Method and system for optimizing content visibility in ai responses

Abstract

The application discloses an AI response content visibility optimization method and system, including: first, using a web crawler to collect multi-modal source content from multiple channels and preprocessing and sorting the theme; then, based on a general sentiment dictionary, quantifying content feature types, locating sentiment words and extracting theme keywords; then, using keyword search to retrieve content, optimizing content by fusing multi-modal features; then, building a knowledge graph, mining content potential associations, and calculating the correlation with user inquiries; then, continuously monitoring and analyzing user interaction flow, extracting key dynamic anchors, calculating semantic distance and converting it into logical assertions; finally, based on a deep learning model, predicting and optimizing content with dynamic anchors as input, generating neural ranking scores to present the results; this method comprehensively and deeply analyzes content, fuses multi-modal features, mines potential associations, tracks user interests, etc., to achieve personalized and accurate display, improve user experience and information acquisition efficiency, and provide intelligent content services for users.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a method and system for optimizing content visibility in AI responses. Background Technology

[0002] In today's era of widespread AI technology application, the quality and visibility of AI-generated response content are crucial to user experience. Currently, while multimodal AI response content is abundant and diverse, it suffers from disorganized content and unclear themes, making it difficult for users to quickly obtain the information they need. Traditional methods often rely on manual annotation to organize themes, which is inefficient and struggles to cover massive amounts of content. Furthermore, the quantification of AI response content features lacks domain-specific sentiment word processing, resulting in inaccurate feature type quantification and an inability to accurately reflect the content's sentiment. Therefore, there is an urgent need for an efficient method to collect and preprocess multimodal source content, organize clear themes, and accurately quantify content features to improve the visibility of AI response content.

[0003] As user interactions with AI become increasingly frequent, real-time, multi-turn interaction streams contain a wealth of user intent and focus information. However, existing AI response systems struggle to effectively monitor and analyze these interaction streams, failing to extract key dynamic anchors and thus accurately grasp current user needs. Furthermore, existing knowledge graph construction methods are inadequate, making it difficult to accurately map optimized content onto the graph and limiting the ability to uncover potential content relationships. There is also a lack of scientifically effective methods for calculating content relevance based on user queries. Moreover, the accuracy of predicting optimized content that satisfies user intent based on existing technologies is low, making it difficult to generate accurate and reasonable neural ranking scores, thus impacting the optimization effect of AI response content visibility. Summary of the Invention

[0004] This application provides a method and system for optimizing content visibility in AI responses. Multimodal data fusion has become the core direction for optimizing content visibility. By parsing structural information such as text, images, and audio in a unified format, personalized recommendations can be achieved, solving the problems of information overload and blurred focus, thereby enhancing the practical value of AI responses.

[0005] This application provides a method for optimizing content visibility in AI responses, the method comprising:

[0006] We utilize web crawling technology to collect multimodal source content of AI responses from multiple channels, and then sort out the themes of the multimodal source content after preprocessing.

[0007] Based on a general sentiment dictionary and domain-specific corpus of document topics, domain-related sentiment words are filtered and added to quantify the feature types of AI response content; the position of sentiment qualitative descriptive words in the text is located, and the text topic keywords are extracted.

[0008] Relevant topic content is retrieved from multimodal source content using topic keywords. After filtering, the most matching content is selected and added to the region based on the semantics and sentiment of the qualitative description region. The text is converted into vector representation, and the feature vectors of different modalities are fused using a concatenation method to form optimized content.

[0009] An initial knowledge graph is constructed by identifying entity concepts and relationships from multimodal source content. Entity concepts in the optimized content are mapped to the graph to uncover potential associations within the content. Based on the semantics of user queries and entity relationships extracted, relevant information is queried in the knowledge graph, and the relevance between the content and the query is calculated accordingly.

[0010] Continuously monitor and analyze real-time, multi-round user interaction flows, extract key dynamic anchors from the interaction flows; compare the dynamic anchors with the content of the initial knowledge graph, calculate the semantic distance between the current user focus and the original content topic, and convert the numerical value into interpretable logical assertions;

[0011] Based on a deep learning model, using dynamic anchor points as input, the system predicts optimized content that satisfies the user's current intent, generates a preliminary neural ranking score, and presents the results.

