Causal attribution-based knowledge graph self-organizing evolution method and system

Abstract

This application discloses a self-organizing evolution method and system for knowledge graphs based on causal attribution, relating to the field of artificial intelligence technology, aiming to solve the problems of insufficient creativity in dialogue systems and the degradation of knowledge graph structures. The method includes: selecting high-level intents from a causal strategy knowledge graph in response to a dialogue state; extracting subgraphs or paths as semantic constraints and causal utility guides; generating response dialogue using the constraints and guides; performing causal attribution analysis on the dialogue trajectory containing the dialogue to determine the causal contribution of the dialogue to business objectives; and performing a self-organizing rewriting operation on the knowledge graph based on the causal contribution. The system is used to implement this method. This application, by constructing an intelligent closed loop, enables the knowledge graph to self-purify and iterate, realizing a transformation from "limited selection" to "autonomous generation" and from "incremental growth" to "self-organizing evolution," thereby improving the long-term performance of the system.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a self-organizing evolution method and system for knowledge graphs based on causal attribution. Background Technology

[0002] In conversational AI applications such as intelligent customer service and intelligent sales, existing technologies typically employ reinforcement learning combined with knowledge graphs to optimize dialogue strategies. In this model, the knowledge graph predefines the states and optional actions in the dialogue process, and the reinforcement learning agent makes choices within a predefined action space to achieve a specific goal. When the strategy needs to be updated, new nodes or edges are usually added to the knowledge graph manually or through simple automated rules.

[0003] However, existing technologies have the following drawbacks:

[0004] First, because decision-making is strictly confined to a predefined discrete action space, the system cannot creatively generate new, potentially better dialogue strategies (such as scripts), making it difficult to cope with open and dynamic dialogue scenarios. At the same time, compliance constraints based on keyword blocking are relatively superficial and cannot completely prevent the generation of illegal content at the semantic level, resulting in insufficient creativity and compliance.

[0005] Secondly, the evolution of knowledge graphs is primarily incremental, meaning that new effective paths are constantly being added. This approach lacks a mechanism for holistic review, abstraction, and elimination of the knowledge system. Over time, the knowledge graph becomes excessively bloated and redundant, with a large number of outdated or inefficient knowledge paths remaining, leading to a chaotic knowledge structure. This negatively impacts the system's decision-making efficiency and long-term performance; in other words, the knowledge base can only "grow" and cannot undergo effective "metabolism," ultimately resulting in the degradation of the knowledge system structure. Summary of the Invention

[0006] The technical problem this application aims to solve is that existing dialogue systems lack creativity due to their reliance on predefined action spaces, and their compliance controls remain superficial. At the same time, the existing knowledge graph update mechanism cannot effectively evolve based on causal feedback from business objectives, leading to the degradation of the knowledge system structure.

[0007] To address the aforementioned technical problems, this application provides a self-organizing evolution method for knowledge graphs based on causal attribution, comprising: responding to the current dialogue state in a dialogue system, selecting a high-level dialogue intent from a causal strategy knowledge graph; based on the current dialogue state and the high-level dialogue intent, extracting a subgraph or path containing predefined nodes and predefined edges from the causal strategy knowledge graph as semantic constraints and causal utility guidance; generating response dialogue using the semantic constraints and the causal utility guidance, and recording the association between the response dialogue and the nodes, as well as the association between the response dialogue and the edges, wherein generating response dialogue further comprises: during the decoding process, distancing candidate content similar to prohibited semantics defined in the semantic constraints, and bringing closer content to the causal utility guidance. The system identifies candidate content similar to the recommended content indicated by the guide; performs causal attribution analysis on dialogue trajectories containing the response phrases to determine the causal contribution of the response phrases to preset business objectives; and performs a self-organizing rewriting operation on the causal strategy knowledge graph based on the causal contribution, the self-organizing rewriting operation including at least one of the following: when multiple dialogue trajectories with similar causal contribution patterns to preset business objectives are identified, creates abstract concept nodes representing the multiple dialogue trajectories in the causal strategy knowledge graph; and reduces or removes nodes or edges in the causal strategy knowledge graph that have causal contributions below a preset threshold, wherein the causal contribution of the node or edge is calculated based on the causal contribution of at least one response phrase generated using the node or edge.

[0008] Optionally, the structure of the causal strategy knowledge graph includes: causal relationship edges for storing probabilistic causal relationships, and abstract concept nodes dynamically created through the self-organizing rewrite operation.

[0009] Optionally, the process of pushing away candidate content that is semantically similar to prohibited content and bringing closer candidate content that is similar to recommended content during the decoding process is achieved using constrained beam search.

[0010] Optionally, the method further includes: optimizing the causal strategy knowledge graph by performing a self-organizing rewrite operation through meta-learning, wherein the self-organizing rewrite operation is regarded as a meta-action, and the change in the overall performance of the dialogue system after the operation is performed is used as a meta-reward.

[0011] Optionally, the method further includes: storing the dialogue trajectory in an experience database; and performing the self-organizing rewrite operation on the causal strategy knowledge graph asynchronously and periodically based on the data accumulated in the experience database.

