extracts a summary of the text, reiterating its own viewpoints and problem-solving strategies. Similarly, the opposing agent re-extracts a set of sentences from the text that are useful for predicting the relationships between events, and reiterates its own viewpoints and problem-solving strategies. Both the affirmative and opposing agents can refute each other's viewpoints. This stage involves a maximum of four rounds of discussion.

[0091] (7) Entering the summary stage, the first speaker of the affirmative side gives the event response based on the information output by their side in the opening statement stage, rebuttal stage and free discussion stage. The event relationships. The opposing side's first speaker's AI agent determines the event relationships based on the information provided by their side during the opening statement, rebuttal, and free discussion phases. The event relationships are determined by the information output by the opposing agents during the opening statements, rebuttals, and free discussion phases. Potential event relationship information. The facilitating agent combines information from the outputs of all agents during the opening, opening statements, rebuttal, and free discussion phases to predict event pairs. Information on possible event relationships.

[0092] (8) In the previous step, the first speaker's agent for the affirmative side, the first speaker's agent for the negative side, the audience agent, and the host agent respectively provided event pairs. Possible event relationship outcomes. Combining these four prediction outcomes, a large model is used to generate the final event relationship outcome.

[0093] The event relationship extraction method provided in this invention was tested on two public datasets, MAVEN-ERE and EventStoryLine. Experimental results show that, compared to methods such as thought chains and self-reflection, the method based on large-model multi-agent interaction significantly improves the performance and accuracy of event relationship extraction. Further analysis revealed that during the debate, the performance of different types of agents improved compared to before the debate.

[0094] Based on the foregoing embodiments, this invention provides an event relationship extraction device. The modules and units included in the device can be implemented by a processor; of course, they can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor (MPU), a digital signal processor (DSP), or a field-programmable gate array (FPGA), etc.

[0095] The event relationship extraction device provided by the present invention is described below. The event relationship extraction device described below and the event relationship extraction method described above can be referred to in correspondence.

[0096] Figure 7 This is a schematic diagram of the event relationship extraction device provided by the present invention. Figure 7 As shown, the device 700 includes an information acquisition module 701 and a relationship extraction module 702, wherein: The information acquisition module 701 is used to acquire the document to be analyzed and the event pair to be analyzed, wherein the event pair to be analyzed is a pair of events in the document to be analyzed. The relation extraction module 702 is used to input the document to be analyzed and the event pair to be analyzed into the relation acquisition model to obtain the target event relation. The target event relation is used to characterize the event relationship between the event pair to be analyzed. The target event relation is obtained by predicting and integrating the event relations by multiple agents in the relation acquisition model. The prediction of the event relations by the multiple agents is obtained by simulating the interaction of a debate competition by the multiple agents. Each agent in the multiple agents is used to represent an instance of a large model.

[0097] In some embodiments, the plurality of intelligent agents includes a host intelligent agent, a pro-side intelligent agent, a con-side intelligent agent, and an audience intelligent agent; the relationship extraction module 702 includes an opinion acquisition unit, an opinion debate unit, and an information integration unit, wherein... The viewpoint acquisition unit is used to acquire the positive debate viewpoints corresponding to the positive agent and the negative debate viewpoints corresponding to the negative agent based on the document to be analyzed and the event pair to be analyzed, wherein the positive debate viewpoints and the negative debate viewpoints are different. The viewpoint debate unit is used to simulate the interaction process between different participants in a debate competition through the host agent, the affirmative agent, the negative agent, and the audience agent, based on the affirmative and negative debate viewpoints, and to obtain the event relationships predicted by each of the multiple agents. The information integration unit is used to integrate the event relationships predicted by the multiple agents to obtain the target event relationship.

[0098] In some embodiments, the viewpoint acquisition unit is specifically used to: generate a summary for the event pair to be analyzed, and use the summary to predict the event relationship to obtain the affirmative debate viewpoint; extract sentences related to the event to be analyzed from the document to be analyzed to obtain a sentence set, and use the sentence set to predict the event relationship to obtain the negative debate viewpoint.

