AI bionic robot double chat content pushing method based on user interest extraction
Patent Information
- Application Number
- CN202611083382.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-10-30
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for pushing chat content to AI-inspired bionic robots based on user interest extraction. Background Technology
[0002] Currently, chat systems for AI-inspired bionic robots primarily employ two technical approaches. The first is the pre-defined dialogue library approach, where developers pre-write numerous dialogue templates and trigger keywords. When the user's input matches a specific keyword, the system retrieves the corresponding pre-defined response from the dialogue library and outputs it. The second approach is the general-purpose large language model approach, which involves calling the application programming interface (API) of a third-party large-scale language model. The large model generates response text in real-time based on user input and then returns it to the robot for output.
[0003] However, both of these approaches have significant shortcomings. The pre-set dialogue library approach has completely fixed content, failing to cover all topics users might mention, and its timeliness depends entirely on the developer's update frequency, making it unable to access real-time external information. While the large language model approach possesses strong generation capabilities, its training data has a deadline, making it unable to answer questions about events occurring after the training data deadline, resulting in content lag. A more critical common flaw is that both approaches operate on a passive response model—the system only generates a reply after receiving a user's question, failing to proactively identify user interests and preferences to initiate new topics or push potentially interesting content, leading to a lack of initiative and engagement in the interaction process.
[0004] Furthermore, even when some systems attempt to introduce content push functionality, their push strategies are often quite simplistic. For example, they might match and push content based solely on a single keyword appearing in a user's historical conversations, neglecting the continuous tracking and modeling of the user's deeper interests. This results in a low degree of relevance between the pushed content and the user's current true interests. Additionally, the insertion of pushed content is often abrupt, disrupting the smooth flow of the conversation and creating a sense of unease for the user, ultimately reducing the overall interactive experience. Summary of the Invention
[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for pushing chat content to AI bionic robot puppets based on user interest extraction, which solves the problems of passive content, poor timeliness, and low personalization in existing AI bionic robot puppet chat systems.
[0006] (II) Technical Solution To achieve the above objectives, the present invention is implemented through the following technical solution: an AI bionic robot puppet chat content push method based on user interest extraction, including extracting user interest features from real-time chat text; The user's long-term interest model is updated based on the aforementioned interest features, wherein the long-term interest model includes information on the user's degree of interest in each topic and time attribute information. The information to be pushed is obtained based on the long-term interest model. In response to the fulfillment of the push timing trigger condition, the information to be pushed is transformed into a conversational message using natural language generation technology, and the conversational message is inserted into the current chat dialogue stream. The structured information content is transformed into a conversational message that conforms to spoken language habits through natural language generation technology and output by inserting it into the dialogue stream, so that the push content is consistent with the normal chat dialogue in terms of language style and appearance, eliminating the user's sense of abruptness. Obtain the user's behavioral response to the session message, and update the long-term interest model based on the behavioral response.
[0007] Preferably, the extraction of user interest features from real-time chat text includes: extracting at least one of keywords, entities, and topics from the real-time chat text; and determining the degree of interest corresponding to each keyword, entity, and topic. The process involves first performing word segmentation and part-of-speech tagging on the real-time chat text, dividing the continuous text sequence into word units with independent semantics, and tagging the part of speech of each word. Based on the word segmentation results, a keyword extraction algorithm (e.g., the TF-IDF algorithm based on word frequency-inverse document frequency, or the TextRank algorithm based on graph ranking) is used to calculate the weight score of each word, and the top-ranked words (e.g., the top 5) are selected as the extracted keywords. Entity identification can be based on sequence labeling methods using conditional random fields or pre-trained language models (e.g., BERT series models) to label and extract entities such as names of people, places, organizations, and works from the text. Topic classification can be implemented based on a text classification model, which maps the input text to one or more topic tags in a preset topic classification system. The classification system can be designed as a tree-like hierarchical structure. For example, the first-level categories include "technology", "entertainment", "sports" and "lifestyle", and each first-level category has several second-level subcategories.
