A retrieval enhancement method for improving the reasoning ability of a large language model in a persuasive dialogue center

Abstract

The application relates to the technical field of large language models, in particular to a retrieval enhancement method for improving the reasoning ability of a large language model in a persuasion dialogue center, which comprises the following steps: obtaining a dialogue history record between a persuader and a persuaded person and a current persuasion dialogue; based on the dialogue history record, a second-order mind reasoning mechanism is used to deduce the current mind theory state of the persuaded person in the current persuasion dialogue; and a structured mind theory knowledge base is constructed; wherein the mind theory knowledge base is used to store the correlation between historical dialogue experience and corresponding mind theory state and persuasion strategy, so as to support the reasoning of the mind theory state; based on the deduced current mind theory state and the dialogue history record, the mind theory knowledge base is queried, and a persuasion strategy for the current persuasion dialogue is generated through a three-stage reasoning framework. The method can improve the accuracy of the large language model in the psychological state reasoning of the persuasion dialogue center and the persuasion strategy.

Description

Technical Field

[0001] Embodiments of this application relate to the field of large language model technology, and in particular to a retrieval enhancement method for improving the intellectual reasoning ability of large language models in persuasive dialogue. Background Technology

[0002] Theory of Mind (ToM) refers to an individual's ability to reason about their own and others' mental states (such as beliefs, desires, and intentions), which is crucial for cognitive processes such as decision-making in human-computer interaction. With the rapid development of large-scale language models, researchers have begun to focus on whether these models possess human-like reasoning abilities regarding mental states in human-computer interaction scenarios such as persuasive dialogues, and have constructed corresponding evaluation benchmarks for this purpose.

[0003] However, existing evaluation methods for theories of mind reveal significant limitations in large-scale language models. First, current benchmarks typically evaluate models directly based on dialogue history and questions related to theories of mind, neglecting the intrinsic interrelationships between beliefs, desires, and intentions. Second, persuasion in real-world social interactions involves continuous dynamic interactions between roles, but existing large-scale language models have been shown to struggle to truly grasp the dynamic changes in mental states throughout the dialogue. While cueing engineering techniques, such as mind chain cues, can enhance the logical reasoning ability of language models to some extent, their role in predicting dynamic mental states and generating persuasion strategies remains insignificant. Summary of the Invention

[0004] In view of this, embodiments of this application propose a retrieval enhancement method to improve the mental reasoning ability of large language models in persuasive dialogues. This method aims to solve the technical problem that large language models have difficulty dynamically modeling and reasoning about the psychological state of the persuaded in persuasive dialogues, resulting in a lack of psychological consistency and relevance in the generated persuasive strategies. Furthermore, through second-order mental reasoning, a structured knowledge base, and a three-thinking mechanism, the method significantly improves the accuracy of mental state reasoning and persuasive strategies of large language models in persuasive dialogues.

[0005] To achieve the above objectives, embodiments of this application propose a retrieval enhancement method to improve the intelligent reasoning ability of large language models in persuasive dialogues, the method comprising: Obtain the history of the conversation between the persuader and the persuadee, as well as the current persuasive dialogue; Based on the dialogue history, a second-order mental reasoning mechanism is used to infer the current theoretical state of the persuadee in the current persuasive dialogue; the theoretical state of mind includes beliefs, desires and intentions. Construct a structured theory of mind knowledge base; wherein, the theory of mind knowledge base is used to store the relationship between historical dialogue experience and corresponding theory of mind states and persuasion strategies, so as to support reasoning about theory of mind states; Based on the inferred current state of mental theory and the history of the dialogue, the mental theory knowledge base is queried, and a persuasion strategy for the current persuasive dialogue is generated through a three-stage reasoning framework.

[0006] To achieve the above objectives, embodiments of this application also propose a retrieval enhancement system for improving the intelligent reasoning ability of large language models in persuasive dialogues, the system comprising: The acquisition module is used to retrieve the history of conversations between the persuader and the persuadee, as well as the current persuasive conversation. The inference module is used to infer the current theoretical state of the persuadee in the current persuasive dialogue based on the dialogue history and using a second-order mental reasoning mechanism; where the theoretical state of mind includes beliefs, desires and intentions; The module is used to build a structured theory of mind knowledge base; the theory of mind knowledge base is used to store the relationship between historical dialogue experience and corresponding theory of mind states and persuasion strategies, so as to support reasoning about theory of mind states. The generation module is used to query the theory of mind knowledge base based on the inferred current state of mind and the history of the dialogue, and generate a persuasion strategy for the current persuasive dialogue through a three-stage reasoning framework.

[0007] To achieve the above objectives, embodiments of this application also propose an electronic device, including: a processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to execute the instructions such that the electronic device can implement the retrieval enhancement method described above for improving the intelligent reasoning ability of a large language model in persuasive dialogue.

[0008] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, enables a retrieval enhancement method as described above for improving the intelligent reasoning ability of a large language model in persuasive dialogue.