[0012] To address the aforementioned technical problems, this application also provides a causal attribution-based knowledge graph self-organizing evolution system for implementing the aforementioned method, comprising: a selection module for selecting high-level dialogue intents from a causal strategy knowledge graph in response to the current dialogue state in the dialogue system; an extraction module for extracting subgraphs or paths containing predefined nodes and predefined edges from the causal strategy knowledge graph based on the current dialogue state and the high-level dialogue intents, as semantic constraints and causal utility guidance; a text generation module; and a generation control module for controlling the text generation module using the semantic constraints and the causal utility guidance to generate response speech, and establish the association between the response speech and the nodes, as well as the association between the response speech and the edges; the generation control module is configured to, during the decoding process, deflect semantics similar to prohibited semantics defined in the semantic constraints. The system includes: candidate content, and narrowing down candidate content similar to the recommended content indicated by the causal utility guidance; a causal attribution module, used to perform causal attribution analysis on dialogue trajectories containing the response phrases to determine the causal contribution of the response phrases to a preset business objective; and a graph optimization component, used to perform a self-organizing rewriting operation on the causal strategy knowledge graph based on the causal contribution, wherein the self-organizing rewriting operation includes at least one of the following: when multiple dialogue trajectories with similar causal contribution patterns to the preset business objective are identified, creating abstract concept nodes representing the multiple dialogue trajectories in the causal strategy knowledge graph; and reducing or removing nodes or edges in the causal strategy knowledge graph that have a causal contribution below a preset threshold, wherein the causal contribution of the node or edge is calculated based on the causal contribution of at least one response phrase generated using the node or edge.

[0013] Optionally, the system further includes: a storage module for storing the causal strategy knowledge graph, the structure of which includes: causal relationship edges for storing probabilistic causal relationships; and abstract concept nodes dynamically created through the self-organizing rewrite operation.

[0014] Optionally, the generation control module is configured to use constrained beam search to push away candidate content that is semantically similar to prohibited content and bring closer candidate content that is similar to recommended content during the decoding process.

[0015] Optionally, the system further includes a meta-learning module, used to optimize the graph optimization component through meta-learning, wherein the self-organizing rewrite operation is regarded as a meta-action, and the change in the overall performance of the dialogue system after the operation is executed is used as a meta-reward.

[0016] Optionally, the system further includes: an experience database for storing the dialogue trajectories; and the graph optimization component is configured to asynchronously and periodically perform self-organizing rewriting operations on the causal strategy knowledge graph based on the data accumulated in the experience database.

[0017] This application has the following beneficial effects:

[0018] 1. This application extracts semantic constraints and causal utility guidance from knowledge graphs, and controls text generation during the decoding process by pushing away candidate content that is semantically similar to prohibited content and bringing closer candidate content that is similar to recommended content, rather than selecting from a preset set of discrete actions. This enables the system to create diverse, highly contextualized dialogues that help ensure deep compliance based on real-time semantic guidance, realizing a paradigm shift from "limited selection" to "autonomous generation". This allows the system to generate novel solutions in real time to deal with open and dynamic dialogue scenarios. The generation control mechanism of this application can achieve deep compliance at the semantic level, significantly improving the system's creativity and compliance assurance capabilities.

[0019] 2. This application introduces a self-organizing rewriting mechanism that includes creating abstract concept nodes (abstraction, generalization) and removing low-contribution nodes or edges (pruning), enabling the knowledge graph to self-purify, iterate, and improve structural efficiency. This mechanism allows the knowledge system to undergo effective "metabolism," fundamentally preventing the problems of knowledge entropy increase and structural degradation caused by long-term operation, and achieving a significant improvement from "incremental growth" to "self-organizing evolution." The updates in this application are driven by explicit causal contributions to business objectives, achieving deep coupling between update behavior and business objective optimization, ensuring that the evolution direction of the knowledge graph is consistent with business needs.

[0020] 3. This application deeply integrates causal inference into script generation, effect attribution, and knowledge graph reconstruction, constructing an intelligent closed loop from strategy generation, effect verification, causal attribution to structural reconstruction. Each evolution of the system is driven by a clear causal contribution to business objectives, rather than simple correlation statistics, which helps improve the accuracy and efficiency of knowledge evolution. The semantic control achieved through the "push-pull" mechanism and the causal contribution-driven graph evolution work together to form a complete and self-consistent technical solution, enabling the system to continuously learn, create, and evolve, significantly improving the system's long-term performance. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application 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 only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating the self-organizing evolution method of knowledge graph based on causal attribution provided in this application embodiment;

[0023] Figure 2 A schematic diagram of the functional module architecture of a knowledge graph self-organizing evolution system based on causal attribution provided in an embodiment of this application;

[0024] Figure 3 This is a timing diagram illustrating the signaling interaction between the components in the embodiments of this application;

[0025] Figure 4 This is a schematic diagram of the structure of the causal strategy knowledge graph in the embodiments of this application.

[0026] Key reference numerals:

[0027] 10: Dialogue Agent; 20: Text Generation Module; 30: Semantic Guidance Controller; 40: Causal Attribution Module; 50: Knowledge Architect Module; 60: Causal Strategy Knowledge Graph; 110: Dialogue State Node; 120: Strategy Intent Node; 130: Business Entity Node; 140: Abstract Concept Node; 210: Causal Relationship Edge. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0029] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0030] For clarity, some of the nouns and terms used in the embodiments of this application will be explained below.

[0031] (1) Causal Strategy Knowledge Graph: This refers to a structured knowledge base used to store nodes and relationships related to dialogue strategies. Unlike traditional knowledge graphs, it not only contains entities and relationships, but is also specifically designed to encode and evolve causal relationships in dialogue strategies, such as the causal strength between a certain strategy action and a certain business result.

[0032] (2) High-level dialogue intent: refers to a relatively macro and abstract goal or direction in the dialogue process, rather than a specific, pre-set script or pre-set words. For example, "soothing the customer's emotions" or "introducing the product advantages" are high-level dialogue intents. They set the overall strategy for the next step of the dialogue, but leave room for creativity in the specific way of expression.

[0033] (3) Semantic constraints: These are the semantic rules and boundaries that must be followed when generating dialogue responses. They are usually used to ensure compliance or avoid inappropriate content. Unlike simple keyword blocking, they define prohibited content areas from the dimension of semantic similarity. For example, it prohibits the generation of any expression that is semantically similar to "promised benefits".

