A method and system for generating a motion sequence and evolving behavior based on a motion knowledge graph for an embodied dialogue agent
By constructing an action knowledge graph and a feedback-driven action sequence generation method, the problem of inconsistent action sequences in embodied dialogue agents is solved, achieving naturalness and controllability in action generation and improving long-term adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF AUTOMATION CHINESE ACAD OF SCI
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies lack unified communicative intent constraints at the whole-round response level in embodied dialogue agents, resulting in inconsistent action sequences, weak semantic control capabilities, and difficulty in continuously optimizing behavioral expressions based on interactive feedback.
An action knowledge graph is constructed, and a unified semantic purpose of the behavior is determined by the entire round of response text. An initial action sequence is generated, and the edge weights and instance representation library of the action knowledge graph are updated after feedback to achieve continuous optimization of the action sequence.
It enhances the naturalness, controllability, and long-term adaptability of embodied dialogue agents' action generation, ensuring that action sequences match semantic content and character style.
Smart Images
Figure CN122452604A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent interaction, digital human-driven systems, graph modeling, natural language processing, and machine learning, and in particular to a method and system for generating action sequences and evolving behaviors based on action knowledge graphs for embodied dialogue agents. Background Technology
[0002] In applications such as service guidance, digital companionship, intelligent question answering, and virtual interaction, embodied dialogue agents typically need to simultaneously output body movements, gestures, and other non-verbal behaviors that match the semantics and personality of the text while generating response text. Existing technologies employ two main approaches: one type matches actions independently to each sentence fragment, lacking a unified communicative intent constraint across the entire response round. This can easily lead to inconsistencies in action expression within the same response round, affecting the naturalness of action sequence transitions and overall semantic consistency. Another type relies primarily on end-to-end black-box generation, which can directly generate action results, but its semantic control is weak, the generation process lacks interpretability, and it is difficult to stably adjust for specific interaction needs. Furthermore, existing technologies underutilize explicit and implicit feedback during the interaction process, making it difficult to continuously update action selection strategies and optimize behavioral expression effects over long-term use.
[0003] Furthermore, action generation during interaction needs to consider not only the matching of current local semantics with actions, but also the natural connection between previous and subsequent actions, consistency of character style, and the diversity of actions across timelines. Without a unified model of action transfer relationships and feedback evolution mechanisms, issues such as repetitive actions, awkward transitions, unstable styles, or difficulty in continuous optimization can easily arise.
[0004] Therefore, it is necessary to propose a method for action sequence generation and behavior evolution of embodied dialogue agents, ensuring the communicative intent of the entire round of responses. Figure 1 Based on this, a unified model is used for action sequence generation and feedback updates, thereby improving the naturalness, controllability and long-term adaptability of action generation. Summary of the Invention
[0005] To address the problems of insufficient semantic adaptability, poor controllability, and difficulty in continuous optimization based on feedback in existing technologies during long-term interactions, this invention proposes a method and system for action sequence generation and behavior evolution based on action knowledge graphs for embodied dialogue agents.
[0006] The core idea of this invention is as follows: First, an action knowledge graph is constructed; then, a unified behavioral semantic purpose is determined based on the entire round of response text, and the initial action is searched and the complete action sequence is expanded under the constraint of this semantic main line; finally, the feedback is quantified after the action is executed, and the edge weights of the action knowledge graph and the instance representation library are updated according to the feedback results, so that the system can continuously evolve in long-term interaction.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents includes the following steps:
[0009] Step S1: Construct an action knowledge graph based on multi-turn user-agent dialogue data, action resources, and annotation results. The action knowledge graph includes behavior semantic category nodes, action nodes, semantic-action adaptation edges, and action-action transition edges. An instance representation library is established for each behavior semantic category node, and an action description library is established for each action node.