[0009] This application proposes a retrieval enhancement method to improve the mental reasoning ability of large language models in persuasive dialogues. The method involves obtaining the dialogue history and current persuasive dialogue between the persuader and the persuadee; based on the dialogue history, a second-order mental reasoning mechanism is used to infer the current mental theory state of the persuadee in the current persuasive dialogue; wherein the mental theory state includes beliefs, desires, and intentions; constructing a structured mental theory knowledge base; wherein the mental theory knowledge base is used to store the correlation between historical dialogue experience and corresponding mental theory states and persuasion strategies to support reasoning about mental theory states; based on the inferred current mental theory state and dialogue history, the mental theory knowledge base is queried, and a persuasion strategy for the current persuasive dialogue is generated through a three-stage reasoning framework. This solution can solve the technical problem that large language models have difficulty dynamically modeling and reasoning about the psychological state of the persuadee in persuasive dialogues, resulting in a lack of psychological consistency and relevance in the generated persuasion strategies. Furthermore, through second-order mental reasoning, a structured knowledge base, and a three-stage reasoning mechanism, the accuracy of mental state reasoning and persuasion strategies of large language models in persuasive dialogues are significantly improved.

[0010] Optionally, based on the dialogue history, a second-order mental reasoning mechanism is used to infer the current mental theory state of the persuadee in the current persuasive dialogue, including: obtaining the explicit language behavior of the persuadee in the dialogue history; and using the explicit language behavior as a starting point, using a second-order mental reasoning mechanism to perform reverse mental state reconstruction to obtain the current mental theory state.

[0011] Optionally, starting from explicit language behavior, a second-order mental reasoning mechanism is adopted to perform reverse mental state reconstruction to obtain the current mental theory state, including: inferring the persuader's intention based on explicit language behavior; inferring the corresponding desire based on the persuader's intention; and inferring the corresponding belief based on the desire, thereby obtaining the current mental theory state.

[0012] Optionally, a structured theory of mind knowledge base is constructed, including: formalizing the theory of mind experience in persuasion scenarios into triplets consisting of dialogue history, theory of mind states, and persuasion strategies; abstracting an abstract scenario representation containing descriptions of the persuadee's intentions from historical dialogue records; constructing knowledge unit quadruplets containing dialogue history, theory of mind states, persuasion strategies, and abstract scenario representations based on the abstract scenario representation and triplets; and constructing a searchable theory of mind knowledge base based on multiple knowledge unit quadruplets.

[0013] Optionally, based on the inferred current theory of mind state and dialogue history, a theory of mind knowledge base is queried, and a persuasion strategy for the current persuasive dialogue is generated through a three-stage reasoning framework. This includes: in the first reasoning stage, based on the dialogue history, the first probability distribution of the persuadee's desire is obtained from the theory of mind knowledge base, and then fused with the second probability distribution obtained from the intuitive judgment of the large language model, and the final desire prediction result is obtained through weighted calculation; in the second reasoning stage, based on the desire prediction result, belief samples related to the desire prediction result are retrieved from the theory of mind knowledge base, and the belief explanation leading to the desire is inferred; in the third reasoning stage, based on the current theory of mind state, the theory of mind knowledge base is retrieved, and the final persuasion strategy is generated through a weighted fusion mechanism.

[0014] Optionally, the first probability distribution The calculation formula is as follows: ; in, It represents the set of all possible states of the desires of the person being persuaded; Represents a set The Middle A specific state of desire; This indicates that the current desire state of the persuadee, inferred from the knowledge base of mental theories, is... The probability of; Second probability distribution The calculation formula is as follows: ; in, This indicates that the large language model infers the current desire state of the persuadee based on the current persuasive dialogue. The probability of; The final desire prediction result is obtained through weighted calculation. The calculation formula is as follows: ; in, Indicates hyperparameters, and .

[0015] Optionally, the value of the desire state includes multiple discrete psychological tendency levels; wherein, the psychological tendency levels include: unwilling, hesitant but inclined to refuse, unclear, hesitant but inclined to accept, and willing. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. Obviously, the following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0017] Figure 1 This is a flowchart of a retrieval enhancement method for improving the intelligent reasoning ability of a large language model in persuasive dialogue, provided in one embodiment of this application; Figure 2 This is a flowchart of a persuasive dialogue based on second-order mental reasoning provided in one embodiment of this application; Figure 3 This is a flowchart illustrating the construction of a theory of mind knowledge base, provided in one embodiment of this application. Figure 4 This is a flowchart of a three-way reasoning method provided in one embodiment of this application; Figure 5 This is a schematic diagram of the structure of a retrieval enhancement system that improves the intelligent reasoning ability of a large language model in persuasive dialogue, provided in another embodiment of this application; Figure 6 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been presented in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0019] Theory of Mind (ToM) refers to an individual's ability to reason about their own and others' mental states (such as beliefs, desires, and intentions). This ability enables people to understand, predict, and adapt to the behavior of others, and is crucial in cognitive processes such as decision-making in human-computer interaction. Persuasive dialogue aims to influence an individual's beliefs, attitudes, or behaviors through communication strategies. Recent work has attempted to develop datasets or facilitate specific goals for law students through persuasion techniques, such as donations, emotional support, and product recommendations, where this ability is essential in human-computer interaction. With the rapid development of Large Language Models (LLMs), researchers have begun to focus on whether these models possess human-like reasoning abilities regarding mental states in human-computer interaction scenarios such as persuasion. To this end, a series of ToM evaluation benchmarks for social scenarios have been constructed to measure the ToM capabilities of models.