[0034] (4) Causal utility guidance: refers to the strategic elements or paths extracted from the causal strategy knowledge graph that have been historically proven to have a significant positive causal contribution to achieving the preset business goals. It plays a "pull" role in generating responses, guiding the generative model to tend towards historically more effective expressions.

[0035] (5) Causal contribution: refers to the net impact of a specific response phrase or strategy element on the achievement of the final business goal (such as user satisfaction, conversion rate) through causal inference or causal attribution analysis. It aims to strip away confounding factors and identify the true driving factors and their contribution.

[0036] (6) Self-organizing rewriting operation: refers to the process of dynamically adjusting and optimizing the structure of the causal strategy knowledge graph. This process is automatically executed by the system based on accumulated experience data. It is not just a simple addition or deletion of nodes or edges, but also includes complex structural evolution behaviors such as abstraction, generalization, pruning, and forgetting, aiming to achieve self-purification and iterative improvement of the knowledge system.

[0037] Reference Figures 1 to 4 This application provides a method and system for the self-organizing evolution of knowledge graphs based on causal attribution. This solution aims to address the problems in existing dialogue systems where dialogue strategies rely on a predefined action space, resulting in a lack of creativity, and the structural degradation of knowledge graphs due to their incremental updates.

[0038] like Figure 1As shown, the knowledge graph self-organizing evolution method based on causal attribution provided in this application (hereinafter referred to as the "method") is based on the construction of an intelligent closed loop from strategy generation, effect verification, causal attribution to structural reconstruction. This method includes multiple steps, forming a continuously iterative and optimized process. The knowledge graph self-organizing evolution method based on causal attribution provided in this application includes the following steps:

[0039] Step S101: Select a high-level dialogue intent; in response to the current dialogue state in the dialogue system, select a high-level dialogue intent from the preset causal strategy knowledge graph 60.

[0040] Step S102: Extract semantic constraints and causal utility guidance; Based on the current dialogue state and the selected high-level dialogue intent, extract a subgraph or path from the causal strategy knowledge graph 60. The subgraph or path contains predefined nodes and predefined edges, and the subgraph or path is used as semantic constraints and causal utility guidance.

[0041] Step S103: Control the generation of response dialogue; using these semantic constraints and causal utility guidance, generate response dialogue. In this process, the association between the response dialogue and the predefined node and the association between the response dialogue and the predefined edge are also recorded. The generation of response dialogue further includes: during the decoding process, pushing away candidate content that is similar to the prohibited semantics defined in the semantic constraints and bringing closer candidate content that is similar to the recommended content indicated by the causal utility guidance.

[0042] Step S104 "Execute the dialogue and record the results" and step S105 "Causal attribution analysis" apply the response script in the actual dialogue and perform causal attribution analysis on the complete dialogue trajectory containing the response script to determine the causal contribution of the response script to the preset business objectives.

[0043] Step S106, graph self-organizing rewriting; based on the calculated causal contribution, a self-organizing rewriting operation is performed on the causal strategy knowledge graph 60 to realize the structural evolution of the knowledge graph; wherein, the self-organizing rewriting operation includes at least one of the following: when multiple dialogue trajectories with similar causal contribution patterns to preset business objectives are identified, an abstract concept node representing the multiple dialogue trajectories is created in the causal strategy knowledge graph; and, the node or edge in the causal strategy knowledge graph that has a causal contribution below a preset threshold is reduced or removed, wherein the causal contribution of the node or edge is calculated based on the causal contribution of at least one response utterance generated using the node or edge.

[0044] As can be seen from the above, the closed-loop process of the method in this application enables the causal strategy knowledge graph 60 to continuously improve and evolve itself.

[0045] To implement the above method, embodiments of this application also provide a knowledge graph self-organizing evolution system based on causal attribution (hereinafter referred to as the "system"). Figure 2 As shown, this causal attribution-based knowledge graph self-organizing evolution system can be integrated into a dialogue system or work collaboratively with a dialogue system. This causal attribution-based knowledge graph self-organizing evolution system includes a selection module, an extraction module, a text generation module 20, a generation control module, a causal attribution module 40, and a graph optimization component. These modules and components work together to implement the entire process of the method.

[0046] Specifically, the function of the selection module corresponds to step S101 in the method. In practical applications, this function can be implemented by the dialogue agent 10. When the user terminal interacts with the dialogue agent 10, the dialogue agent 10 analyzes the current dialogue context, such as the user's latest input, historical dialogue records, user profile, etc., to determine the current dialogue state. Based on the current dialogue state, the dialogue agent 10 queries the causal strategy knowledge graph 60 and selects at least one high-level dialogue intent from it, such as selecting only the most suitable high-level dialogue intent. For example, when a user expresses hesitation about the price of a product, the dialogue agent 10 may select "handle price objections" as the most suitable high-level dialogue intent.

[0047] The extraction module corresponds to step S102 in the method. Based on the high-level dialogue intent determined by the selection module and the current dialogue state, the extraction module revisits the causal strategy knowledge graph 60. The extraction module retrieves subgraphs or paths in the causal strategy knowledge graph 60 that are relevant to the current context. These subgraphs or paths contain predefined nodes (e.g., nodes representing specific product characteristics or sales techniques) and predefined edges (edges representing logical or causal relationships between them). The extracted information is divided into two parts: one part serves as semantic constraints, defining the red lines for generated content; the other part serves as causal utility guidance, pointing towards high-value strategy directions. In the system architecture, this function can be performed by the dialogue agent 10 or a dedicated logic unit.