[0010] Step S2: Extract features from the text in the instance representation library using a language model, map them to a unified semantic vector space, and store the vector results;
[0011] Step S3: Obtain the complete response text for the current round, and perform semantic similarity matching by using the semantic similarity between the complete response text for the current round and the instance expression text under each behavioral semantic category to obtain the unified behavioral semantic purpose for the current round;
[0012] Step S4: Based on the unified behavioral semantic purpose, recall candidate initial actions along the semantic-action adaptation edge, and combine the semantic matching degree between the first sentence fragment at the beginning of the current round reply and the action description, the matching degree between the current character description and the character applicable to the action, the action-action transfer edge connection relationship and edge weight between the last action of the previous round and the candidate initial action, and the cross-round repetition penalty of the candidate initial action relative to the last action of the previous round, sort the candidate initial actions to obtain multiple initial action candidates;
[0013] Step S5: Divide the complete reply text of the current round into several sentence fragments. Starting from the initial action candidate, perform graph search based on the local semantic matching degree between the sentence fragment and the action description, the action-action transition edge weight, the matching degree between the action and the current character description, and the intra-sequence repetition penalty to generate multiple sets of candidate action sequences. Then, perform global adjudication on the candidate action sequences to obtain the final action sequence and the starting position of the semantic fragment corresponding to each action.
[0014] Step S6: Based on the final action sequence and the starting position of the semantic fragment, output the text starting position and time starting position of each action in the current round of complete response text, and generate timing instructions that can directly drive the digital human to perform actions.
[0015] A system for action sequence generation and behavior evolution based on action knowledge graphs for embodied dialogue agents includes:
[0016] The action knowledge graph construction module is used to collect multi-turn user-agent dialogue samples and action resources, label each round of response with a unified behavioral semantic purpose, and segment the response text into several sentence fragments; on this basis, an action knowledge graph including behavioral semantic category nodes, action nodes, semantic-action adaptation edges and action-action transition edges is constructed, and an instance expression library and an action description library are established, and the instance expression text is pre-encoded offline.
[0017] The action sequence generation module, after obtaining the complete response text of the current round, first determines the unified semantic purpose of the behavior by matching the similarity between the instance representation library and the complete response text of the current round. Then, based on the unified semantic purpose of the behavior, it recalls candidate actions from the corresponding semantic category node along the semantic-action adaptation edge. Combining the matching degree between the first sentence fragment of the current round response and the action description, the current persona description, the action-action transfer edge connection relationship and edge weight between the last action of the previous round and the candidate initial action, and the cross-round repetition penalty, the candidate actions are filtered and sorted to obtain multiple initial action candidates. Starting from the multiple initial action candidates, the module expands position by position to obtain multiple sets of candidate action sequences by combining the local semantic matching between the sentence fragment and the action description, the action-action transfer edge weight, the persona matching, and the intra-sequence repetition penalty. Finally, the language model makes a global decision under the constraints of the current round text and the candidate action sequence group to obtain the final action sequence and the starting position of the semantic fragment corresponding to each action.
[0018] The alignment and driving module is used to determine the text start position of each action in the current round of complete response text based on the final action sequence and the start position of the semantic fragment corresponding to each action, and to obtain the corresponding time start position by combining the voice alignment timestamp, generate action execution instructions, and drive the digital human to execute.
[0019] A computing device includes: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device performs the action sequence generation and behavior evolution method based on action knowledge graph for embodied dialogue agents.
[0020] The beneficial effects of the present invention are as follows: The above-mentioned method of the present invention enables the embodied dialogue agent to generate action sequences that match the semantic content and role style when generating response text, and can continuously optimize the action generation strategy based on interaction feedback, thereby improving the naturalness, smoothness and long-term adaptability of action expression. Attached Figure Description
[0021] Figure 1 This is an overall flowchart of the method and system for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue intelligent agents, which is the subject of this invention. Detailed Implementation
[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0023] This invention provides a method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents. Figure 1 A schematic flowchart of the method is shown. (e.g.) Figure 1 As shown, the method includes:
[0024] Step S1: Offline Construction of Action Knowledge Graph. In this step, multi-turn user-agent dialogue data and action resources are first collected and labeled in units of one round of responses. A unified behavioral semantic purpose is determined for each round of responses, and the response text is segmented into several sentence fragments to determine the length of subsequent action sequences and the corresponding local text content at each position. Then, an action knowledge graph is constructed, which includes at least two types of nodes and two types of edges: the two types of nodes are behavioral semantic category nodes and action nodes; the two types of edges are semantic-action adaptation edges and action-action transition edges, where the initial edge weight is defined as 1.0 as a neutral baseline weight. Behavioral semantic category nodes are used to represent stable high-level communicative purposes, and action nodes are used to represent executable actions; semantic-action adaptation edges are used to represent a set of actions adapted to a certain semantic category, and action-action transition edges are used to represent natural transition relationships between actions. In addition, an action description library is established for action nodes, and an instance representation library is established for behavioral semantic category nodes.