[0008] Determining the interest level corresponding to each keyword, entity, and topic includes: analyzing at least one of the following in the real-time chat text: tone features (such as the frequency of interjections), sentiment tendency (positive sentiment usually indicates higher interest), and discussion depth (the degree to which users actively engage in the topic); and determining the interest level based on the analysis results. Tone features, sentiment tendency, and discussion depth reflect the user's true level of attention to a topic from different dimensions. Incorporating this information into the interest level calculation significantly improves the accuracy of interest judgment. Sentiment tendency is scored using a sentiment analysis model, classifying the text as positive, neutral, or negative, or using continuous values. Positive sentiment usually indicates that the user holds a positive attitude towards the topic, corresponding to a higher level of interest. Discussion depth can be quantified as: the number of consecutive rounds of user comments on the same topic, the information density in the user's comments (measured by text length and the proportion of content words), and whether the user actively asks questions to explore the topic in depth. The final value of the interest level is the weighted sum or product of the scores of the above dimensions. The weights can be adjusted experimentally based on actual application scenarios. The range of interest level can be normalized to between 0 and 1, where 0 represents no interest and 1 represents extremely strong interest.
[0009] Preferably, updating the user's long-term interest model based on the interest features includes: weighting and sorting the extracted interest features according to interest level, time freshness, and frequency of occurrence to obtain a ranking result; updating the weight of the corresponding topic in the long-term interest model according to the ranking result; and effectively distinguishing the topics that the user is currently most concerned about, the topics mentioned occasionally, and the topics that the user has been consistently following for a long time by combining the ranking mechanism of the three dimensions, thus providing an accurate priority basis for subsequent push decisions. For example, if a topic in the newly extracted interest features already has a corresponding node in the long-term interest model, then the existing interest level value of the node and the newly calculated interest level value are subjected to exponential moving average or other smoothing fusion processing to avoid the excessive influence of an occasional mention in a single chat on the model; at the same time, the last access timestamp of the node is updated to the current time, and its frequency value is incremented. If the newly extracted interest features include a new topic that does not yet exist in the long-term interest model, then a new node is created for the topic in the model, and an initial interest level value is assigned, the current time is used as the last access timestamp, and the frequency value is initialized to 1. The long-term interest model update should also include decay processing for outdated interests. A global decay operation should be triggered at a preset update cycle (e.g., every 24 hours, every 7 days): all nodes in the long-term interest model are traversed, and for nodes whose last access timestamp is more than a preset decay threshold (e.g., 30 days), their interest level value is reduced by a preset decay factor (e.g., multiplied by 0.85). When the interest level value of an interest falls below a preset minimum threshold (e.g., 0.05) after multiple decays, the node can be marked as dormant or removed from the active model to control the unlimited growth of the model size.
[0010] The weighted overall score can be calculated using the following formula: Score(topic)=α×Interest_Level(topic)+β×Time_Recency(topic)+γ×Frequency(topic); Wherein, α, β, and γ are the weights for the interest level dimension, the time freshness dimension, and the frequency dimension, respectively, and their sum can be 1. Interest_Level(topic) is the interest level value of the topic (normalized to 0-1), Time_Recency(topic) is the freshness decay function value calculated based on the last access time (e.g., exp(-Δt / τ), where Δt is the difference between the current time and the last access time, and τ is the time decay constant), and Frequency(topic) is the normalized value of the frequency of occurrence (e.g., min(number of occurrences / preset upper limit, 1)). The specific values of α, β, and γ can be adjusted according to the different emphases of "current enthusiasm," "recency effect," and "long-term preference" in the actual application scenario. For example, for news and information push scenarios, the value of β can be appropriately increased to emphasize freshness; for audio-visual entertainment recommendation scenarios, the value of γ can be appropriately increased to take into account long-term preference. This application does not limit the specific values of α, β, and γ in its embodiments. Preferably, the long-term interest model is constructed with topics as nodes and the relationships between topics as edges. Each node includes at least one of the following attributes: degree of interest, last access time, and frequency of occurrence. The specific data structure of the long-term interest model is a graph structure in accordance with the above method.