[0048] The generation control module and text generation module 20 work together to complete step S103 of the method. Figure 2In the example, the generation control module is specifically embodied as a semantic guidance controller 30. The semantic guidance controller 30 receives the semantic constraints and causal utility guidance provided by the extraction module and uses this information to regulate the behavior of the text generation module 20. The text generation module 20 is typically a Large Language Model (LLM). The semantic guidance controller 30 ensures that when generating the response speech, the text generation module 20 neither violates the red lines defined by the semantic constraints nor deviates too far from the direction indicated by the causal utility guidance. After generating the response speech, the generation control module also establishes and stores the association between the response speech and the nodes and edges of the causal strategy knowledge graph 60 on which the response speech was generated, providing a basis for subsequent causal attribution analysis. Specifically, during the decoding process (i.e., selecting the next word or token from the probability distribution) controlled by the text generation module 20, the semantic guidance controller 30 actively intervenes to push away candidate content similar to the prohibited semantics defined in the semantic constraints and bring closer candidate content similar to the recommended content indicated by the causal utility guidance.

[0049] The causal attribution module 40 corresponds to step S105 in this method. After the dialogue agent 10 completes at least one interaction with the user, a complete dialogue trajectory is formed, which includes all interaction content from beginning to end and the final business result (e.g., whether the user places an order, whether they are satisfied, etc.). The causal attribution module 40 is responsible for in-depth analysis of this dialogue trajectory. It uses techniques such as causal inference to accurately calculate the causal contribution of specific response phrases generated by the generation control module to the final business result throughout the entire dialogue process. For example, it may analyze that the phrase "mentioning risk hedging mechanisms" contributes +0.3 to the customer's final investment decision.

[0050] The function of the map optimization component corresponds to step S106 in this method. Figure 2In the example, the graph optimization component can be embodied as a knowledge architect module 50. The knowledge architect module 50 receives causal contribution data analyzed by the causal attribution module 40. Based on this data, it maintains and evolves the knowledge stored in the causal strategy knowledge graph 60. This is a self-organizing rewrite operation that adjusts the structure of the causal strategy knowledge graph 60 based on the accumulated causal contribution values ​​in at least one dialogue trajectory. For example, it might enhance the weights of nodes or edges that consistently generate high positive causal contributions, or adjust the topology of the causal strategy knowledge graph 60. The self-organizing rewrite operation based on this application includes at least one of the following: when multiple dialogue trajectories with similar causal contribution patterns to a preset business objective are identified, creating abstract concept nodes representing the multiple dialogue trajectories in the causal strategy knowledge graph; and reducing or removing nodes or edges in the causal strategy knowledge graph that have causal contributions below a preset threshold, wherein the causal contribution of the node or edge is calculated based on the causal contribution of at least one response utterance generated using the node or edge.

[0051] In a preferred embodiment, the self-organizing rewriting operation in the above embodiments is further refined. This operation is performed by the graph optimization component (such as...). Figure 2 The Knowledge Architect module (50) in the knowledge graph executes operations that include creating abstract concept nodes and performing structural pruning. These two operations correspond to the inductive generalization and the elimination and forgetting of knowledge, respectively, and together constitute the "metabolism" mechanism of the knowledge graph.

[0052] Specifically, when the Knowledge Architect module 50 analyzes a large number of dialogue trajectories and identifies multiple dialogue trajectories with similar causal contribution patterns to preset business goals, it creates a new abstract concept node 140 in the causal strategy knowledge graph 60 to represent these similar trajectories or strategies. For example, in the smart car sales scenario, the system discovers that three different specific strategies—"offering free full-car window tinting," "providing low-interest financing options," and "extending the warranty period"—all exhibit similar positive causal contribution patterns that significantly increase the signing rate when addressing customer objections to "high prices." The Knowledge Architect module 50 identifies the commonality in the functionality of these strategies, namely, "compensating for the customer's psychological price difference through added value without directly lowering the price," and creates an abstract concept node 140 named "Value Compensation Strategy" in the graph accordingly. Simultaneously, it establishes subordinate connections from this abstract concept node 140 to the intent nodes of the three specific strategies mentioned above. This operation achieves knowledge induction and generalization.

[0053] On the other hand, the self-organizing rewriting operation also includes reducing or removing nodes or edges in the causal strategy knowledge graph 60 that have a causal contribution below a preset threshold. The causal contribution of a node or edge is calculated based on the cumulative causal contribution of at least one response phrase generated using that node or edge, and may be combined with factors such as usage frequency and time decay. For example, in a smart investment advisory scenario, a strategy initially considered effective, "listing historical annual returns of products," may, as the system runs, find that the causal attribution module 40 continuously discovers that the causal contribution of this strategy to customer signing is very low, or even negative (e.g., customer churn due to information overload). The graph optimization component continuously reduces the utility weight of the strategy node and its related edges based on these negative attribution results. When the weight falls below a preset "forgetting threshold," the graph optimization component automatically performs a pruning operation, removing the node and its associated edges from the causal strategy knowledge graph 60. This operation achieves self-purification of knowledge.

[0054] The technical effect of the aforementioned self-organizing rewriting operation lies in its significant improvement from "incremental growth" to "self-organizing evolution." By introducing a rewriting mechanism that includes abstraction, generalization, and pruning, the knowledge graph can self-purify, iterate, and improve structural efficiency. The policy abstraction function enables the system to learn by analogy and creatively apply inductively derived high-level strategies to generate entirely new solutions. The structural pruning function fundamentally prevents the problems of knowledge entropy increase and structural degradation caused by long-term operation, ensuring the conciseness and efficiency of the knowledge base and maintaining the high performance of system decision-making.