[0099] In some embodiments, the interaction flow between different participants in the debate includes an opening phase, a statement phase, a rebuttal phase, a free discussion phase, and a summary phase. The viewpoint debate unit includes an opening component, a statement component, a rebuttal component, a discussion component, and a summary component. The opening component is used to introduce the debate rules, process, theme, roles of the affirmative and negative intelligent agents, documents to be analyzed, event pairs to be analyzed, the affirmative debate viewpoints corresponding to the affirmative intelligent agents, and the negative debate viewpoints corresponding to the negative intelligent agents through the host intelligent agent during the opening phase. The declaring component is used during the declaring phase, where the first debater of the affirmative team and the first debater of the negative team speak in turn, stating their respective debate viewpoints and problem-solving strategies. The problem-solving strategies are used to characterize the strategies adopted for predicting the event relationships between the event pairs to be analyzed. The rebuttal component is used in the rebuttal phase, where the affirmative and negative intelligent agents take turns questioning each other's argument and problem-solving approaches. The discussion component is used during the free discussion phase, where the affirmative and negative agents clarify their respective argumentative viewpoints. The summary component is configured to, during the summary phase, have the first speaker of the affirmative team and the first speaker of the negative team predict a first event relationship and a second event relationship based on all information from the interaction process, respectively; have the audience speaker predict a third event relationship based on all information from the interaction process, the first event relationship, and the second event relationship; and have the host speaker predict a fourth event relationship based on the first event relationship, the second event relationship, and the third event relationship.

[0100] In some embodiments, the rebuttal component is specifically used for: the second speaker of the affirmative agent to raise questions about the shortcomings of the debate viewpoints and problem-solving approaches put forward by the first speaker of the negative agent; The second debater of the opposing team responds to the questions raised by the second debater of the affirmative team, while the second debater of the opposing team also raises questions about the shortcomings of the second debater of the affirmative team's argumentative viewpoints and problem-solving approaches. The third debater of the opposing team responds to the questions raised by the second debater of the affirmative team, while the third debater of the opposing team also raises questions about the shortcomings of the third debater of the affirmative team's argumentation and problem-solving approaches. The affirmative and negative agents engage in question-and-answer sessions in this manner until all agents in both the affirmative and negative agents participate in the question-and-answer process.

[0101] In some embodiments, the discussion component is specifically used for: combining previous interaction information, the affirmative agent regenerating a summary of the event pair to be analyzed, and reiterating its own argument and problem-solving approach; the negative agent re-extracting the set of sentences related to the event pair to be analyzed from the document to be analyzed, and reiterating its own argument and problem-solving approach. The affirmative and negative AI agents refute each other's arguments, with a maximum of four rounds of rebuttal.

[0102] In some embodiments, the apparatus further includes a model setting module, which is specifically used to: set multiple agents in the relationship acquisition model, wherein the multiple agents are set with different positions, and the multiple agents include a positive agent, a negative agent, a host agent, and an audience agent; The rules for the debate process in the relationship acquisition model are set, and the debate includes an opening stage, a debate stage, and a summary stage. The debate stage includes an opening statement stage, a rebuttal stage, and a free discussion stage. The relationship acquisition model sets up debate rules, which are used to instruct the affirmative and negative agents to predict event relationships, adopt debate methods, and take debate steps based on the document to be analyzed, the event pairs to be analyzed, and the debate methods to be adopted before the debate.

[0103] In this embodiment of the invention, the performance of large models in the task of event relationship extraction can be effectively improved, and the accuracy of the event relationship extraction results can be guaranteed.

[0104] Figure 8 This is a schematic diagram of the physical structure of the electronic device provided by the present invention. For example... Figure 8 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute an event relationship extraction method. This method includes: acquiring a document to be analyzed and a pair of events to be analyzed, wherein the pair of events to be analyzed is a pair of events in the document to be analyzed; inputting the document to be analyzed and the pair of events to be analyzed into a relationship acquisition model to obtain a target event relationship, wherein the target event relationship is used to characterize the event relationship between the pair of events to be analyzed, wherein the target event relationship is obtained by multiple agents in the relationship acquisition model predicting and integrating the event relationships respectively, and the multiple agents predicting the event relationships respectively are obtained by the multiple agents simulating a debate interaction, and each of the multiple agents represents an instance of a large model.

[0105] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the event relationship extraction method provided by the above methods. The method includes: acquiring a document to be analyzed and a pair of events to be analyzed, wherein the pair of events to be analyzed is a pair of events in the document to be analyzed; inputting the document to be analyzed and the pair of events to be analyzed into a relationship acquisition model to obtain a target event relationship, wherein the target event relationship is used to characterize the event relationship between the pair of events to be analyzed, wherein the target event relationship is obtained by predicting and integrating the event relationships by multiple agents in the relationship acquisition model, wherein the prediction of the event relationships by the multiple agents is obtained by simulating a debate interaction by the multiple agents, and each of the multiple agents is used to characterize an instance of a large model.