[0011] Preferably, updating the user's long-term interest model based on interest characteristics includes: within a preset period, reducing the interest weight of nodes not mentioned in the long-term interest model and increasing the interest weight of newly appearing nodes. This not only records the user's interest in each independent topic, but also captures the semantic relationship between topics, thereby expanding the push scope along the relationship edge when making push decisions and discovering content that the user may be interested in.
[0012] Preferably, the step of obtaining the information to be pushed based on the long-term interest model includes: generating search keywords based on the long-term interest model, wherein the generation strategy can adopt single interest point retrieval or combined interest point retrieval; obtaining information related to the search keywords through a network interface; filtering the obtained information to obtain the information to be pushed, wherein single interest point retrieval refers to selecting the topic node with the highest current interest level value from the long-term interest model, and using the name of the topic itself or its synonyms as search keywords; combined interest point retrieval refers to selecting multiple topic nodes with the highest interest level value (e.g., the first three), combining their names or using them as independent search terms to initiate parallel search requests, and finally merging and deduplicating the search results. In the scenario of introducing a graph structure interest model, the generation of search keywords can also be extended along the associated edges in a one-hop manner, that is, from the node with the highest interest level value along its associated edges, the topic names of adjacent nodes with an association weight exceeding a preset edge weight threshold (e.g., 0.6) are included as additional search keywords.
[0013] Preferably, the filtering of the acquired information includes: sorting the acquired information based on at least one of the following: timeliness, relevance to the search keywords, and credibility of the information source; filtering the information to be pushed based on the sorting results; automatically generating search conditions from a long-term interest model and acquiring external information via a network interface, thus opening up a data pathway from internal interest perception to external content acquisition; the timeliness score (TimeScore) is calculated based on the difference between the information's publication time and the current time, with a smaller difference resulting in a higher score; the calculation method can use the exponential decay function TimeScore=exp(-Δt / T_half), where T_half is the timeliness half-life constant (e.g., 24 hours for news information and 7 days for knowledge content). The relevance score (RelScore) is obtained by calculating the semantic matching degree of the search keywords in the information title and summary text; specifically, it can use cosine similarity calculation based on word vectors or sentence vectors, or relevance scoring based on deep semantic matching models (such as DSSM, BERT). The Credibility Score is based on the historical credibility assessment of the information source domain. Credibility assessments can be based on third-party fact-checking databases or domain expert annotations, and the score range is 0 to 1. The TotalScore is a weighted sum of the scores from the above three dimensions: TotalScore=w1×TimeScore+w2×RelScore+w3×CredScore Where w1, w2, and w3 represent the weights for timeliness, relevance, and credibility, respectively. In chat content push scenarios targeting AI-powered bionic robot puppets, the relevance dimension is typically assigned the highest weight (e.g., w2 = 0.5) to ensure the pushed content closely matches user interests; the timeliness dimension is assigned the second highest weight (e.g., w1 = 0.3) to ensure content freshness; and the credibility dimension is assigned a basic weight (e.g., w3 = 0.2) to filter out false information. After sorting, the system selects several pieces of information with the highest overall scores (e.g., the top 3) as the candidate set of information to be pushed in this round.
[0014] Preferably, the push timing triggering condition includes at least one of the following: The current chat is silent for more than a preset duration; the relevance of the current chat content to the information to be pushed reaches a preset threshold; the user's expression of interest in a specific topic is detected from the real-time chat text. The setting of multiple triggering conditions enables the AI bionic robot to find a suitable push entry point in various scenarios such as silence gaps, natural topic transitions, and user active expression, rather than mechanically waiting for a fixed signal.
[0015] Preferably, the behavioral response includes: continuing to discuss the topic related to the information to be pushed, switching to other topics, and not responding; Updating the long-term interest model based on the behavioral response includes: When the behavioral response is to continue the discussion, the weight of the corresponding interest point is increased; when the behavioral response is to change the topic, the weight of the corresponding interest point is decreased; when the behavioral response is no response, the silent feedback is recorded and the push timing trigger conditions are adjusted. Through refined classification and differentiated processing of user feedback, the system can more accurately understand the user's true attitude towards different push content and push timing, thereby continuously optimizing the push strategy.