[0055] like Figure 3 As shown, the interactions between the system's components are clearly defined in terms of time. The interaction between the user and the dialogue system is real-time. After a user sends a message, the dialogue system quickly completes a series of real-time processes, such as intent selection and script generation, and replies. After each interaction, the dialogue system stores information such as the dialogue trajectory and business results in the experience database. In contrast, the knowledge architect's work is asynchronous. It periodically reads accumulated data from the experience database and performs complex causal attribution analysis and graph rewriting operations. These operations are computationally intensive and unsuitable for real-time interaction. After completing the rewriting operation, the knowledge architect pushes the updated causal strategy knowledge graph to the dialogue system for use. This architecture, separating real-time and asynchronous operations, ensures smooth user interaction while enabling the system to utilize sufficient data for deep thinking and self-evolution.

[0056] The working principle of this method and system lies in transforming the relatively static knowledge graph of traditional dialogue systems into a dynamic, self-evolving entity. By introducing causal attribution, the system's learning is no longer based on simple correlation statistics (e.g., after saying A, B occurs more frequently), but rather delves into the true causal effect of A on B. This makes knowledge accumulation and optimization more accurate and robust. Simultaneously, by shifting the decision-making process from "choosing from a finite set of actions" to "creating under semantic constraints and causal guidance," the system gains the ability to generate novel and effective strategies, greatly enhancing its flexibility and creativity in dealing with open and dynamic environments.

[0057] The aforementioned basic approach achieves significant technical benefits. First, it realizes a paradigm shift from "limited choice" to "autonomous generation." The system is no longer confined to preset scripts or strategies, but can create diverse, highly contextualized response scripts that ensure deep compliance based on real-time semantic and causal utility guidance. This enables the immediate generation of novel solutions to address complex and ever-changing dialogue scenarios. Second, it constructs a deeply coupled causal-driven closed loop, deeply integrating causal inference into every stage of reward attribution and knowledge graph reconstruction. This ensures that every evolution of the system is driven by explicit, quantifiable causal relationships, rather than vague correlations. This helps improve the accuracy and efficiency of knowledge evolution, reliably guaranteeing the system's long-term performance.

[0058] Furthermore, such as Figure 4 As shown, the structure of the causal strategy knowledge graph 60 includes causal relationship edges 210 for storing probabilistic causal relationships, and abstract concept nodes 140 dynamically created through self-organizing rewrite operations. Other types of nodes are also shown in the graph, such as dialogue state nodes 110 (e.g., "customer inquires about high-risk products"), strategy intent nodes 120 (e.g., "mentioning risk hedging mechanisms"), and business entity nodes 130 (e.g., the name of a specific financial product).

[0059] Among them, the causal edge 210 not only represents the connection relationship between nodes, but also stores quantified, probabilistic causal relationships. For example, the weight or attribute of the causal edge 210 connecting the dialogue state node 110 and the policy intent node 120 can be represented as "under the dialogue state node 110, executing the policy intent node 120 has a 70% probability of producing a positive causal contribution of 0.5 units to the business objective". This probabilistic representation allows causal attribution and graph rewriting operations to be based on a solid mathematical foundation.

[0060] like Figure 4As shown, abstract concept nodes 140 can be connected to at least one more specific policy intent node 120. These nodes are not predefined manually at the initial stage, but are automatically created by the graph optimization component based on experience during system operation. This ability to dynamically create abstract concepts allows the graph's ontology itself to evolve, rather than simply adding or removing instances. Therefore, abstract concept nodes 140 embody the dynamic nature of the graph structure.

[0061] This graph structure, comprising probabilistic causal relationship edges 210 and dynamic abstract concept nodes 140, provides structural support for the system's self-organizing evolution. Probabilistic causal relationships make all decisions and evolutionary operations more precise and interpretable. The ability to dynamically create abstract concepts endows the knowledge system with inductive and generalizing capabilities.

[0062] In another preferred embodiment, step S103 of generating the response script in the above embodiment is further refined. This step is controlled by a generation control module (such as...). Figure 2 The semantic guidance controller 30 and the text generation module 20 work together to complete this process. This process can be implemented using various techniques, such as constrained beam search. During each decoding step, for multiple candidate words provided by the text generation module 20, the generation control module 30 calculates the similarity between the current sequence formed by each candidate word and the prohibited semantics. If the similarity exceeds a certain threshold, the score of that candidate path will be significantly reduced, equivalent to being "pushed away." Simultaneously, it also calculates the similarity between the current sequence and the recommended content, and rewards candidate paths with high similarity, equivalent to being "pulled closer."

[0063] The term "similarity" as used in this application refers to the semantic closeness between two text fragments (e.g., the currently generated candidate phrase and the preset prohibited semantic text, or the candidate phrase and the recommended content text). Similarity can be calculated using various semantic similarity calculation methods known in the art, including but not limited to: methods based on vector space models (such as converting text into vectors using a bag-of-words model or TF-IDF weighting and then calculating cosine similarity, Euclidean distance, or Manhattan distance), methods based on distributed representations (such as encoding text into semantic vectors using pre-trained language models like Word2Vec, GloVe, or BERT and then calculating cosine similarity or dot product similarity between vectors), and text similarity calculation methods based on word overlap or edit distance. Those skilled in the art can select one or more suitable methods to calculate the similarity based on the computational resource requirements and accuracy needs of the actual application scenario.

[0064] Regarding the "similarity judgment criteria," this application adopts a threshold comparison method for determination. Specifically, a first similarity threshold is pre-configured for the semantic constraints, and a second similarity threshold is pre-configured for the causal utility guidance. During the decoding process, when the similarity between candidate content and prohibited semantics exceeds the first similarity threshold, the candidate content is determined to be "similar" to the prohibited semantics, triggering a push-away operation; when the similarity between candidate content and recommended content exceeds the second similarity threshold, the candidate content is determined to be "similar" to the recommended content, triggering a close-away operation. The first and second similarity thresholds can be empirical values, fixed values ​​calibrated through experiments, or adaptive thresholds that can be dynamically adjusted according to dialogue state, user profile, or system performance.