[0012] Preferably, the step of identifying entity concepts and relationships from preprocessed multimodal source content to construct an initial knowledge graph includes: performing word segmentation on user queries, splitting sentences into individual words, and labeling the part of speech of each word; determining the grammatical relationships between the components of the sentence through syntactic analysis; extracting entity relationships from the semantic analysis results, using a convolutional neural network to take the semantically analyzed sentence features as model input, and outputting entity relationship categories; matching the extracted entities with nodes in the knowledge graph, and performing queries in the knowledge graph using a graph query language based on the extracted entity relationships.

[0013] Preferably, the step of using a graph query language in the knowledge graph specifically involves: statistically analyzing the co-occurrence of entities in the query results and entities in the user query, calculating the proportion of co-occurring entities to the total number of entities in the query and the total number of entities in the query results; vectorizing the text information in the user query and query results, and calculating the cosine similarity between the two vectors; calculating the semantic similarity between entities in the query and entities in the query results using a word vector model; and combining the correlation based on entity co-occurrence and the correlation based on semantic similarity, using the following formula: .

[0014] Preferably, the deep learning model-based approach using dynamic anchors as input includes the following steps: The dynamic anchor is the smallest semantic unit representing the user's current focus; the deep learning model, using dynamic anchors as input, quickly predicts which optimized content satisfies the user's current intent, generating a preliminary neural ranking score; based on the knowledge graph, business rules, and semantic distance, symbolic reasoning is performed to verify the reasonableness of the ranking results from the neural prediction channels; if the results from the two channels are consistent, the result is directly adopted and an explanation is generated; if the results from the two channels conflict, an arbitration mechanism is initiated to prioritize rule constraints and generate a counterfactual explanation.

[0015] Preferably, the continuous monitoring and analysis of real-time, multi-round user interaction flows includes: collecting all user interaction data in different sessions; using deep behavioral analysis technology to deeply analyze user interaction data and behavioral patterns; employing a temporal neural network model to encode the identified user behavior patterns, transforming behavioral features in different scenarios into a unified feature representation for fusion, and extracting feature vectors of users' core value orientations; constructing a unique value graph for each user based on the extracted user value features; performing value semantic analysis on the question-and-answer content of user interactions each time content visibility is optimized, matching the parsed value orientation with the value graph, and calculating the value fit; dynamically adjusting the semantic distance based on the value fit, and then updating the content visibility ranking in a floating manner.

[0016] Preferably, the step of constructing a unique value graph for each user includes: the value graph uses abstract value concepts as nodes, and connects each node through weighted edges, where the weight of the edge represents the strength of the association between value concepts, and changes in the edge reflect the evolution of value concepts; specifically: determining abstract value concepts as nodes in the value graph, assuming the node set is... , where k is the number of nodes; through statistical analysis of a small number of samples, the weights of the edges between each node are initialized. Construct the initial weight matrix When a user generates new interaction data, the value map is updated based on the feature vector V extracted from the new interaction data; let the update function be U, and the updated weight matrix be... .

[0017] Preferably, the calculation of value fit includes: performing value semantic analysis on user question-and-answer content each time content visibility optimization is performed, and extracting the value tendency feature vector Q of the question-and-answer content through a deep learning model; matching the parsed value tendency feature vector Q with the node features in the value graph, and calculating the value fit S; assuming that the nodes in the value graph... The feature vector is The value fit S is calculated by comparing Q with each... The cosine similarity is obtained, that is Take the maximum value As the ultimate value alignment.

[0018] Preferably, the collection of all user interaction data in different sessions includes: recording and preprocessing each user's interaction behavior data during the use of the AI ​​response system; extracting features reflecting individual value orientations from the preprocessed data, and calculating the individual's value orientation weights for each feature based on the frequency and importance of user interactions on specific features using a variant of the TF-IDF algorithm; dividing users into different subgroups based on the relationships and interactions between users; calculating the subgroup's value orientation for each feature by comprehensively considering the value orientations of individuals within the subgroup; integrating the value orientations of each subgroup to calculate the overall group's value orientation for each feature; merging the value orientations of individuals, subgroups, and the entire group to form a comprehensive multi-level value assessment system; and sorting and filtering the content in the AI ​​response system based on the merged multi-level value assessment system to optimize content visibility.