[0020] In recent years, large-scale language models (LLMs) have made groundbreaking progress in the field of natural language understanding and generation, providing a solid foundation for the development of intelligent dialogue systems. However, in complex social interaction scenarios, especially in tasks involving persuasive dialogue, existing models still have significant shortcomings. First, traditional dialogue generation models mostly rely on surface semantic information for reasoning, lacking the ability to model the psychological dynamics of participants and failing to accurately grasp the internal psychological changes of the persuadee at the levels of belief, desire, and intention. Second, existing methods often generate persuasive statements only through explicit linguistic features, ignoring the underlying Theory of Mind (ToM) mechanisms in social cognition, resulting in persuasive strategies that lack specificity and psychological consistency. Third, due to the lack of structured knowledge support to support reasoning about mental states and the generation of behavior, models often struggle to conduct progressive psychological analysis and strategy optimization when faced with semantically ambiguous, unclear, or attitudinally wavering situations, affecting the coherence and effectiveness of the persuasion process.

[0021] Specifically, existing methods for evaluating ToM (To-M) reveal a series of significant limitations of LLMs (Large Language Models). First, existing benchmarks directly evaluate LLMs by asking ToM questions about the large language model based on the dialogue history, thus neglecting the interrelationships between beliefs, desires, and intentions. Second, persuasion in real-world social behavior is often accompanied by continuous dynamic interactions between roles; however, LLMs have been shown to struggle to truly understand the dynamic changes in mental states throughout the dialogue. Although cueing engineering techniques such as thought chain cues can enhance the logical reasoning ability of language models to some extent, their role in predicting dynamic mental states and generating persuasion strategies remains insignificant.

[0022] In view of this, embodiments of this application propose a retrieval enhancement method to improve the mental reasoning ability of large language models in persuasive dialogues. This method aims to solve the technical problem that large language models have difficulty dynamically modeling and reasoning about the psychological state of the persuaded in persuasive dialogues, resulting in a lack of psychological consistency and relevance in the generated persuasive strategies. Furthermore, through second-order mental reasoning, a structured knowledge base, and a three-thinking mechanism, the method significantly improves the accuracy of mental state reasoning and persuasive strategies of large language models in persuasive dialogues.

[0023] Specifically, the method provided in the embodiments of this application constructs a persuasive dialogue framework that combines psychological rationality and semantic generation capabilities by hierarchically modeling the psychological states of the persuader and the persuadee, using reverse reasoning and knowledge-enhanced reasoning. This framework includes three core components: First, it proposes a formal definition of a second-order ToM problem and a reverse psychological state modeling mechanism to achieve hierarchical reconstruction of the persuadee's beliefs, desires, and intentions; second, it constructs a structured ToM knowledge base, supporting interpretable reasoning and strategy generation of psychological states through experiential abstraction and semantic retrieval mechanisms; third, it designs a three-stage reasoning framework, the "Three-Thinking Mechanism," integrating model intuition and knowledge experience to achieve a progressive generation from psychological states to strategic behaviors. Through this mechanism, the present invention effectively enhances the mental reasoning and persuasive strategy planning capabilities of large language models in social interaction scenarios, making the persuasive dialogue process more consistent with human psychological logic and language behavior characteristics. See the following embodiments for details.

[0024] One embodiment of this application proposes a retrieval enhancement method to improve the intelligent reasoning ability of a large language model in persuasive dialogue, applied to an electronic device, wherein the electronic device can be a terminal or a server. This embodiment and the following embodiments will use a server as an example for description. The implementation details of the retrieval enhancement method to improve the intelligent reasoning ability of a large language model in persuasive dialogue proposed in this embodiment are described below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0025] The specific process of the retrieval enhancement method for improving the reasoning ability of large language models in persuasive dialogue centers proposed in this embodiment can be as follows: Figure 1 As shown, it includes: Step 101: Obtain the history of the conversation between the persuader and the persuadee, as well as the current persuasive conversation.

[0026] For example, contextual information from persuasive dialogues can be received and preprocessed. The dialogue history can contain all textual content of interactions between the persuader and the persuadee prior to the current round. The dialogue history serves as the input foundation for the entire reasoning process, characterizing the specific context and dynamics of the persuasion.

[0027] For example, by hierarchically representing and reasoning backward about the psychological states of the persuader and the persuadee, persuasive dialogue is formalized into a dialogue history. Theory of mind state and persuasion strategies The triplet .