[0065] For example, in a robo-advisor scenario, the semantic constraint is "prohibiting promises of returns," while the causal utility guidance is "emphasizing risk hedging mechanisms." When words like "guarantee" or "sure profit" appear in the candidate content of the text generation module 20, the generation control module 30 will identify that they are highly semantically similar to "promised returns" and suppress these candidate paths. When words like "hedging" or "dynamic portfolio rebalancing" appear in the candidate content, the generation control module 30 will identify that they are semantically related to "risk hedging mechanisms" and increase the priority of these candidate paths.

[0066] This technology, which performs real-time push-pull control during the decoding process, achieves deeper and more proactive control over the generated content. Compared to solutions that filter content after generating the complete script, this solution fundamentally avoids the generation of non-compliant content and more effectively guides the generation process towards high-value areas. It ensures that the final output script is not only compliant, but also that its content and expression have been optimized for causal effectiveness, thereby improving the quality and effectiveness of the script at its source.

[0067] Furthermore, to optimize the evolutionary capability of the entire system itself, this application also provides a meta-learning-based optimization mechanism. This mechanism uses meta-learning to optimize graph optimization components (such as...) Figure 1 The knowledge architect module 50 in the graph is optimized by performing a self-organizing rewrite operation on the causal strategy knowledge graph. In this mechanism, the self-organizing rewrite operation performed by the graph optimization component (e.g., deciding when to abstract, when to prune, and the parameters of the operation such as pruning thresholds) is regarded as a "meta-action".

[0068] In the system architecture, a meta-learning module can be added, which can be integrated within the Knowledge Architect module 50. The meta-learning module is responsible for learning meta-policies. The input to this policy is the current state of the knowledge graph, the statistical characteristics of experience data, etc., and the output is a specific meta-action. After the system executes this meta-action, the meta-learning module observes and calculates the overall performance change of the entire dialogue system over a relatively long time window (e.g., one week or one month), such as the slope of the average business reward increase and the month-on-month growth rate of user satisfaction. This performance change is used as the "meta-reward" for the meta-action.

[0069] For example, the meta-learning module faces a decision: given five successful paths with similar causal patterns, should it immediately create an abstract node (meta-action A), or observe and accumulate 10 more similar paths before creating it (meta-action B)? Suppose the meta-policy chooses meta-action A. After one month, system evaluation shows that executing this action improved the dialogue system's average conversion rate by 5% month-over-month. This "5% improvement" is taken as the meta-reward for meta-action A. By continuously repeating the process of "trying meta-actions - observing meta-rewards," the meta-learning module can use reinforcement learning algorithms (such as Proximal Policy Optimization, PPO) to update its meta-policy network, thereby learning which graph rewriting operation is most efficient in different situations.

[0070] The technical effect of this meta-learning optimization mechanism lies in its ability to enable the system to learn and optimize its own evolutionary strategies. The system in this application can not only learn specific dialogue strategies (first-order learning), but also learn how to more effectively learn and evolve its knowledge structure (second-order learning, i.e., meta-learning). This gives the graph optimization component the ability to self-optimize, dynamically adjusting its evolutionary strategies to adapt to different stages of system development (e.g., more aggressively abstracting strategies in the early stages of system operation to accelerate knowledge formation, and focusing more on refined pruning to improve efficiency in the mature stage of the system), thereby maximizing the overall long-term performance of the system.

[0071] The following embodiment, which integrates all the above-mentioned technical features, will fully demonstrate the workflow and overall technical effects of this application. This embodiment uses an intelligent sales assistant as an example, which is deployed for online sales of a high-end electronic product.

[0072] The core of this system is based on Figure 2 The architecture shown is constructed as follows. The structure of its causal strategy knowledge graph 60 is as follows: Figure 4As shown, it includes dialogue state nodes 110, policy intent nodes 120, business entity nodes 130, dynamically created abstract concept nodes 140, and causal relationship edges 210 storing probabilistic causal relationships. The system's graph optimization component, namely the knowledge architect module 50, not only performs abstraction and pruning operations, but is also optimized by a meta-learning module. The generation of the system's response dialogue is managed by the semantic guidance controller 30 through push-pull control during the decoding process, which manages the text generation module 20.

[0073] In a specific interactive scenario, a potential customer (user) says while interacting with the intelligent sales assistant (conversation system): "This product is more expensive than I expected."

[0074] The system's workflow is as follows:

[0075] 1. Selecting a high-level intent: After receiving user input, the dialogue agent 10 analyzes the data and determines the current dialogue state as "the customer has raised a price objection." Based on this state, it queries the causal strategy knowledge graph 60 and selects "handle the price objection" as the high-level dialogue intent.

[0076] 2. Extraction of Semantic Constraints and Causal Utility Guidance: Based on the high-level dialogue intent and the current dialogue state, the system extracts information from the causal strategy knowledge graph 60. Semantic constraints are identified as "prohibited semantics": "direct price reduction," "belittling competitors." Causal utility guidance points to an abstract concept node 140 named "non-price value compensation." This abstract concept node was automatically created by the system through self-organizing rewriting operations in the past. It connects multiple historically proven specific strategies, such as "emphasizing after-sales service," "giving away accessories," and "providing interest-free installments." This abstract concept node itself carries a high weight for historical causal contribution.