[0019] Following the established strategy of introducing authoritative content and optimizing fluency, the text to be optimized is initially modified. Using a multimodal feature representation model, the initially optimized text is analyzed again to establish a comprehensive evaluation index system for authority and fluency.

[0020] Preferably, the formation of a comprehensive multi-level value assessment system specifically includes: continuously collecting various interaction data of users in the AI ​​response system, extracting features that reflect the group's value orientation, arranging the extracted group value orientation features in chronological order to form time series data; training a long-term value evolution model using historical time series data, adjusting the model parameters to accurately fit historical data; predicting future value orientation trends of the group using the trained long-term value evolution model; defining individual and group value difference measurement indicators to measure the difference between individual value orientation and the group's average value orientation; setting warning thresholds for each indicator, indicating value conflict when the indicator value exceeds the threshold, and then simulating different value integration and content recommendation strategies by combining the prediction results of the long-term value evolution model and the value conflict warning information.

[0021] Furthermore, embodiments of the present invention also provide an optimization system for content visibility in AI responses, the system comprising:

[0022] The data collection module is used to collect multimodal source content of AI responses from multiple channels using web crawling technology, and then sort out the themes of the multimodal source content after preprocessing.

[0023] The analysis module is used to filter and add domain-related sentiment words based on a general sentiment dictionary combined with domain corpus of document topics, quantify the feature types of AI response content, locate the position of sentiment qualitative descriptive words in the text, and extract text topic keywords;

[0024] The fusion module is used to retrieve relevant topic content from multimodal source content using topic keywords, and after filtering, selects the most matching content based on the semantics and sentiment of the qualitative description area and adds it to that area; it converts the text into vector representation and uses a concatenation method to fuse the feature vectors of different modalities to form optimized content;

[0025] The mining module is used to identify entity concepts and relationships from preprocessed multimodal source content to construct an initial knowledge graph, map entity concepts in optimized content to the graph, and mine potential associations in the content; based on the semantics of user queries and extract entity relationships, relevant information is queried in the knowledge graph, and the relevance between the content and the query is calculated accordingly;

[0026] The computation module is used to continuously monitor and parse real-time, multi-round user interaction streams, extract key dynamic anchors from the interaction streams, compare the dynamic anchors with the content of the initial knowledge graph, calculate the semantic distance between the current user focus and the original content topic, and convert the numerical value into interpretable logical assertions.

[0027] The prediction module is used to predict optimized content that meets the user's current intent based on a deep learning model with dynamic anchor points as input, generate a preliminary neural ranking score, and present the results.

[0028] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0029] By collecting and deeply analyzing multimodal content from multiple channels, we can comprehensively and accurately grasp information; identify emotional and quantitative features to provide strong support for optimization; integrate multimodal features to optimize content, making it rich and attractive; construct knowledge graphs to mine potential connections and expand information dimensions; and dynamically adjust the display order and impression score based on relevance and user queries to achieve personalized and accurate display, greatly improving user experience and information acquisition efficiency. Through dynamic anchor points and semantic distance calculation, we can achieve real-time tracking of user interests and accurate judgment of intent drift, thereby providing users with a more personalized and intelligent content service experience.

[0030] By deeply analyzing user interaction patterns and combining them with temporal neural networks to extract core value orientation features and construct a unique value graph, we can accurately grasp users' intrinsic value orientations. When optimizing content visibility, we match the value orientation of Q&A content with the value graph to calculate the fit, and dynamically adjust the content ranking accordingly. This ensures that the content presented to users better aligns with their core value orientations, effectively improving the match between content and user needs, enhancing user experience, and increasing the efficiency and accuracy of information acquisition.