[0016] Preferably, the push timing triggering condition is met, including: actively generating a topic guidance statement based on at least one topic with a high degree of interest in the long-term interest model; and outputting the topic guidance statement in the form of a conversation message.
[0017] (III) Beneficial Effects This invention provides a method for pushing chat content to an AI-powered bionic robot based on user interest extraction. It has the following beneficial effects: 1. By extracting interest features from real-time chat text and continuously updating the long-term interest model, a two-layer modeling approach of real-time perception and long-term tracking of user interests is achieved. This ensures that the pushed content highly matches the user's current real interests, significantly improving personalization. The establishment of the proactive push mechanism, coupled with the ability to access external information online, fundamentally changes the existing chat system's passive response mode. It enables the AI-inspired robot to proactively initiate topics and introduce timely content, effectively extending user interaction time.
[0018] 2. The combination of multiple settings for push notification timing triggers with natural language generation technology ensures that push content blends naturally into the conversation flow in both style and timing, eliminating the abruptness of traditional push methods and improving the smoothness and naturalness of the user experience. Based on a user behavior-driven feedback update loop, the system possesses self-optimization capabilities. As usage time increases, the understanding of user preferences becomes more accurate, and the push strategy continuously improves, creating a positive cumulative effect. Detailed Implementation
[0019] 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.
[0020] This invention provides a method for pushing chat content to an AI-powered bionic robot based on user interest extraction. Includes the following steps: Step 1: Real-time extraction of user interest points Natural language processing is performed on real-time chat content to extract keywords, entities and topics mentioned by users. Real-time chat text can be obtained by automatic speech recognition of voice signals collected by smart interactive terminals, or by directly receiving text input by users through input devices such as keyboards and touch screens. Chat text streams are usually transmitted to the cloud service layer sentence by sentence or segment by segment in turn of conversation. Analyze users' tone, emotional inclination, and depth of discussion to determine their level of interest in the topic; The extracted points of interest are weighted and sorted according to their degree of interest, time freshness, and frequency of occurrence to construct a real-time list of user interests; Step 2: Construction and Update of User Interest Graph Based on users' historical chat history and real-time interest lists, a long-term interest graph of users is constructed. Interest graphs use topics as nodes and the relationships between topics as edges. Each node contains attributes such as interest level, last visit time, and frequency of occurrence. Regularly update the user interest graph, reducing the weight of interests that have not been mentioned for a long time, and increasing the weight of newly emerging interests.
[0021] Step 3: Online search and filtering of relevant information Generate search keywords based on the user's real-time interest list and long-term interest graph; The system uses a backend network interface to call a search engine to retrieve the latest information related to the keywords. The retrieved information is filtered and preprocessed: advertisements, vulgar and false information are filtered out, and the core content, publication time and source credibility of the information are extracted; Information is sorted based on its timeliness, relevance, and source credibility to select the most suitable content for push notifications.
[0022] Step 4: Judging the timing of the chat and naturally integrating the content Real-time monitoring of chat status to determine if there is a suitable time to send a push notification: When there is a brief silence in the chat, when the current chat topic is highly relevant to the information to be pushed, or when the user clearly expresses interest in a certain topic, natural language generation technology is used to transform the filtered information content into conversational expressions. The converted content is seamlessly integrated into the current chat conversation, avoiding abrupt insertion; Step 5: Feedback and Optimization of Push Notifications Record the content, timing, and user feedback (such as continued discussion, topic change, no response, etc.) for each push notification. Adjust the weighting of interest points and push strategies based on user feedback; We continuously optimize the content generation model to improve the naturalness and relevance of the pushed content.
[0023] Implementation plan: Implementation Plan 1: Topic-Related Push Notifications. When a user mentions "What are some good movies to watch lately?" in a chat, the system extracts the interest point "movies," searches online for information on recently released popular movies, and naturally incorporates it into the answer: "There are a few well-received movies released recently, such as 'XX,' which has a Douban rating of 8.5 and is a sci-fi thriller. Are you interested?" Implementation Plan 2: Proactive Push During Silence. When there is silence for more than 10 seconds in a chat, the system selects a topic with high interest and recent news from the user's interest graph and proactively initiates a chat: "By the way, that band you used to like recently released a new album. I listened to it and it's pretty good."