[0077] 3. Controlled Generation of Response Script: The semantic guidance controller 30 receives the aforementioned semantic constraints and causal utility guidance. It controls the text generation module 20 to begin generating response scripts. During the decoding process, when the candidate sequence contains "We can give you a 10% discount," the semantic guidance controller 30 identifies its semantic similarity to "direct price reduction" and suppresses it. Simultaneously, it increases the weight of semantic directions related to the abstract concept of "non-price value compensation." Combining the current customer profile (e.g., the customer previously mentioned concern about battery life), the system ultimately generates a novel, combined response script: "Sir, I understand your feelings. Although our prices are uniform nationwide, to address your concerns about battery life, we can apply for a dedicated accessory package for you that includes an original spare battery and a fast charger. This package is worth over 500 yuan. Would that be acceptable to you?" This script is creative because it combines abstract guidance with a specific context.

[0078] 4. Perform the dialogue and record the result: The customer is satisfied with the proposal and ultimately completes the purchase. The dialogue system records the entire dialogue process, as well as the positive business outcome of "transaction successful," and stores this data in the experience database.

[0079] 5. Causal Attribution Analysis: During the asynchronous processing phase, the causal attribution module 40 analyzed this successful dialogue trajectory. It used a causal inference model to calculate that the specific phrase "propose a customized accessory package" contributed +0.8 to the "successful transaction," and further attributed this contribution to its superior abstract concept node 140, "non-price value compensation," and the graph path on which this generation was based.

[0080] 6. Graph Self-Organizing Rewriting: After receiving the results of the causal attribution analysis, the Knowledge Architect module 50 performs a rewriting operation. First, it strengthens the weight of the causal edge 210 from the "customer raises price objection" state to the "non-price value compensation" node. Second, it analyzes another historical strategy, "repeatedly emphasizing product performance," and finds that its causal contribution in multiple "price objection" scenarios is consistently negative, and its cumulative weight has fallen below the preset pruning threshold. Therefore, it removes this node from the graph.

[0081] 7. Meta-learning Optimization: The meta-learning module observed that the Knowledge Architect module 50 recently adopted a meta-strategy of "creating an abstract node when there are 5 similar causal pattern paths." It evaluated that within one month of implementing this meta-strategy, the overall system's success rate in handling price objections improved by 3%. This "3% improvement" was used as a positive meta-reward to strengthen the current meta-strategy, making it more inclined to perform knowledge abstraction quickly in the future.

[0082] Through the complete workflow integrating all the technical features described above, this embodiment achieves significant comprehensive technical effects. Firstly, regarding creativity and compliance, by combining abstract concept guidance and push-pull control during decoding, the system can not only generate highly personalized solutions never before seen in the database (such as accessory packages addressing battery life concerns), but also completely eliminate non-compliant expressions at the semantic level, achieving a unity of "autonomous generation" and "helping to ensure deep compliance." Secondly, regarding the health of the knowledge system, through the synergy of abstraction and pruning operations, the causal strategy knowledge graph 60 achieves efficient "metabolism," continuously generating higher-level intelligence (such as "non-price value compensation" strategies) through inductive generalization, while promptly eliminating outdated and ineffective knowledge, maintaining a streamlined and efficient structure. Finally, regarding long-term evolutionary capability, the meta-learning mechanism enables the system to "learn how to learn," dynamically adjusting its evolutionary strategy according to its development stage, ensuring continuous and optimal growth in long-term system performance. The synergistic effect of these results makes the dialogue system provided in this application significantly improved in terms of intelligence, adaptability, and long-term benefits compared to existing technologies.

[0083] The methods and systems proposed in this application have a wide range of applications and can be applied to various fields that require dialogue and user interaction to achieve specific business goals. Their core advantage lies in their ability to transform a dialogue system from a tool that follows fixed rules into an intelligent partner capable of self-evolution, continuous learning, and creation.

[0084] In the field of intelligent customer service, this application can be used to handle complex customer complaints and issues. Traditional customer service robots often rely on knowledge base matching, which is insufficient for effectively handling issues beyond their scope. By applying this application, the system can creatively propose solutions through multiple rounds of interaction with users, while adhering to compliance requirements (e.g., not making promises beyond its authority). Through causal attribution of the solution's effectiveness, the system can continuously learn which soothing techniques and explanations are most effective in resolving specific types of complaints, and gradually abstract these successful experiences into new strategies, thereby continuously improving the resolution rate of complex problems and user satisfaction.

[0085] In the field of intelligent sales and marketing, this application helps improve sales conversion rates. As shown in the aforementioned embodiments, whether recommending financial products or selling physical goods, the system can flexibly address various objections and concerns from customers. It no longer simply recites product manuals, but is guided by causal utility, proactively mentioning features or value points that have historically proven to be most effective in resonating with similar customers. Through self-organizing evolution, the sales strategy knowledge graph becomes increasingly adept at understanding customers, and may even generate effective combinations of sales strategies.

[0086] In the fields of education and personalized tutoring, this application can be used to build an adaptive intelligent tutoring system. Students' answers and questions constitute the dialogue trajectory, while "mastering knowledge points" is the business objective. The system can generate different explanation methods, questioning styles, or motivational phrases, and analyze the true impact of these teaching interactions on student learning outcomes through causal attribution analysis. For example, the system might discover that for a certain type of student, "giving a real-life example" has a much higher causal contribution than "reciting a definition." Based on this, the system will rewrite its teaching strategy map, forming a causally validated personalized teaching methodology for different student groups.

[0087] In the field of healthcare consultation, compliance is paramount. The semantic constraints and decoding control mechanisms of this application provide an effective safeguard, ensuring that all information provided by virtual doctors or health assistants strictly adheres to medical regulations and guidelines. Furthermore, its evolutionary capabilities enable the system to learn from interactions with a large number of patients how to more effectively conduct health education, medication reminders, and lifestyle recommendations. For example, the system can analyze the causal contribution of different communication styles to improving patient medication adherence and optimize its communication strategies, thereby improving patient health outcomes.