[0028] Based on this formal definition, embodiments of this application divide the solution process for persuasive dialogue into two stages: the first stage is the solution of the second-order ToM state, that is, the persuader based on the dialogue history. Inferring the Theory of Mind State of the Person Being Persuaded The second stage is the generation of persuasion strategies, that is, after inferring the state of mind... Then, we further solve for persuasion strategies that match this psychological state. Alternatively, it can generate more adaptive and targeted persuasive statements.

[0029] Step 102: Based on the dialogue history, a second-order mental reasoning mechanism is used to infer the current mental theory state of the persuadee in the current persuasive dialogue.

[0030] The theory of mind states include beliefs, desires, and intentions.

[0031] Understandably, in order to address the lack of mental theory modeling in persuasion scenarios in existing technologies, embodiments of this application propose a persuasion dialogue modeling method based on a second-order mental reasoning mechanism, which aims to perform formal modeling and reverse reasoning of mental states in persuasion dialogues.

[0032] For example, the theory of mind state Further broken down into beliefs ,desire With intention Mental state Represented as a triple .belief Reflects an individual's cognitive assumptions about the world or events; desires Indicates an individual's desired goals or preferences; intentions. It is an action tendency formed based on beliefs and desires.

[0033] Understandably, in traditional first-order ToM reasoning, individuals typically model their mental state through forward reasoning: beliefs combined with external context generate desires, desires then form intentions, and ultimately lead to specific behaviors. For example, when an individual believes that "hiking helps to relax," they will have the desire to "go hiking," which further forms the intention to "go hiking this weekend," ultimately leading to actual action. In the second-order ToM scenario of persuasive dialogue, the situation is more complex. The persuadee's verbal behavior is itself an overt action; therefore, the persuader needs to infer their internal mental state through backward reasoning.

[0034] The model can dynamically reconstruct the psychological structure of a persuadee by inferring their intentions, desires, and beliefs from their language behavior, and then generate persuasion strategies that fit the current situation by combining persuasion theory. This mechanism simulates the psychological reasoning process of humans in social interactions, enabling the large language model to have a more accurate psychological understanding and strategy generation ability in complex social contexts.

[0035] In one possible embodiment, step 102 includes: obtaining the explicit language behavior of the persuaded person in the dialogue history; and, starting from the explicit language behavior, using a second-order mental reasoning mechanism to perform reverse mental state reconstruction to obtain the current mental theory state.

[0036] In one possible embodiment, starting from explicit language behavior, a second-order mental reasoning mechanism is used to perform reverse mental state reconstruction to obtain the current mental theory state, including: inferring the persuader's intention based on explicit language behavior; inferring the corresponding desire based on the persuader's intention; and inferring the corresponding belief based on the desire, thereby obtaining the current mental theory state.

[0037] For example, reverse mental state modeling can be performed. Unlike traditional forward reasoning from belief to intention, this application starts with the explicit linguistic behavior already exhibited by the persuadee (i.e., the persuadee's statements in the dialogue history) and performs reverse reasoning: first, inferring the intention. Analyze the overt language and behavior of the persuaded to infer their most immediate behavioral tendency or action plan. Then, infer their desires. Based on the inferred intention, reverse reasoning is used to generate the underlying motivation or goal for that intention. Finally, beliefs are inferred. This allows us to infer the underlying cognitions or assumptions that support the desire and intention, thereby arriving at the current state of our mental theory.

[0038] For example, such as Figure 2 As shown, in the scenario of "persuading a friend to go hiking on the weekend," if the person being persuaded responds with "I've been feeling a bit tired lately and want to rest," the model starts by inferring their intention from this statement. It could be that they "refused to climb the mountain and chose to rest"; thus, one can infer their desires. It was "a desire for rest rather than exercise"; ultimately, the underlying reason for this desire was deduced. belief It might be that "rest is more valuable than exercise when fatigued." Through this reverse reasoning chain of "intention → desire → belief," one can then understand the current mental state of the person being persuaded. Hierarchical reconstruction.

[0039] This embodiment enables dynamic consistency between mental state and linguistic behavior when outputting the current persuasive statement, thereby significantly improving the mental reasoning ability and strategy generation effect of the large language model in persuasive dialogue. This framework not only enhances the model's social cognitive ability in complex human-computer interactions but also provides basic structural support for subsequent dynamic ToM updates and retrieval enhancement mechanisms.

[0040] Step 103: Construct a structured knowledge base of mental theories.

[0041] The theory of mind knowledge base is used to store the correlation between historical dialogue experiences and corresponding theory of mind states and persuasion strategies, in order to support reasoning about theory of mind states.

[0042] Understandably, this step enables the model to perform experience analogies and retrieval enhancements.

[0043] For example, when faced with social interaction tasks, humans often recall similar past experiences and reason and make decisions based on existing knowledge and past situations. For instance, in a persuasive dialogue scenario, when we try to convince a friend to go out for hot pot, they don't directly agree or refuse, but instead respond, "I feel like I need to think about it." In human understanding, this expression usually implies hesitation in their current behavioral decision. From this, we can infer that they are uncertain about their intention to "go for hot pot"; further reasoning, their desire might lean towards other activities, and their belief reflects that they consider other activities to have higher value or meaning.

