A Method for Image-Based Retrieval in AI Bionic Robots Based on Chat Keyword Decomposition

By constructing a graph memory database based on chat keyword decomposition, the problems of low storage structure and low retrieval efficiency of AI bionic robot puppets are solved, realizing an efficient and logically related memory system with basic reasoning ability and automatic update function.

CN123115305APending Publication Date: 2026-11-17SHENZHEN BANYINGSUIXING TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611061129.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing AI bionic robot puppet memory systems suffer from problems such as low degree of storage structuring, low retrieval efficiency, weak reasoning ability, and difficulty in memory updating.

Method used

A keyword-based approach is adopted to extract core entities, relational words, and attribute words from chat content through word segmentation, part-of-speech tagging, and named entity recognition. A graph memory database is constructed, and a graph traversal algorithm is used for retrieval and reasoning. A forgetting mechanism is introduced to handle information conflicts.

Benefits of technology

It achieves structured storage and efficient retrieval, improves the logical connection of memories and the coherence of dialogue, possesses basic reasoning ability, and automatically updates and maintains the accuracy of memories.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

This invention provides an AI-based bionic robot graph memory retrieval method based on chat keyword decomposition, relating to the fields of artificial intelligence natural language processing and knowledge graph technology. It includes the following steps: chat content preprocessing and keyword decomposition, performing word segmentation, part-of-speech tagging, and named entity recognition on real-time acquired chat text to extract core entities, relational words, and attribute words from the chat content, and filtering stop words; graph memory database construction, using the extracted core entities as nodes of an attribute graph, relational words as edges of the attribute graph, and attribute words as attribute values ​​of corresponding nodes to construct a graph memory database. This invention uses a graph structure to store memory information, effectively extracting and preserving the relationships between entities, making memory no longer fragmented text but a logically connected knowledge network; the time complexity of graph traversal retrieval is far lower than that of text keyword matching retrieval, achieving millisecond-level response even with large-scale memory data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence natural language processing and knowledge graph technology, specifically to an AI bionic robot diptych memory retrieval method based on chat keyword decomposition. Background Technology

[0002] Memory capacity is a crucial indicator of the intelligence level of AI-powered bionic robot puppets. Currently, most bionic robot puppet memory systems employ a text-based linear storage method, storing chat content as plain text in a database and retrieving relevant content through keyword matching. This storage and retrieval method has the following significant drawbacks: 1. Low degree of storage structuring: Plain text storage cannot effectively extract and save entity, relationship and attribute information in chat content, resulting in fragmented memory and lack of contextual association; 2. Low retrieval efficiency: Keyword matching retrieval requires traversing a large amount of text data, and the retrieval speed will decrease significantly as the amount of data to be remembered increases; 3. Weak reasoning ability: Unable to make simple logical inferences and associations based on existing memory information, resulting in poor dialogue coherence and frequent "answers that are not what are asked"; 4. Difficulty in updating memories: When new information conflicts with old memories, it cannot be automatically recognized and updated, leading to memory errors. Summary of the Invention

[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an AI bionic robot puppet memory retrieval method based on chat keyword decomposition, which solves the problems of low storage structure, low retrieval efficiency, and weak reasoning ability in existing AI bionic robot puppet memory systems.

[0004] (II) Technical Solution To achieve the above objectives, this invention employs the following technical solution: an AI bionic robot diptych memory retrieval method based on chat keyword decomposition, step one: chat content preprocessing and keyword decomposition: The system performs word segmentation, part-of-speech tagging, and named entity recognition on the chat text acquired in real time, extracts the core entities, relation words, and attribute words in the chat content, and filters out stop words. Step 2: Building the graph memory database: The extracted core entities are used as nodes in the attribute graph, relation words are used as edges in the attribute graph, and attribute words are used as attribute values ​​of the corresponding nodes to construct a graph memory database. When new chat content is generated, the graph memory database is automatically updated. If the entity already exists, its attributes and frequency of occurrence are updated. If the relation already exists, its confidence and update time are updated. If the entity or relation does not exist, the corresponding node or edge is added. Step 3: Memory Retrieval and Reasoning The system breaks down the current user's question into keywords, extracts the search keywords, searches for nodes that match the search keywords in the graph memory database, retrieves all nodes and edges that are directly or indirectly connected to the node using a graph traversal algorithm, sorts the search results based on the confidence of the edges and the frequency of the nodes, and generates an answer containing contextual information based on the retrieved nodes and edges.