[0025] In one embodiment, the action description library may include information such as applicable semantic categories, applicable scenarios, applicable character descriptions, keywords and typical expressions, and style tags, to support semantic filtering and local matching of candidate actions. The instance expression library stores high-quality representative text expressions corresponding to each action semantic category, for subsequent full-round semantic judgment and incremental library writing.
[0026] Step S2: Perform offline precoding on the instance representation library. In this step, the instance representation texts under each behavioral semantic category are used to extract features using a language model, mapped to a unified semantic vector space, and the vector results are stored. Preferably, a fixed number of representative expressions can be retained for each behavioral semantic category, and vector precoding is completed offline to reduce the computational load during online execution. The offline precoding results are used for both online semantic judgment and feedback update phases for near-duplication detection and difference filtering of newly added expressions.
[0027] Step S3: Perform full-round semantic determination. In this step, the system receives the complete response text of the current round and performs semantic similarity matching with the instance expression text under each behavioral semantic category to obtain the unified behavioral semantic purpose of the current round, ensuring consistency in the main communicative line of the entire round of responses. This invention first determines a unified behavioral semantic purpose at the full-round level, and then expands the search for subsequent actions under the constraint of this main line. This process may include: mapping the complete response text of the current round and the instance expression text under each behavioral semantic category to the same semantic vector space, calculating the similarity between the texts, and determining the unified behavioral semantic purpose of the current round based on the aggregated similarity of several instance expressions in each behavioral semantic category that are most similar to the complete response text of the current round. In one embodiment, several instance expressions in each behavioral semantic category that are most similar to the complete response text of the current round can be aggregated, and the behavioral semantic category with the highest aggregated similarity can be used as the unified behavioral semantic purpose of the current round.
[0028] Step S4: Perform initial action search. In this step, based on the unified behavioral semantic purpose obtained in Step S3, the system recalls candidate initial actions from the corresponding behavioral semantic category node along the semantic-action adaptation edge, forming an initial action pool. Subsequently, the system further refines and sorts the candidate actions in the initial action pool by considering the semantic matching degree between the first sentence fragment of the current round's response and the action description, the matching degree between the current persona description and the persona applicable to the action, the existence of an action-action transition edge and its weight between the last action of the previous round and the current candidate initial action, and the cross-round repetition penalty of the current candidate initial action relative to the last action of the previous round, resulting in multiple initial action candidates. Through this design, the semantic node is responsible for limiting the search scope, thereby reserving multiple high-quality starting points for the subsequent generation of complete action sequences.
[0029] In one embodiment, candidate initial action The scoring function can be expressed as:
[0030] ,
[0031] in, This represents the first sentence segment of the current round's complete response text after it has been segmented. This indicates the current character setting. This indicates the last action in the sequence of actions executed in the previous round. This represents the semantic matching score between the first sentence fragment at the beginning of the current round of responses and the candidate initial action description. This indicates the matching score between the candidate's initial action and the current character description. This represents the transition adaptation score between the last action of the previous round and the current candidate initial action. This indicates a penalty for repeated actions across rounds. , , , These are the corresponding weighting coefficients.