[0088] Furthermore, this application also has broad application prospects in professional fields such as human-machine collaboration, psychological counseling, and legal consultation. It enables machines to interact with humans not merely as information retrieval tools, but as intelligent agents capable of understanding causality, creating, and continuously improving themselves, thus playing a more important role in broader and more complex tasks.

[0089] Those skilled in the art will understand that the various modules in the system of the above embodiments can be implemented by a computer program, which can be stored in a computer-readable storage medium and, when executed, performs the steps of the method of the above embodiments; and the various modules can also be implemented by hardware such as integrated circuits.

[0090] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A self-organizing evolution method for knowledge graphs based on causal attribution, characterized in that, include: In response to the current dialogue state in the dialogue system, select high-level dialogue intents from the causal policy knowledge graph; Based on the current dialogue state and the high-level dialogue intent, extract a subgraph or path containing predefined nodes and predefined edges from the causal strategy knowledge graph as semantic constraints and causal utility guidance. Using the semantic constraints and the causal utility guidance, a response dialogue is generated, and the association between the response dialogue and the node and the association between the response dialogue and the edge are recorded. The generation of response dialogue further includes: during the decoding process, pushing away candidate content that is similar to the prohibited semantics defined in the semantic constraints, and bringing closer candidate content that is similar to the recommended content indicated by the causal utility guidance. A causal attribution analysis is performed on the dialogue trajectory containing the response script to determine the causal contribution of the response script to the preset business objectives. Based on the causal contribution, a self-organizing rewriting operation is performed on the causal strategy knowledge graph, wherein the self-organizing rewriting operation includes at least one of the following: When multiple dialogue trajectories with similar causal contribution patterns to preset business objectives are identified, abstract concept nodes representing these multiple dialogue trajectories are created in the causal strategy knowledge graph; and, Reduce or remove nodes or edges in the causal strategy knowledge graph that have a causal contribution below a preset threshold, wherein the causal contribution of the node or edge is calculated based on the causal contribution of at least one response phrase generated using the node or edge.

2. The method according to claim 1, characterized in that, The structure of the causal strategy knowledge graph includes: causal relationship edges for storing probabilistic causal relationships, and abstract concept nodes dynamically created through the self-organizing rewrite operation.

3. The method according to claim 1, characterized in that, The process of pushing away candidate content that is semantically similar to prohibited content and bringing closer candidate content that is similar to recommended content during decoding is achieved using constrained beam search.

4. The method according to claim 1, characterized in that, Also includes: The causal strategy knowledge graph is optimized by performing a self-organizing rewrite operation through meta-learning, wherein the self-organizing rewrite operation is regarded as a meta-action, and the change in the overall performance of the dialogue system after the operation is performed is used as a meta-reward.

5. The method according to any one of claims 1-4, characterized in that, Also includes: The dialogue trajectory is stored in an experience database; the self-organizing rewriting operation on the causal strategy knowledge graph is performed asynchronously and periodically based on the data accumulated in the experience database.

6. A self-organizing evolutionary system for knowledge graphs based on causal attribution, characterized in that, For implementing the method as described in any one of claims 1-5, comprising: The selection module is used to select high-level dialogue intents from the causal policy knowledge graph in response to the current dialogue state in the dialogue system. The extraction module is used to extract a subgraph or path containing predefined nodes and predefined edges from the causal strategy knowledge graph based on the current dialogue state and the high-level dialogue intent, as semantic constraints and causal utility guidance. Text generation module; A generation control module is used to control the text generation module using the semantic constraints and the causal utility guidance to generate response statements and establish the association between the response statements and the nodes, as well as the association between the response statements and the edges. The generation control module is configured to, during the decoding process, push away candidate content that is similar to the prohibited semantics defined in the semantic constraints and bring closer candidate content that is similar to the recommended content indicated by the causal utility guidance. The causal attribution module is used to perform causal attribution analysis on the dialogue trajectory containing the response words in order to determine the causal contribution of the response words to the preset business objectives. A graph optimization component is used to perform a self-organizing rewriting operation on the causal strategy knowledge graph based on the causal contribution, wherein the self-organizing rewriting operation includes at least one of the following: When multiple dialogue trajectories with similar causal contribution patterns to preset business objectives are identified, abstract concept nodes representing these multiple dialogue trajectories are created in the causal strategy knowledge graph; and, Reduce or remove nodes or edges in the causal strategy knowledge graph that have a causal contribution below a preset threshold, wherein the causal contribution of the node or edge is calculated based on the causal contribution of at least one response phrase generated using the node or edge.

7. The system according to claim 6, characterized in that, Also includes: A storage module is used to store the causal strategy knowledge graph, the structure of which includes: causal relationship edges for storing probabilistic causal relationships; And, abstract concept nodes dynamically created through the self-organizing rewrite operation.

8. The system according to claim 6, characterized in that, The generation control module is configured to use constrained beam search to push away candidate content that is semantically similar to prohibited content and bring closer candidate content that is similar to recommended content during the decoding process.

9. The system according to claim 6, characterized in that, Also includes: The meta-learning module is used to optimize the graph optimization component through meta-learning, wherein the self-organizing rewrite operation is regarded as a meta-action, and the change in the overall performance of the dialogue system after the operation is executed is used as the meta-reward.

10. The system according to any one of claims 6 to 9, characterized in that, Also includes: An experience database is used to store the dialogue trajectories; the graph optimization component is configured to asynchronously and periodically perform self-organizing rewriting operations on the causal strategy knowledge graph based on the data accumulated in the experience database.

Images