[0005] Preferably, the nodes in step two include entity name, entity type, frequency of occurrence, first occurrence time, most recent occurrence time, and confidence level attributes; each edge includes relation type, confidence level, establishment time, update time, and number of visits attributes.

[0006] Preferably, it also includes step four, memory conflict handling and forgetting mechanism: when new information conflicts with existing memory information, the establishment time and confidence level of the conflicting information are compared, and the information with the newer time and higher confidence level is retained; for nodes and edges that have not been accessed for a long time and have a low frequency of occurrence, their confidence level is automatically reduced, and when the confidence level is lower than a preset threshold, they are deleted from the graph memory database.

[0007] Preferably, the core entities include, but are not limited to, person name entities, place name entities, organization entities, time entities, numerical entities, and thing entities; the relation words include, but are not limited to, sentiment relation words, social relation words, action relation words, attribute relation words, and causal / conditional relation words; the attribute words include, but are not limited to, color attribute words, size attribute words, sentiment attribute words, time attribute words, and quantity attribute words.

[0008] Preferably, in step two, a graph database is used as the physical storage engine, and a full-text index of node names, a B-tree index of node types, a B-tree index of edge types, and a time index are established.

[0009] Preferably, the node matching strategy in step three includes exact matching, fuzzy matching, and semantic matching; the graph traversal algorithm adopts an improved breadth-first search algorithm and supports multi-hop reasoning.

[0010] Preferably, the ranking of search results in step three comprehensively considers edge confidence, node frequency, access time decay, and semantic similarity. The comprehensive ranking score is calculated using the following formula: Score=α*confidence+β*frequency_norm+γ*time_decay+δ*semantic_similarity, where α, β, γ, and δ are weight coefficients.

[0011] Preferably, the conflict resolution strategies in step four include a time-first strategy, a confidence-first strategy, a frequency-first strategy, and a comprehensive strategy; the comprehensive score calculation formula for the comprehensive strategy is as follows: Conflict_Score = w1 * time_score + w2 * confidence + w3 * frequency_norm, where time_score = 1 / (1 + number of days of time difference), and w1, w2, and w3 are weighting coefficients.

[0012] Preferably, the confidence decay formula for the forgetting mechanism described in step four is: confidence_new=confidence_old*e^(-λ*Δt), where λ is the decay coefficient and Δt is the time interval since the last visit.

[0013] Preferably, step four also includes a memory consolidation mechanism: for frequently accessed nodes and edges, their confidence is increased, and the consolidation formula is: confidence_new=min(confidence_old+0.01*access_count,1.0).

[0014] (III) Beneficial Effects This invention provides an AI bionic robot diptych memory retrieval method based on chat keyword decomposition. It has the following beneficial effects: 1. Structured storage with strong contextual association: The graph structure is used to store memory information, which effectively extracts and preserves the relationships between entities, so that memory is no longer fragmented text, but a knowledge network with logical connections.

[0015] 2. High retrieval efficiency and fast response speed: The time complexity of graph traversal retrieval is much lower than that of text keyword matching retrieval. Even when faced with large-scale memory data, it can achieve millisecond-level response.

[0016] 3. Possesses basic reasoning ability: The graph-based memory database supports simple logical reasoning and associative thinking, and can answer questions that require combining multiple pieces of memory information, significantly improving the coherence and intelligence of the dialogue.