[0032] In one embodiment, the semantic matching score between the first sentence fragment at the beginning of the current round of responses and the candidate initial action description is... This can be obtained in the following way: First, use a pre-trained language model to extract the first sentence fragment of the current round of replies. The semantic feature vectors of the candidate initial action description text are compared, and then the relationship between the two is calculated. The distance is calculated and converted into a matching score using a preset mapping function; the smaller the semantic distance, the higher the corresponding matching score.
[0033] In one embodiment, the matching score between the candidate initial action and the current character description This can be obtained in the following way: First, use a pre-trained language model to extract the current persona description. The semantic feature vectors of the character adaptation description text corresponding to the candidate initial action are then calculated, and the relationship between the two is then determined. The distance is calculated and mapped to a matching score; the more consistent the candidate's initial action is with the current character description, the higher the corresponding score.
[0034] In one embodiment, the transfer adaptation score Based on the action knowledge map, the last action of the previous round can be used. With the current candidate initial action The existence of an action-action transition edge and the weight of that transition edge are determined. When a transition edge exists between the two and its weight is high, it indicates a high degree of adaptation from the last action of the previous round to the current candidate initial action, and the corresponding transition adaptation score is also high. The specific value of "higher", the mapping relationship between the transition edge weight and the degree of adaptation and the transition adaptation score can be preset by those skilled in the art according to the application scenario.
[0035] In one embodiment, the cross-wheel repetition penalty term It can be determined based on whether the current candidate initial action is the same as the last action of the previous round; the higher the degree of repetition of the candidate actions, the greater the corresponding penalty value.
[0036] In one embodiment, the weighting coefficient , , , The relative influence of semantic matching of the first segment, character matching, action transfer adaptation, and cross-round repetition penalty on the initial action score can be adjusted by experience setting, manual parameter tuning, or learning based on training data.
[0037] Step S5: Perform action sequence expansion and global adjudication. In this step, the complete response text of the current round is segmented into several sentence fragments, and a graph search is performed along the action knowledge graph starting from multiple initial action candidates to obtain multiple sets of candidate action sequences. For each candidate action at a position, the system considers the following factors simultaneously: first, the degree of local semantic matching between the current sentence fragment and the action description; second, the action-action transition edge weight between the action and the previous action; third, the degree of matching between the action and the current persona description; and fourth, whether there are repetitions in adjacent actions within the current round and in the already generated action paths, thus forming an intra-sequence repetition penalty. Preferably, a search strategy of retaining multiple high-scoring paths at each position can be adopted, retaining several candidate paths with high cumulative scores at each position, ultimately obtaining multiple sets of complete candidate action sequences.
[0038] In step S5, the graph search adopts a position-by-position path expansion strategy. For each position, the initial candidate action is comprehensively calculated for the semantic mismatch cost between the sentence fragment and the candidate action description, the inconsistency cost between the candidate action and the current character description, the transfer cost between the previous position action and the current candidate action, and the intra-sequence repetition penalty cost of the current candidate action relative to the previous position action. The search objective function or energy function of the candidate action sequence is constructed, and the action sequence that minimizes the energy function is retained to form a candidate action sequence set.
[0039] Specifically, in one embodiment, suppose the complete response text in the current round is segmented into T sentence fragments, and the corresponding candidate action sequence is denoted as... ,in This represents the candidate action at position T. Let T represent the T-th sentence segment. Then, the sequence energy function of the candidate action sequence A can be expressed as:
[0040] ,
[0041] in, This represents the cost of semantic mismatch between the current sentence fragment and the candidate action description. This indicates the cost of inconsistency between the candidate action and the current character description. This represents the transition cost between the previous action and the current candidate action. This represents the cost of intra-sequence repetition penalty for the current candidate action relative to the action at the previous position. , , , These are the corresponding weight coefficients. The action at the first position in the sequence... The determination of the action has been completed by the initial action search in step S4, so the action transition cost and the intra-sequence repetition penalty cost are calculated starting from t=2.