[0107] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).

[0108] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the event relationship extraction method provided by the above methods. The method includes: acquiring a document to be analyzed and a pair of events to be analyzed, wherein the pair of events to be analyzed is a pair of events in the document to be analyzed; inputting the document to be analyzed and the pair of events to be analyzed into a relationship acquisition model to obtain a target event relationship, wherein the target event relationship is used to characterize the event relationship between the pair of events to be analyzed, wherein the target event relationship is obtained by predicting and integrating event relationships by multiple agents in the relationship acquisition model, and the multiple agents predicting event relationships are obtained by simulating debate interactions among the multiple agents, each of the multiple agents representing an instance of a large model.

[0109] The aforementioned computer-readable storage medium may be any combination of one or more computer-readable media. A computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM) or flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0110] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0111] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, radio frequency (RF), etc., or any suitable combination thereof.

[0112] Computer program code for performing the operations described herein can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as "C" or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0113] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0115] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting event relationships, characterized in that, include: Obtain the document to be analyzed and the event pair to be analyzed, wherein the event pair to be analyzed is a pair of events in the document to be analyzed; The document to be analyzed and the event pair to be analyzed are input into the relationship acquisition model to obtain the target event relationship. The target event relationship is used to characterize the event relationship between the event pairs to be analyzed. The target event relationship is obtained by predicting and integrating the event relationships by multiple agents in the relationship acquisition model. The prediction of the event relationships by the multiple agents is obtained by simulating a debate interaction by the multiple agents. Each agent in the multiple agents is used to represent an instance of a large model. The plurality of intelligent agents includes a host intelligent agent, a positive intelligent agent, a negative intelligent agent, and an audience intelligent agent. The step of obtaining the target event relationship by inputting the document to be analyzed and the event to be analyzed into an input relationship acquisition model includes: Based on the document to be analyzed and the event pair to be analyzed, obtain the positive debate viewpoints corresponding to the positive agent and the negative debate viewpoints corresponding to the negative agent, wherein the positive debate viewpoints and the negative debate viewpoints are different; Based on the arguments of the affirmative and negative sides, the interaction process between different participants in a debate is simulated by the host agent, the affirmative agent, the negative agent, and the audience agent, obtaining the event relationships predicted by each of the multiple agents. The interaction process between different participants in the debate includes a summary stage. In the summary stage, the first speaker agents of the affirmative and negative sides predict a first event relationship and a second event relationship based on all the information of the interaction process, respectively. The audience agent predicts a third event relationship based on all the information of the interaction process, the first event relationship, and the second event relationship. The host agent predicts a fourth event relationship based on the first event relationship, the second event relationship, and the third event relationship. The target event relationship is obtained by integrating the event relationships predicted by the multiple agents. The step of obtaining the affirmative debate viewpoints corresponding to the affirmative agent and the negative debate viewpoints corresponding to the negative agent based on the document to be analyzed and the event pair to be analyzed includes: A summary is generated for the event pairs to be analyzed, and the event relationships are predicted using the summary to obtain the affirmative side's argument. Sentences related to the event to be analyzed are extracted from the document to be analyzed to obtain a sentence set, and the event relationship is predicted using the sentence set to obtain the opposing argument.

2. The event relationship extraction method according to claim 1, characterized in that, The interaction process between different participants in the debate also includes an opening phase, a statement phase, a rebuttal phase, and a free discussion phase. Based on the arguments of the affirmative and negative sides, the interaction process between different participants in the debate is simulated by the moderator agent, the affirmative agent, the negative agent, and the audience agent. The event relationships predicted by each of the multiple agents are obtained, including: In the opening phase, the host AI introduces the debate rules, process, theme, roles of the affirmative and negative AI, documents to be analyzed, event pairs to be analyzed, the affirmative debate viewpoints corresponding to the affirmative AI, and the negative debate viewpoints corresponding to the negative AI. During the opening statement phase, the first debater of the affirmative team and the first debater of the negative team speak in turn, stating their respective debate viewpoints and problem-solving strategies. The problem-solving strategies are used to characterize the strategies adopted to predict the event relationships between the event pairs to be analyzed. During the rebuttal phase, the affirmative and negative intelligent agents take turns questioning each other's arguments and problem-solving approaches. During the free discussion phase, the affirmative and negative AI agents clarified their respective arguments.