[0017] 4. Automatic memory update and forgetting: It can automatically handle information conflicts to ensure the accuracy of memory; it introduces a forgetting mechanism to avoid the infinite expansion of the memory database and improve the system's operating efficiency. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a method for AI bionic robot diptych memory retrieval based on chat keyword decomposition, comprising the following steps: Step 1: Chat content preprocessing and keyword decomposition: The real-time acquired chat text is segmented using a deep learning-based word segmentation model (such as a BERT-based word segmenter). The cleaned text is then segmented, part-of-speech tagging is performed, and named entity recognition is conducted to extract core entities, relational words, and attribute words from the chat content. Stop words are filtered out, specifically meaningless function words, modal particles, and auxiliary words, while retaining valid information. The stop word list can be dynamically expanded according to actual application scenarios. Core entities include, but are not limited to, personal name entities, place name entities, organizational entity entities, time entities, numerical entities, and thing entities. See below for details: Personal Entities (PER): such as Xiaoming, Zhang San, Teacher Li, etc.; Place names (LOCs): such as Beijing, Shanghai, West Lake, etc.; Organizational entities (ORGs): such as Tsinghua University, Alibaba, etc. Time entities: such as 2024, yesterday, 3 PM, etc.; Numerical entities (NUM): such as 25 years old, 100 yuan, 3 kilometers, etc.; Entities (THING): such as apples, basketballs, movies, etc. Relationship terms include, but are not limited to, sentiment relation terms, social relation terms, action relation terms, attribute relation terms, and causal / conditional relation terms. See below for details: Emotional relationships: such as liking, disliking, loving, hating, etc.; Social relationships: such as friends, colleagues, family, classmates, etc.; Action relationships: such as having been there, owning, purchasing, making, etc.; Attribute relationships: such as, belonging to, having, etc.; Causal / conditional relationships: such as because, therefore, if, etc.

[0020] Attribute words include, but are not limited to, color attribute words, size attribute words, sentiment attribute words, time attribute words, and quantity attribute words; see the following for details: Color attributes: such as red, blue, green, etc.; Size attributes: such as tall, small, wide, etc.; Emotional attributes: such as interesting, boring, happy, sad, etc.; Time attributes: such as new, old, ancient, etc.; Quantity attributes: such as many, few, all, etc.

[0021] Step 2: Building the graph memory database: The extracted core entities are used as nodes in the attribute graph, relation words are used as edges in the attribute graph, and attribute words are used as attribute values ​​for corresponding nodes. A graph memory database is constructed, using the graph database as the physical storage engine, and full-text indexes for node names, B-tree indexes for node types, B-tree indexes for edge types, and time indexes are established. When new chat content is generated, the graph memory database is automatically updated. If the entity already exists, its attributes and frequency of occurrence are updated; if the relation already exists, its confidence and update time are updated; if the entity or relation does not exist, the corresponding node or edge is added. That is, if the core entity extracted in step one is used as a node in the graph, a new node is created when the entity appears for the first time, with an initial frequency of occurrence of 1 and a confidence of 0.8, and the first occurrence time and the most recent occurrence time are recorded; if the entity already exists, its frequency of occurrence (frequency += 1) and the most recent occurrence time (last_seen = current time) are updated, and its attribute list is updated according to the new context information. The nodes mentioned above include attributes such as entity name, entity type, frequency of occurrence, first occurrence time, most recent occurrence time, and confidence level; each edge includes attributes such as relation type, confidence level, establishment time, update time, and number of visits. The relation terms extracted in step one are used as edges in the graph. If a relation appears for the first time, a new edge is created, the initial confidence is set to 0.8, the creation time and update time are recorded, and the number of visits is 0. If a relation already exists, its confidence is updated (confidence=min(confidence+0.05,1.0)), the update time is updated, and the number of visits is increased. The attribute words extracted in Step 1 are used as attribute values ​​for the corresponding nodes and stored in the node's attribute list. Attribute value updates follow these rules: if the attribute name does not exist, the attribute is added; if the attribute name already exists and the attribute values ​​are different, the confidence levels of the old and new information are compared, and the value with the higher confidence level is retained; if the attribute name already exists and the attribute values ​​are the same, the frequency of occurrence of that attribute is increased.