[0042] In one embodiment, the semantic mismatch cost This can be obtained in the following way: First, use a pre-trained language model to extract the current sentence segments. The semantic feature vectors of the candidate action description text are compared, and then the Euclidean distance between them is calculated. This distance is used as the semantic mismatch cost, or the distance is transformed into a cost through a preset mapping function. The better the sentence fragment matches the candidate action description, the smaller the corresponding semantic mismatch cost.
[0043] In one embodiment, the cost of the inconsistency in character design The following method can be used to obtain the semantic feature vectors of the current character description r and the character adaptation description text corresponding to the candidate action, respectively, using a pre-trained language model. Then, the Euclidean distance between the two is calculated and the distance is used as the character inconsistency cost, or the distance is transformed into a cost value through a preset mapping function. The more consistent the candidate action is with the current character description, the smaller the corresponding character inconsistency cost.
[0044] In one embodiment, the transfer cost It can be based on the action of the previous position in the action knowledge map. With current candidate actions The action-action transition edges between the two actions are determined. When there is a transition edge between them and the edge weight is high, the corresponding transition cost is low; when there is no transition edge between them or the edge weight is low, the corresponding transition cost is high. Preferably, the transition cost can be obtained based on the edge weight through a monotonically decreasing mapping function, so that the larger the edge weight, the smaller the transition cost.
[0045] In one embodiment, the intra-sequence repetition penalty cost Actions can be performed based on the previous position. With current candidate actions Whether they are the same or belong to a preset set of similar actions is determined; when and When two actions are identical, or highly similar in action category, action semantic label, or action performance, the repeat penalty cost within the corresponding sequence is set to a larger value; when the two actions are significantly different, the repeat penalty cost within the corresponding sequence is set to a smaller value or zero. The threshold for high similarity can be determined by those skilled in the art based on actual needs, for example, it can be set to 95% similarity. A larger value refers to a value that is larger than the value in the case of "significant differences between the two actions," and its specific value can be flexibly set by those skilled in the art, as can the value of "smaller value."
[0046] In one embodiment, the weighting coefficient , , , These are used to adjust the influence of local semantic matching, action transfer relationship, persona consistency, and intra-sequence repetition penalty on the sequence energy function, respectively. The weight coefficients can be set empirically, manually tuned, or learned based on training data.
[0047] The goal of graph search is to find the action sequence that minimizes the sequence energy function from the set of candidate action sequences, i.e.:
[0048] ,
[0049] in, This indicates the final selected candidate action sequence.
[0050] In one embodiment, after graph search, the present invention performs a global decision on multiple sets of candidate action sequences based on a language model. During the global decision, the language model does not freely generate arbitrary action sequences. Instead, under the constraints of the current round's complete response text and the given set of candidate action sequences, the language model performs a selection-based judgment task on the set of candidate action sequences, selecting the set that best matches the semantics of the current text and providing the starting position of the corresponding semantic segment for alignment with the starting positions of subsequent action texts and time sequences. By transforming the divergent generation process of the language model into a constrained selection or classification process, the stability and accuracy of the global decision can be improved.
[0051] Step S6: Action Alignment, Feedback Quantization, and Graph Update. In this step, the system determines the text start position of each action in the current round of text based on the final action sequence and the semantic fragment start position of the global decision output. It then combines this with the speech alignment timestamp to obtain the corresponding time start position, generating a timing instruction that can directly drive the digital human's actions. After the action is executed, the system does not directly write the original feedback into the graph; instead, it first forms an empirical feedback unit. This empirical feedback unit can record at least: the current round of text and its segmentation results, the unified behavioral semantic purpose, the candidate action sequence set, the final executed action sequence, the semantic fragment start position, the text position and time position alignment results, explicit feedback, implicit feedback, and the execution result.
[0052] Explicit and implicit feedback are collected to quantify the feedback results and form a comprehensive feedback score. Based on the comprehensive feedback score, the current semantic-action adaptation edge and action-action transfer edge are multiplicatively adjusted. At the same time, new expressions are deduplicated and filtered for differences. The current round text or its representative expression that meets the conditions is written into the instance expression library of the corresponding behavior semantic category, thereby realizing the continuous updating of the action knowledge graph and behavior evolution.