[0044] Inspired by this human decision-making model, we believe that human experience in solving the ToM problem in a persuasive context can be formalized into a triple. .in, Indicates the history of the conversation. This represents the individual's ToM state in this context, while This refers to the strategies or verbal behaviors employed by the persuader. We hypothesize that the process by which humans recall, compare, and infer similar situations in their brains can essentially be viewed as retrieving and matching data from a knowledge base rich in ToM (Too Much Influence) experiences. Through this knowledge retrieval mechanism, the system can automatically infer the psychological state of the person being persuaded based on past interaction experiences and generate more targeted persuasion strategies.

[0045] In one possible embodiment, step 103 includes: formalizing the mental theory experience in the persuasion scenario into a triple consisting of a dialogue history, a mental theory state, and a persuasion strategy; abstracting an abstract scenario representation containing a description of the persuadee's intention from the historical dialogue record; constructing a knowledge unit quadruple containing a dialogue history, a mental theory state, a persuasion strategy, and an abstract scenario representation based on the abstract scenario representation and the triples; and constructing a searchable mental theory knowledge base based on multiple knowledge unit quadruples.

[0046] Understandably, current large language models face certain difficulties in intent inference and often fail to accurately grasp the psychological dynamics of the persuaded. Therefore, embodiments of this application introduce an abstract representation mechanism, using an abstract scenario to describe the current dialogue history. This transforms the situation into a representative persuasion scenario, explicitly including the intentions of the person being persuaded. .

[0047] In this context, the value of the desire state includes multiple discrete psychological tendency levels; the psychological tendency levels include: unwilling, hesitant but inclined to refuse, unclear, hesitant but inclined to accept, and willing.

[0048] This design enables the model to establish discrete mental state mappings within continuous semantic persuasive dialogues, ultimately defining the knowledge in the knowledge base as a... The quadruple.

[0049] Furthermore, there is a certain causal dependency between beliefs and desires. To reflect this relationship, beliefs can be defined as the internal causes that lead to that desired state. For example, the belief that "doing other things is more valuable" is the fundamental basis for the hesitant or even refusing behavior of the persuaded individual. Finally, This refers to the persuasive strategies (or persuasive statements) adopted in response to the ToM state, including but not limited to stating viewpoints, asking for reasons, and providing emotional feedback. By constructing this structured ToM knowledge base, the system can accurately model the psychological state of the persuadee and dynamically generate strategies in persuasive dialogues, thus more closely resembling human social cognition and decision-making mechanisms.

[0050] For example, the construction of a theory of mind knowledge base stems from the simulation of the human thought mechanism that involves reasoning by recalling similar experiences in social interactions. For instance... Figure 3 As shown, Figure 3 This application provides a flowchart for constructing a theory of mind knowledge base, as illustrated in its embodiments. Each experience in the knowledge base is structured as a quadruple. The construction process is as follows: 1. Abstraction and Representation: For a historical dialogue experience, the first step is to examine its dialogue history. Abstraction is performed to generate an abstract scenario description that includes the intentions of the person being persuaded. This helps to extract representative persuasion situational patterns from specific and varied dialogues.

[0051] 2. Discretization of desire: The desire of the person being persuaded in the current situation. The psychological states were quantified into five discrete levels: "unwilling," "hesitant but inclined to refuse," "unclear," "hesitant but inclined to accept," and "willing." This provides a standardized dimension for matching and comparing psychological states.

[0052] 3. Define causal beliefs: ... Clearly defined as the cause of this desire The underlying causes or explanations of the state lead to an explicit model of the causal relationship between beliefs and desires.

[0053] 4. Relevance Strategies: Record the persuasive strategies actually employed by the persuader in this historical context in response to the aforementioned mental states. Effective persuasion strategies .

[0054] Example: In the scenario of "persuading someone to eat hot pot," based on past experiences of the other party's response, "I feel I need to think about it," the knowledge base construction process is as follows: Abstract scenario: This can be summarized as "persuading a friend to participate in a dinner party, but the friend hesitates."

[0055] intention The response suggests that the person was "hesitant about whether to participate."

[0056] desire Based on their hesitant attitude, mark them as "hesitant but inclined to refuse" or "unclear".

[0057] belief Infer the reasons for this desire, such as "believing that other activities are more valuable at the moment" or "not being very interested in hot pot".

[0058] Strategy Record effective persuasive phrases used at the time, such as, "That restaurant has a new tomato hot pot; didn't you say you wanted to try it last time?"

[0059] Ultimately, this experience is based on the quadruple ( Hesitation Hesitant but inclined to refuse. They believe other activities are more valuable. (Mentioning new flavors as attractive) is stored in the mental theory knowledge base.