3. The event relationship extraction method according to claim 2, characterized in that, During the rebuttal phase, the affirmative and negative AI agents take turns questioning each other's arguments and problem-solving approaches, including: The second debater of the affirmative team raises questions about the shortcomings of the debate viewpoints and problem-solving approaches put forward by the first debater of the negative team. The second debater of the opposing team responds to the questions raised by the second debater of the affirmative team, while the second debater of the opposing team also raises questions about the shortcomings of the second debater of the affirmative team's argumentative viewpoints and problem-solving approaches. The third debater of the opposing team responds to the questions raised by the second debater of the affirmative team, while the third debater of the opposing team also raises questions about the shortcomings of the third debater of the affirmative team's argumentation and problem-solving approaches. The affirmative and negative agents engage in question-and-answer sessions in this manner until all agents in both the affirmative and negative agents participate in the question-and-answer process.

4. The event relationship extraction method according to claim 2, characterized in that, During the free discussion phase, both the affirmative and negative AI agents clarified their respective arguments, including: Based on the previous interaction information, the affirmative agent regenerates the summary of the event pair to be analyzed and reiterates its own argument and problem-solving approach. The negative agent re-extracts the set of sentences related to the event pair to be analyzed from the document to be analyzed and reiterates its own argument and problem-solving approach. The affirmative and negative AI agents refute each other's arguments, with a maximum of four rounds of rebuttal.

5. The event relationship extraction method according to claim 1, characterized in that, Before obtaining the document to be analyzed and the event pair to be analyzed, the method further includes: The relationship acquisition model is configured with multiple agents, each with a different stance. These agents include a pro-side agent, a con-side agent, a host agent, and an audience agent. The rules for the debate process in the relationship acquisition model are set, and the debate includes an opening stage, a debate stage, and a summary stage. The debate stage includes an opening statement stage, a rebuttal stage, and a free discussion stage. The relationship acquisition model sets up debate rules, which are used to instruct the affirmative and negative agents to predict event relationships, adopt debate methods, and take debate steps based on the document to be analyzed, the event pairs to be analyzed, and the debate methods to be adopted before the debate.

6. An event relationship extraction device, characterized in that, include: The information acquisition module is used to acquire the document to be analyzed and the event pair to be analyzed, wherein the event pair to be analyzed is a pair of events in the document to be analyzed. The relation extraction module is used to input the document to be analyzed and the event pair to be analyzed into the relation acquisition model to obtain the target event relation. The target event relation is used to characterize the event relationship between the event pair to be analyzed. The target event relation is obtained by predicting and integrating the event relations by multiple agents in the relation acquisition model. The prediction of the event relations by the multiple agents is obtained by simulating the interaction of a debate competition by the multiple agents. Each agent in the multiple agents is used to represent an instance of a large model. The multiple intelligent agents include a host intelligent agent, a pro-side intelligent agent, a con-side intelligent agent, and an audience intelligent agent. The relationship extraction module includes an opinion acquisition unit, an opinion debate unit, and an information integration unit. The viewpoint acquisition unit is used to acquire the positive debate viewpoints corresponding to the positive agent and the negative debate viewpoints corresponding to the negative agent based on the document to be analyzed and the event pair to be analyzed, wherein the positive debate viewpoints and the negative debate viewpoints are different. The viewpoint debate unit is used to simulate the interaction process between different participants in a debate competition based on the affirmative and negative debate viewpoints, through the host agent, affirmative agent, negative agent, and audience agent, to obtain the event relationships predicted by each of the multiple agents. The interaction process between different participants in the debate competition includes a summary stage. In the summary stage, the first speaker agents of the affirmative and negative agents respectively predict a first event relationship and a second event relationship based on all the information of the interaction process. The audience agent predicts a third event relationship based on all the information of the interaction process, the first event relationship, and the second event relationship. The host agent predicts a fourth event relationship based on the first event relationship, the second event relationship, and the third event relationship. The information integration unit is used to integrate the event relationships predicted by the multiple intelligent agents to obtain the target event relationship; The viewpoint acquisition unit is specifically used for: generating a summary for the event pair to be analyzed, and using the summary to predict the event relationship to obtain the affirmative debate viewpoint; extracting sentences related to the event to be analyzed from the document to be analyzed to obtain a sentence set, and using the sentence set to predict the event relationship to obtain the negative debate viewpoint.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the event relationship extraction method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the event relationship extraction method as described in any one of claims 1 to 5.

Images