[0053] Specifically, in one embodiment, the system first quantizes explicit and implicit feedback into numerical feedback components. The explicit feedback includes likes, dislikes, and human ratings, where likes correspond to positive feedback values, dislikes to negative feedback values, and human ratings to the rating itself or its normalized result. The implicit feedback includes at least whether the user interrupted the action, where no interruption corresponds to a zero value, and an interruption corresponds to a negative penalty value. The explicit and implicit feedback are mapped to feedback components via quantization functions and then weighted and fused to obtain a comprehensive feedback score F, which is used to perform subsequent graph update operations. Preferably, it can be expressed as:
[0054] ,
[0055] in, This indicates an explicit feedback component. This represents implicit feedback components. and This represents the corresponding fusion weight coefficient.
[0056] After obtaining the comprehensive feedback score F, the semantic-action adaptation edges and action-action transition edges are updated. Let the weights of the semantic-action adaptation edges before and after the update be respectively... and The weights of the action-action transition edge before and after the update are respectively and The update process can then be represented as:
[0057] ,
[0058] ,
[0059] in, and To update the step size coefficient, This represents a truncation function that constrains the updated edge weights within a preset legal range. Employing an exponential update method ensures that the strengthening and weakening effects of positive and negative feedback on edge weights remain continuous and smooth, while avoiding the edge weight anomalies that can occur with linear update methods when feedback amplitudes are large.
[0060] Furthermore, if the overall feedback score F is higher than the preset threshold... Furthermore, the current round text or its representative expression corresponds to the set of existing instance expressions under the corresponding behavioral semantic category. The difference between them is greater than a preset threshold. Then, consider writing the current round text or its representative expression into the instance expression library. The writing conditions can be expressed as follows:
[0061] If and only if and ,
[0062] Where x represents the current round of text or its representative expression, Represents x and the set of instances. The difference between them is measured. In one embodiment, a pre-trained language model can be used to extract x and the set. The semantic feature vectors of each instance in the set are calculated, and then the Euclidean distance between x and the most similar instance in the set is calculated. This minimum distance is used as a measure of dissimilarity. The larger the dissimilarity, the lower the degree of repetition between the current expression and existing instance expressions.
[0063] This condition indicates that a write operation will be performed; otherwise, the write operation will not be performed. The system can use this condition to deduplicate and filter newly added expressions, thereby enabling the continuous evolution of the graph structure and instance expression library.
[0064] This invention also proposes an action sequence generation and behavior evolution system based on action knowledge graph for embodied dialogue agents, comprising:
[0065] The action knowledge graph construction module is used to collect multi-turn user-agent dialogue samples and action resources, label each round of response with a unified behavioral semantic purpose, and segment the response text into several sentence fragments; on this basis, behavioral semantic category nodes, action nodes, semantic-action adaptation edges and action-action transition edges are established, and an instance expression library and an action description library are constructed, and the instance expression text is pre-encoded offline.
[0066] The action sequence generation module, after obtaining the complete response text of the current round, first determines the unified semantic purpose of the behavior by matching the similarity between the instance representation library and the complete response text of the current round. Then, based on the unified semantic purpose of the behavior, it recalls candidate actions from the corresponding semantic category node along the semantic-action adaptation edge. Combining the matching degree between the first sentence fragment of the current round response and the action description, the current persona description, the action-action transfer edge connection relationship and edge weight between the last action of the previous round and the candidate initial action, and the cross-round repetition penalty, the candidate actions are filtered and sorted to obtain multiple initial action candidates. Starting from the multiple initial action candidates, the module expands position by position to obtain multiple sets of candidate action sequences by combining the local semantic matching between the sentence fragment and the action description, the action-action transfer edge weight, the persona matching, and the intra-sequence repetition penalty. Finally, the language model makes a global decision under the constraints of the current round text and the candidate action sequence group to obtain the final action sequence and the starting position of the semantic fragment corresponding to each action.