[0060] Step 104: Based on the inferred current mental theory state and dialogue history, query the mental theory knowledge base and generate a persuasion strategy for the current persuasive dialogue through a three-stage reasoning framework.

[0061] For example, after completing problem modeling and the construction of the ToM knowledge base, we further propose a mental reasoning framework called "Three-step Reasoning" (also known as a three-stage reasoning framework) to simulate the psychological reasoning process of humans in social interactions. This mechanism aims to achieve accurate estimation of the ToM state of the persuadee and the generation of effective persuasion strategies based on the constructed ToM knowledge base through progressive speculation and verification.

[0062] Understandably, this mechanism simulates the progressive thinking process of humans in social interactions. Relying on the established ToM knowledge base, it gradually completes the inference of the psychological state of the persuadee and the planning of strategies through the fusion of experience retrieval and intuition.

[0063] For example, firstly, the system abstracts persuasion patterns based on dialogue history and retrieves semantically similar experience samples as the query target to obtain the probability distribution of the persuadee's desires. Simultaneously, it integrates the immediate judgment distribution of the large language model and obtains the final desire prediction result through weighted calculation. Subsequently, after obtaining the persuadee's attitude tendency, the system infers from relevant belief samples in the knowledge base to generate a psychological explanation for the formation of this desire (second consideration). Finally, after synthesizing the ToM state of intention, desire, and belief, the system generates the persuasion strategy most suitable for the current context based on the weighted retrieval and probability fusion mechanism of the knowledge base (third consideration). Through this three-stage reasoning mechanism, the model achieves hierarchical progressive reasoning from psychological state to strategic behavior, possessing both psychological rationality and semantic generation capabilities, significantly improving the mental understanding and decision-making level of the large language model in persuasive dialogues.

[0064] In one possible embodiment, step 103 includes: a persuasion strategy for sexual dialogue, comprising: in a first reasoning stage, based on the dialogue history, retrieving a first probability distribution of the persuadee's desire from the theory of mind knowledge base, and fusing it with a second probability distribution obtained from the intuitive judgment of a large language model, and obtaining a final desire prediction result through weighted calculation; in a second reasoning stage, based on the desire prediction result, retrieving belief samples related to the desire prediction result from the theory of mind knowledge base, and inferring to generate a belief explanation that leads to the desire; in a third reasoning stage, based on the current theory of mind state, retrieving the theory of mind knowledge base, and generating a final persuasion strategy through a weighted fusion mechanism.

[0065] For example, such as Figure 4 As shown, Figure 4 The flowchart of a three-stage reasoning method provided in the embodiments of this application is used to simulate human progressive thinking, integrate model intuition and knowledge base experience, and generate a strategy in three stages: The first reasoning stage (inferring the persuadee's desires): The system first abstracts from the current dialogue history and uses this as a query to retrieve semantically similar historical experience samples from the theory of mind knowledge base. From these samples, an empirical distribution of the persuadee's desires (i.e., the first probability distribution) can be statistically obtained. Simultaneously, the large language model itself, based on its parameterized knowledge, also provides an intuitive distribution of the current desires (i.e., the second probability distribution). The system weighted and fused these two distributions to obtain the final optimized desire prediction result, thus completing the initial judgment of the persuadee's attitude tendency.

[0066] In one possible embodiment, the first probability distribution The calculation formula is as follows: ; in, It represents the set of all possible states of the desires of the person being persuaded; Represents a set The Middle A specific state of desire; This indicates that the current desire state of the persuadee, inferred from the knowledge base of mental theories, is... The probability of; Second probability distribution The calculation formula is as follows: ; in, This indicates that the large language model infers the current desire state of the persuadee based on the current persuasive dialogue. The probability of; The final desire prediction result is obtained through weighted calculation. The calculation formula is as follows: ; in, Indicates hyperparameters, and .

[0067] The second stage of reasoning (inferring the persuadee's beliefs): After obtaining the desire, it uses this as a key clue to query the theory of mind knowledge base again, retrieving historical experience samples with the same or similar desire tags. From these samples, the system extracts and analyzes the corresponding belief descriptions, and then infers and generates explanations of the psychological reasons leading to the current desire, that is, inferring the current belief state. This allows the model not only to know what the other party "wants," but also to understand "why they think that way."

[0068] The third reasoning stage (generating a persuasive strategy): After synthesizing the complete current mental state, the system performs final reasoning. Using this triple as the query, a weighted search is conducted in the knowledge base to find the historical experience that best matches the current mental state. Employing a similar fusion mechanism as the first reasoning stage, the system combines the retrieved historical strategy distribution with the strategy intuition generated by the model itself. Through weighted calculation, a final strategy that best fits the current mental state and context, and is the most persuasive, is generated and output.

[0069] Understandably, the Sansi Mental Reasoning Framework enables a complete reasoning chain from inference of psychological state to generation of behavioral strategies. It retains the hierarchical cognitive characteristics of human ToM reasoning while also taking into account the model's semantic generation capabilities, thus making persuasive dialogue more psychologically reasonable and interactively adaptable.