[0024] Implementation Plan 3: Personalized Push Notifications. Based on users' long-term interest profiles, the system identifies that users have a high interest in the topics of "technology" and "travel," regularly retrieves the latest information in these two fields, and pushes it to users at appropriate times.
[0025] 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. A method for pushing chat content to an AI-powered bionic robot based on user interest extraction, characterized in that, This includes extracting users' interest features from real-time chat text; The user's long-term interest model is updated based on the aforementioned interest features, wherein the long-term interest model includes information on the user's degree of interest in each topic and time attribute information. The information to be pushed is obtained based on the long-term interest model. In response to the fulfillment of the push timing trigger condition, the information to be pushed is converted into a conversational message using natural language generation technology, and the conversational message is inserted into the current chat conversation stream; Obtain the user's behavioral response to the session message, and update the long-term interest model based on the behavioral response.
2. The AI bionic robot chat content push method based on user interest extraction according to claim 1, characterized in that, The step of extracting user interest features from real-time chat text includes: extracting at least one of keywords, entities, and topics from the real-time chat text; and determining the degree of interest corresponding to each of the keywords, entities, and topics. Determining the degree of interest corresponding to each keyword, entity, and topic includes: analyzing at least one of the tone features, sentiment tendencies, and discussion depth in the real-time chat text; and determining the degree of interest based on the analysis results.
3. The AI bionic robot chat content push method based on user interest extraction according to claim 2, characterized in that, Updating the user's long-term interest model based on the interest features includes: weighting and sorting the extracted interest features according to the degree of interest, time freshness, and frequency of occurrence to obtain the sorting result, and updating the weight of the corresponding topic in the long-term interest model according to the sorting result.
4. The AI bionic robot chat content push method based on user interest extraction according to claim 3, characterized in that: The long-term interest model is constructed with topics as nodes and the relationships between topics as edges. Each node contains at least one of the following attributes: degree of interest, last access time, and frequency of occurrence.
5. The AI bionic robot chat content push method based on user interest extraction according to claim 4, characterized in that: Updating a user's long-term interest model based on interest characteristics includes: within a preset period, reducing the interest weight of nodes not mentioned in the long-term interest model and increasing the interest weight of newly appearing nodes.
6. The AI bionic robot chat content push method based on user interest extraction according to claim 5, characterized in that, The step of obtaining the information to be pushed based on the long-term interest model includes: generating search keywords based on the long-term interest model; obtaining information related to the search keywords through a network interface; and filtering the obtained information to obtain the information to be pushed.
7. The AI bionic robot chat content push method based on user interest extraction according to claim 6, characterized in that, The process of filtering the acquired information includes: sorting the acquired information based on at least one of the following: timeliness of the information, relevance to the search keywords, and credibility of the information source; and filtering the information to be pushed based on the sorting results.
8. The method for pushing chat content to an AI bionic robot based on user interest extraction according to claim 7, characterized in that, The push notification timing triggering conditions include at least one of the following: The current chat is silent for more than a preset duration; the relevance of the current chat content to the information to be pushed reaches a preset threshold; The user's expression of interest in a specific topic was detected from the real-time chat text.
9. The AI bionic robot chat content push method based on user interest extraction according to claim 8, characterized in that, The behavioral responses include: continuing to discuss the topic related to the information to be pushed, switching to other topics, and not responding; Updating the long-term interest model based on the behavioral response includes: When the behavioral response is to continue the discussion, the weight of the corresponding interest point is increased; when the behavioral response is to change the topic, the weight of the corresponding interest point is decreased; when the behavioral response is no response, the silent feedback is recorded and the push timing triggering conditions are adjusted.
10. The method for pushing chat content to an AI bionic robot based on user interest extraction according to claim 9, characterized in that, The conditions for triggering the push notification are met, including: actively generating a topic guidance statement based on at least one topic with a high degree of interest in the long-term interest model; and outputting the topic guidance statement in the form of a conversation message.