[0067] The alignment and driving module is used to determine the text start position of each action in the current round of text based on the final action sequence and the starting position of the semantic fragment corresponding to each action, and to obtain the corresponding time start position by combining the voice alignment timestamp, and to generate action execution instructions to drive the digital human to complete body movements, gestures or other non-verbal behavior outputs.
[0068] The system may also include a feedback quantification and graph update module, which collects explicit and implicit feedback after an action is executed, quantifies the feedback results, and forms a comprehensive feedback score. Based on the comprehensive feedback score, it performs multiplicative adjustments on the current semantic-action adaptation edges and action-action transition edges, and performs deduplication and difference filtering on newly added expressions. The newly added expressions that meet the conditions are written into the instance expression library of the corresponding behavior semantic category, thereby realizing the continuous updating of the action knowledge graph and behavior evolution.
[0069] The present invention also provides a computing device, comprising: at least one processor and a memory storing program instructions; when the program instructions are read and executed by the processor, the computing device causes the computing device to execute the action sequence generation and behavior evolution method based on action knowledge graph for embodied dialogue agents.
[0070] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments can be conceived within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of explaining or limiting the subject matter of the invention.
Claims
1. A method for action sequence generation and behavior evolution based on action knowledge graphs for embodied dialogue agents, characterized in that, Including the following steps: Step S1: Construct an action knowledge graph based on multi-turn user-agent dialogue data, action resources, and annotation results. The action knowledge graph includes behavior semantic category nodes, action nodes, semantic-action adaptation edges, and action-action transition edges. An instance representation library is established for each behavior semantic category node, and an action description library is established for each action node. Step S2: Extract features from the text in the instance representation library using a language model, map them to a unified semantic vector space, and store the vector results; Step S3: Obtain the complete response text for the current round, and perform semantic similarity matching by using the semantic similarity between the complete response text for the current round and the instance expression text under each behavioral semantic category to obtain the unified behavioral semantic purpose for the current round; Step S4: Based on the unified behavioral semantic purpose, recall candidate initial actions along the semantic-action adaptation edge, and combine the semantic matching degree between the first sentence fragment at the beginning of the current round reply and the action description, the matching degree between the current character description and the character applicable to the action, the action-action transfer edge connection relationship and edge weight between the last action of the previous round and the candidate initial action, and the cross-round repetition penalty of the candidate initial action relative to the last action of the previous round, sort the candidate initial actions to obtain multiple initial action candidates; Step S5: Divide the complete reply text of the current round into several sentence fragments. Starting from the initial action candidate, perform graph search based on the local semantic matching degree between the sentence fragment and the action description, the action-action transition edge weight, the matching degree between the action and the current character description, and the intra-sequence repetition penalty to generate multiple sets of candidate action sequences. Then, perform global adjudication on the candidate action sequences to obtain the final action sequence and the starting position of the semantic fragment corresponding to each action. Step S6: Based on the final action sequence and the starting position of the semantic fragment, output the text starting position and time starting position of each action in the current round of complete response text, and generate timing instructions that can directly drive the digital human to perform actions.
2. The method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents according to claim 1, characterized in that, The action description library mentioned in step S1 includes at least the applicable semantic category, applicable scenario, applicable persona description, keywords and typical expressions, and style tags of the action, which are used to support the semantic filtering and local matching of candidate actions.
3. The method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents according to claim 1, characterized in that, Step S3, the whole round semantic determination, includes: mapping the current round complete response text and the instance expression text under each behavior semantic category to the same semantic vector space, calculating the similarity between texts, and determining the unified behavior semantic purpose of the current round based on the aggregated similarity of several instance expressions in each behavior semantic category that are most similar to the current round complete response text.
4. The method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents according to claim 1, characterized in that, In step S5, the graph search adopts a position-by-position path expansion strategy. For each position, the initial candidate action is comprehensively calculated for the semantic mismatch cost between the sentence fragment and the candidate action description, the inconsistency cost between the candidate action and the current character description, the transfer cost between the previous position action and the current candidate action, and the intra-sequence repetition penalty cost of the current candidate action relative to the previous position action. The energy function of the candidate action sequence is constructed, and the action sequence that minimizes the energy function is retained as the candidate action sequence set.