[0070] Combining all the above steps, here's an example to illustrate the entire process: Input (corresponding to step 101): Conversation history: Persuader A says, "That sci-fi movie just came out this weekend, it has a really high rating, let's go see it together?"; Persuaded B replies, "Sci-fi movie... I've been wanting to watch something light and funny lately." Second-order reasoning (corresponding to step 102): The model infers backward from the persuadee B's response to infer their intention. The criteria for this profile are: "Refuse to watch science fiction films and prefer comedy films"; desire tendency is "hope to obtain light entertainment"; belief may be "thinking that science fiction films are not light enough" or "comedy films can relieve stress better".

[0071] Knowledge base query and reasoning (corresponding to steps 103 and 104): First thought: Using the abstracted dialogue scenario of "recommending a movie but being politely declined" as an example, we searched the knowledge base, combined with experience and model intuition, and confirmed that the desire was "wanting to watch an easy movie".

[0072] Second thought: Based on the search of the mental theory knowledge base of desire information, it was found that the belief under similar desires is "seeking relaxation", thus strengthening the belief inference that "the current need is for relaxing entertainment rather than mentally taxing".

[0073] The third approach involves querying the theory of mind knowledge base based on intentions, desires, and beliefs. This might retrieve historical strategies such as, "I've heard there's a new comedy that's really funny, how about we check it out?" The model integrates this experience with the current context to generate the final persuasive strategy: "Oh, I want to relax a bit. There's a new comedy called 'XX' that's got good reviews and is really funny, how about we watch that instead?" Through the above steps, this invention achieves a complete closed loop from understanding the dialogue and inferring the deep psychological state to finally generating targeted and psychologically logical persuasion strategies, significantly improving the mental reasoning and strategy generation capabilities of the large language model in persuasive dialogues.

[0074] This application proposes a retrieval enhancement method to improve the mental reasoning ability of large language models in persuasive dialogues. The method involves obtaining the dialogue history between the persuader and the persuadee; based on the dialogue history, employing a second-order mental reasoning mechanism to infer the current mental theory state of the persuadee in the current persuasive dialogue; wherein the mental theory state includes beliefs, desires, and intentions; constructing a structured mental theory knowledge base; wherein the mental theory knowledge base stores the correlation between historical dialogue experiences and corresponding mental theory states and persuasion strategies to support reasoning about mental theory states; based on the inferred current mental theory state and dialogue history, querying the mental theory knowledge base, and generating a persuasion strategy for the current persuasive dialogue through a three-stage reasoning framework; this solution can solve the technical problem that large language models struggle to dynamically model and reason about the psychological state of the persuadee in persuasive dialogues, resulting in a lack of psychological consistency and relevance in the generated persuasion strategies. Furthermore, through second-order mental reasoning, a structured knowledge base, and a three-stage reasoning mechanism, it significantly improves the accuracy of psychological state reasoning and persuasion strategies in persuasive dialogues using large language models.

[0075] The steps described above are for clarity only. In implementation, they can be combined into one step, or some steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0076] Another embodiment of this application proposes a retrieval enhancement system to improve the intelligent reasoning ability of a large language model in persuasive dialogue. The details of this retrieval enhancement system are described below for ease of understanding and are not essential for implementing this example. Figure 5 This is a schematic diagram of the structure of a retrieval enhancement system proposed in this embodiment to improve the intelligent reasoning ability of a large language model in persuasive dialogue, including: Module 210 is used to retrieve the dialogue history between the persuader and the persuadee. The inference module 220 is used to infer the current theoretical state of the persuadee in the current persuasive dialogue based on the dialogue history and using a second-order mental reasoning mechanism; wherein, the theoretical state of mind includes beliefs, desires and intentions; Module 230 is used to build a structured theory of mind knowledge base; wherein, the theory of mind knowledge base is used to store the relationship between historical dialogue experience and corresponding theory of mind states and persuasion strategies, so as to support reasoning about theory of mind states; The generation module 240 is used to query the mental theory knowledge base based on the inferred current mental theory state and dialogue history, and generate a persuasion strategy for the current persuasive dialogue through a three-stage reasoning framework.

[0077] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above method embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above method embodiments.

[0078] It is worth mentioning that all modules and units involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units do not exist in this embodiment.

[0079] Another embodiment of this application provides an electronic device, such as Figure 6 As shown, it includes a processor 31 and a memory 32. The memory 32 stores instructions that the processor 31 can execute. When the processor 31 is configured to execute the instructions, the electronic device can realize a retrieval enhancement method for improving the intelligent reasoning ability of a large language model in persuasive dialogue, as described in the above method embodiment.

[0080] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges, connecting various circuits of one or more processors and the memory. The bus can also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0081] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0082] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, enables a retrieval enhancement method for improving the intelligent reasoning ability of a large language model in persuasive dialogue, as described in the above method embodiments.