[0022] Step 3: Memory Retrieval and Reasoning The current user's question is broken down into keywords, and search keywords are extracted. Nodes matching the search keywords are searched in a graph memory database. A graph traversal algorithm is used to retrieve all nodes and edges directly or indirectly connected to the selected node. The search results are sorted based on edge confidence and node frequency. An answer containing contextual information is generated based on the retrieved nodes and edges. The node matching strategies include exact matching (node ​​name exactly matches the search keyword), fuzzy matching (edit distance between node name and search keyword is less than a threshold, such as 2), and semantic matching (semantic similarity between node name and search keyword is calculated using a pre-trained language model; similarity scores above a threshold (such as 0.85) are considered matches). The graph traversal algorithm uses an improved breadth-first search algorithm, supporting multi-hop reasoning. The specific steps of the breadth-first search algorithm are as follows: Initialization: Add matching nodes to the queue and set the initial depth to 0; Traversal: Remove a node from the queue, obtain all its outgoing and incoming edges, add adjacent nodes to the queue, and increment the depth by 1; Termination condition: Terminate when the queue is empty or the maximum traversal depth (e.g., 3 levels) is reached; Result collection: Collect all nodes and edges visited during the traversal to form a subgraph.

[0023] In the above, based on the retrieved nodes and edges, the following reasoning operation is performed before answering: Direct association reasoning: If the query node is directly connected to the target node, the association information is returned directly; Indirect association reasoning: If the query node is indirectly connected to the target node through an intermediate node, the reasoning path is constructed and returned; Multi-hop reasoning: Supports multi-hop reasoning (e.g., what do my friend's friend like), implemented through multiple graph traversals; Attribute reasoning: Reasoning is performed based on the attribute information of the node (e.g., what are some red fruits).

[0024] Step 4: Memory Conflict Handling and Forgetting Mechanisms When new information conflicts with existing memory information, the creation time and confidence level of the conflicting information are compared, and the information with the newer time and higher confidence level is retained; for nodes and edges that have not been accessed for a long time and appear infrequently, their confidence level is automatically reduced, and when the confidence level is lower than a preset threshold, they are deleted from the graph memory database. Memory consolidation mechanism: Increase the confidence of frequently accessed nodes and edges: confidence_new=min(confidence_old+0.01*access_count,1.0).

[0025] In step three, the ranking of search results comprehensively considers edge confidence, node frequency, access time decay, and semantic similarity. The formula for calculating the comprehensive ranking score is as follows: Score = α * confidence + β * frequency_norm + γ * time_decay + δ * semantic_similarity, where α, β, γ, and δ are weight coefficients. The higher the confidence, the higher the ranking; the higher the frequency, the higher the ranking; the more recently visited the node, the higher the ranking; and the more relevant the node is to the query semantics, the higher the ranking.

[0026] Step four's conflict resolution strategies include a time-priority strategy (comparing the creation time of conflicting information and retaining the more recent information), a confidence-priority strategy (comparing the confidence levels of conflicting information and retaining the information with higher confidence levels), a frequency-priority strategy (comparing the frequency of occurrence of conflicting information and retaining the information with higher frequency), and a comprehensive strategy (considering time, confidence, and frequency comprehensively, calculating a comprehensive score, and retaining the information with the higher score); the formula for calculating the comprehensive score of the comprehensive strategy is: Conflict_Score=w1*time_score+w2*confidence+w3*frequency_norm Where time_score=1 / (1+time difference in days), it means that the newer the information, the higher the score, and w1, w2, w3 are weight coefficients.

[0027] The confidence decay formula for the forgetting mechanism in step four is: confidence_new=confidence_old*e^(-λ*Δt) Where λ is the decay coefficient, Δt is the time interval since the last access, automatic deletion: when the confidence of a node or edge is lower than a preset threshold (e.g., 0.3), it is deleted from the graph memory database; archive storage: deleted nodes and edges can be archived to the historical database for later recovery or auditing.

[0028] Implementation Plan 1: Basic Fact Memory. When a user says "I like to eat apples," the system extracts the entities "I," "apple," and the relation "like to eat," constructs nodes "I" and "apple," and establishes an edge of type "like to eat" between them. When the user subsequently asks "What fruit do I like to eat?", the system quickly finds "apple" by searching the outgoing edges of node "I" and generates the answer.

[0029] Implementation Plan 2: Associative Reasoning Memory. When a user says "Xiaoming is my friend" and "Xiaoming likes to play basketball," the system constructs the nodes "I," "Xiaoming," and "basketball," and establishes two edges: "I-friend-Xiaoming" and "Xiaoming-likes-basketball." When the user asks "What sport does my friend like?", the system infers "Xiaoming likes to play basketball" through graph traversal and generates the answer.