5. The method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents according to claim 1, characterized in that, S6 also includes: after the action is executed, collecting explicit and implicit feedback to quantify the feedback results and form a comprehensive feedback score; and multiplicatively adjusting the current semantic-action adaptation edge and action-action transfer edge according to the comprehensive feedback score, while deduplicating and filtering newly added expressions, and writing the current round complete response text or its representative expression that meets the conditions into the instance expression library of the corresponding behavior semantic category, thereby realizing the continuous updating of the action knowledge graph and behavior evolution.
6. The method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents according to claim 5, characterized in that, The explicit feedback includes likes, dislikes, and human ratings, while the implicit feedback includes whether the user interrupted. The explicit and implicit feedback are mapped to feedback components by quantization functions and then weighted and fused to obtain a comprehensive feedback score.
7. The method for action sequence generation and behavior evolution based on action knowledge graph for embodied dialogue agents according to claim 6, characterized in that, After obtaining the comprehensive feedback score F, the semantic-action adaptation edges and action-action transition edges are updated: Let the weights of the semantic-action adaptation edges before and after the update be respectively... and The weights of the action-transition edge before and after the update are respectively and The update process is then represented as: ， ， in, and To update the step size coefficient, This represents a truncation function that constrains the updated edge weights within a preset legal range.
8. A system for action sequence generation and behavior evolution based on action knowledge graphs for embodied dialogue agents, characterized in that, include: The action knowledge graph construction module is used to collect multi-turn user-agent dialogue samples and action resources, label each round of response with a unified behavioral semantic purpose, and segment the response text into several sentence fragments; on this basis, an action knowledge graph including behavioral semantic category nodes, action nodes, semantic-action adaptation edges and action-action transition edges is constructed, and an instance expression library and an action description library are established, and the instance expression text is pre-encoded offline; The action sequence generation module is used to first determine the unified action semantic purpose by matching the similarity between the instance expression library and the current round complete response text after obtaining the complete response text of the current round, and then recall candidate actions from the corresponding semantic category node along the semantic-action adaptation edge according to the unified action semantic purpose. By combining the matching degree between the first sentence fragment of the current round's response and the action description, the current persona description, the action-action transfer edge connection relationship and edge weight between the last action of the previous round's executed sequence and the candidate initial action, as well as the cross-round repetition penalty, the candidate actions are screened and sorted to obtain multiple initial action candidates. Starting from these multiple initial action candidates, by combining the local semantic matching between the sentence fragment and the action description, the action-action transfer edge weight, the persona matching, and the intra-sequence repetition penalty, multiple sets of candidate action sequences are obtained position by position. Then, the language model makes a global decision under the constraints of the current round's text and the candidate action sequence group to obtain the final action sequence and the starting position of the semantic fragment corresponding to each action. The alignment and driving module is used to determine the text start position of each action in the current round of complete response text based on the final action sequence and the start position of the semantic fragment corresponding to each action, and to obtain the corresponding time start position by combining the voice alignment timestamp, generate action execution instructions, and drive the digital human to execute.
9. The action sequence generation and behavior evolution system based on action knowledge graph for embodied dialogue agents according to claim 8, characterized in that, Also includes: The Feedback Quantification and Graph Update module is used to collect explicit and implicit feedback after an action is executed, quantify the feedback results, and form a comprehensive feedback score. Based on the comprehensive feedback score, the current semantic-action adaptation edge and action-action transfer edge are multiplicatively adjusted. At the same time, new expressions are deduplicated and filtered for differences. New expressions that meet the conditions are written into the instance expression library of the corresponding behavior semantic category, thereby realizing the continuous updating of the action knowledge graph and behavior evolution.
10. A computing device, characterized in that, include: At least one processor and a memory storing program instructions; When the program instructions are read and executed by the processor, the computing device performs the action sequence generation and behavior evolution method based on action knowledge graph for embodied dialogue agents as described in any one of claims 1-7.