[0083] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0084] Those skilled in the art will understand that the above embodiments are specific implementations of this application, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A retrieval enhancement method for improving the intellectual reasoning ability of large language models in persuasive dialogues, characterized in that, The method includes: Obtain the history of the conversation between the persuader and the persuadee, as well as the current persuasive dialogue; Based on the dialogue history, a second-order mental reasoning mechanism is used to infer the current theoretical state of the persuadee in the current persuasive dialogue; the theoretical state of mind includes beliefs, desires and intentions. Construct a structured theory of mind knowledge base; wherein, the theory of mind knowledge base is used to store the relationship between historical dialogue experience and corresponding theory of mind states and persuasion strategies, so as to support reasoning about theory of mind states; Based on the inferred current state of mental theory and the history of the dialogue, the mental theory knowledge base is queried, and a persuasion strategy for the current persuasive dialogue is generated through a three-stage reasoning framework.

2. The method according to claim 1, characterized in that, The method of inferring the current theoretical state of the persuadee in the current persuasive dialogue, based on the dialogue history and employing a second-order mental reasoning mechanism, includes: Obtain explicit language behavior of the persuaded person in the conversation history; Starting from explicit language behavior, a second-order mental reasoning mechanism is adopted to perform reverse mental state reconstruction and obtain the current mental theory state.

3. The method according to claim 2, characterized in that, Starting from explicit language behavior, a second-order mental reasoning mechanism is employed to perform reverse mental state reconstruction, resulting in the current mental theory state, including: Inferring the persuadee's intentions based on explicit language behavior; Based on the intentions of the persuadee, infer their corresponding desires in reverse; Based on desires, we reason about the corresponding beliefs, and thus arrive at the current state of our mental theory.

4. The method according to claim 1, characterized in that, The construction of a structured knowledge base of mental theories includes: The mental theory experience in persuasion scenarios is formalized into a triad consisting of dialogue history, mental theory state, and persuasion strategy; Abstract scene representations that contain descriptions of the persuadee's intentions, extracted from historical dialogue records; Based on abstract scene representation and triples, a knowledge unit quadruple is constructed, which includes dialogue history, mental theory state, persuasion strategy and abstract scene representation. Based on multiple knowledge unit quadruples, a searchable knowledge base of mental theories is constructed.

5. The method according to claim 1, characterized in that, Based on the inferred current state of mental theory and the dialogue history, the system queries the mental theory knowledge base and generates a persuasive strategy for the current persuasive dialogue using a three-stage reasoning framework, including: In the first reasoning stage, based on the dialogue history, the first probability distribution of the persuaded person's desire is obtained by retrieving the knowledge base of mental theory, and then it is fused with the second probability distribution obtained by the intuitive judgment of the large language model. The final desire prediction result is obtained through weighted calculation. In the second reasoning stage, based on the desire prediction result, belief samples related to the desire prediction result are retrieved from the mind theory knowledge base, and reasoning is used to generate the belief explanation that leads to the desire. In the third reasoning stage, based on the current state of mental theory, the mental theory knowledge base is retrieved, and the final persuasion strategy is generated through a weighted fusion mechanism.

6. The method according to claim 5, characterized in that, First probability distribution The calculation formula is as follows: ； in, It represents the set of all possible states of the desires of the person being persuaded; Represents a set The Middle A specific state of desire; This indicates that the current desire state of the persuadee, inferred from the knowledge base of mental theories, is... The probability of; Second probability distribution The calculation formula is as follows: ； in, This indicates that the large language model infers the current desire state of the persuadee based on the current persuasive dialogue. The probability of; The final desire prediction result is obtained through weighted calculation. The calculation formula is as follows: ； in, Indicates hyperparameters, and .

7. The method according to any one of claims 1 to 6, characterized in that, The value of the desire state includes multiple discrete psychological tendency levels; among which, the psychological tendency levels include: unwilling, hesitant but inclined to refuse, unclear, hesitant but inclined to accept, and willing.

8. A retrieval enhancement system for improving the intelligent reasoning ability of large language models in persuasive dialogues, characterized in that, The system includes: The acquisition module is used to retrieve the history of conversations between the persuader and the persuadee, as well as the current persuasive conversation. The inference module is used to infer the current theoretical state of the persuadee in the current persuasive dialogue based on the dialogue history and using a second-order mental reasoning mechanism; where the theoretical state of mind includes beliefs, desires and intentions; The module is used to build a structured theory of mind knowledge base; the theory of mind knowledge base is used to store the relationship between historical dialogue experience and corresponding theory of mind states and persuasion strategies, so as to support reasoning about theory of mind states. The generation module is used to query the theory of mind knowledge base based on the inferred current state of mind and the history of the dialogue, and generate a persuasion strategy for the current persuasive dialogue through a three-stage reasoning framework.

9. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to, when executing the instructions, enable the electronic device to implement a retrieval enhancement method for improving the intelligent reasoning ability of a large language model in persuasive dialogue as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement a retrieval enhancement method as described in any one of claims 1 to 7 to improve the intelligent reasoning ability of a large language model in persuasive dialogue.

Images