[0030] Implementation Plan 3: Memory Update and Conflict Handling. When a user first says "I am 25 years old" and then says "I am 26 years old", the system compares the creation time of the two pieces of information, retains the newer "26 years old" information, and updates the age attribute of the node "I".

[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI bionic robot image memory retrieval method based on chat keyword decomposition, characterized in that, Includes the following steps: Step 1: Chat content preprocessing and keyword decomposition: The system performs word segmentation, part-of-speech tagging, and named entity recognition on the chat text acquired in real time, extracts the core entities, relation words, and attribute words in the chat content, and filters out stop words. Step 2: Building the graph memory database: The extracted core entities are used as nodes in the attribute graph, relation words are used as edges in the attribute graph, and attribute words are used as attribute values ​​of the corresponding nodes to construct a graph memory database. When new chat content is generated, the graph memory database is automatically updated. If the entity already exists, its attributes and frequency of occurrence are updated. If the relation already exists, its confidence and update time are updated. If the entity or relation does not exist, the corresponding node or edge is added. Step 3: Memory Retrieval and Reasoning The system breaks down the current user's question into keywords, extracts the search keywords, searches for nodes that match the search keywords in the graph memory database, retrieves all nodes and edges that are directly or indirectly connected to the node using a graph traversal algorithm, sorts the search results based on the confidence of the edges and the frequency of the nodes, and generates an answer containing contextual information based on the retrieved nodes and edges.

2. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 1, characterized in that, The nodes in step two include entity name, entity type, frequency of occurrence, first occurrence time, most recent occurrence time, and confidence level attributes; each edge includes relation type, confidence level, establishment time, update time, and number of visits attributes.

3. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 2, characterized in that, It also includes step four, memory conflict handling and forgetting mechanism: when new information conflicts with existing memory information, the establishment time and confidence level of the conflicting information are compared, and the information with the newer time and higher confidence level is retained; for nodes and edges that have not been accessed for a long time and have a low frequency of occurrence, their confidence level is automatically reduced, and when the confidence level is lower than a preset threshold, they are deleted from the graph memory database.

4. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 3, characterized in that: The core entities include, but are not limited to, personal name entities, place name entities, organizational entity entities, time entities, numerical entities, and object entities; the relational terms include, but are not limited to, sentiment relational terms, social relational terms, action relational terms, attribute relational terms, and causal / conditional relational terms. The attribute words include, but are not limited to, color attribute words, size attribute words, sentiment attribute words, time attribute words, and quantity attribute words.

5. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 4, characterized in that: In step two, a graph database is used as the physical storage engine, and a full-text index of node names, a B-tree index of node types, a B-tree index of edge types, and a time index are established.

6. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 5, characterized in that: The node matching strategies in step three include exact matching, fuzzy matching, and semantic matching; the graph traversal algorithm adopts an improved breadth-first search algorithm and supports multi-hop reasoning.

7. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 6, characterized in that: In step three, the ranking of search results comprehensively considers edge confidence, node frequency, access time decay, and semantic similarity. The formula for calculating the comprehensive ranking score is as follows: Score=α*confidence+β*frequency_norm+γ*time_decay+δ*semantic_similarity Where α, β, γ, and δ are weighting coefficients.

8. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 7, characterized in that: The conflict resolution strategies described in step four include time-priority strategy, confidence-priority strategy, frequency-priority strategy, and a comprehensive strategy; the comprehensive score calculation formula for the comprehensive strategy is as follows: Conflict_Score=w1*time_score+w2*confidence+w3*frequency_norm Where time_score = 1 / (1 + number of days of time difference), w1, w2, and w3 are weighting coefficients.

9. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 8, characterized in that: The confidence decay formula for the forgetting mechanism described in step four is: confidence_new=confidence_old*e^(-λ*Δt) Where λ is the decay coefficient and Δt is the time interval since the last access.

10. The AI ​​bionic robot diptych memory retrieval method based on chat keyword decomposition according to claim 9, characterized in that, Step 4 also includes a memory consolidation mechanism: for frequently accessed nodes and edges, their confidence is increased, and the consolidation formula is: confidence_new=min(confidence_old+0.01*access_count,1.0).