[0031] By constructing a multi-domain knowledge graph, the text is accurately linked to authoritative knowledge expressed in unofficial statements, effectively improving the accuracy and credibility of the content. By integrating multimodal features and optimizing the introduction strategy, the authoritativeness and text fluency are balanced, avoiding logical breaks caused by information supplementation or modal introduction. Finally, the optimization effect is quantified through a comprehensive evaluation index system, achieving a synergistic improvement in authoritative knowledge coverage, language coherence, and multimodal integration, significantly enhancing the information value and readability of the text.

[0032] By constructing a multi-layered value assessment system that comprehensively considers the value orientations of individuals, subgroups, and the entire group, the needs of users at different levels can be grasped more comprehensively and accurately. When dealing with multi-user scenarios, the system effectively integrates multiple value systems, avoids value conflicts, and makes the content recommended by the AI ​​response system more aligned with the overall needs of the group. This significantly improves the matching degree between content and user needs, optimizes content visibility, and provides users with a higher-quality and more targeted service experience. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the method for optimizing content visibility in AI responses according to an embodiment of the present invention.

[0034] Figure 2 This is a flowchart illustrating the system for optimizing content visibility in AI responses according to an embodiment of the present invention. Detailed Implementation

[0035] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.

[0036] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0038] Example 1: Figure 1 This is a flowchart illustrating the method for optimizing content visibility in AI responses according to an embodiment of the present invention.

[0039] like Figure 1 As shown, the method for optimizing content visibility in AI responses includes the following steps:

[0040] S1 utilizes web crawling technology to collect multimodal source content of AI responses from multiple channels, and then sorts out the themes of the multimodal source content after preprocessing.

[0041] Specifically, the TF-IDF algorithm is used to calculate the importance of each word in the text and identify candidate topic words to determine the text topic.

[0042] The multimodal source content includes, but is not limited to, text, images, and audio. The received source content is analyzed in depth to identify the article's theme, paragraph structure, and key information points (including sentence segmentation, word segmentation, part-of-speech tagging, and named entity recognition).

[0043] By leveraging web crawling technology, specific rules are set for different channels (such as web pages, API interfaces, etc.) to capture multimodal data such as text, images, and audio from AI responses. For example, for web pages, the tag elements containing AI response content are located and extracted by analyzing the web page structure; for audio data, speech recognition technology can be used to initially convert it into text for subsequent unified processing.

[0044] The collected multimodal source content is preprocessed. The unified encoding format is UTF-8, and special characters, extra spaces and other irrelevant information in the text are removed; the images are uniformly converted to common formats (such as JPEG) and adjusted to a uniform size for subsequent analysis and processing; the text converted from speech recognition is formatted in a unified manner, while retaining key parameters such as the sampling rate and number of channels of the original audio.

[0045] The importance of each word in the text is calculated using the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm. For a document set D containing N documents, the TF-IDF value of word t is calculated using the following formula:

[0046]

[0047] Where TF(t, d) represents the frequency of word t in document d, and the calculation formula is: , It is the number of times word t appears in document d. It is the sum of the occurrences of all words in document d; IDF(t) is the inverse document frequency, calculated using the formula: , It is the total number of documents in document set D. This represents the number of documents containing the word t. Based on the TF-IDF values, words with higher weights in each document are selected as candidate topic words. Then, a clustering algorithm (K-Means clustering) is used to aggregate similar candidate topic words to form the document's topic.

[0048] For text content, paragraph division is determined by analyzing formatting information such as line breaks and indentation, combined with semantic coherence analysis (e.g., using cosine similarity to calculate the semantic similarity between adjacent sentences). The formula for calculating the cosine similarity between sentence i and sentence j is:

[0049]

[0050] in, and These are the weights of the k-th word in sentences i and j, respectively (which can be calculated using TF-IDF values), and n is the total number of words in the sentence.

[0051] The text is segmented into sentences and words using a natural language processing tool (jieba word segmentation library). A part-of-speech tagging method based on a statistical model (Hidden Markov Model, HMM) is employed to determine the part of speech of a word based on its probability of occurrence in the context. Named entity recognition is performed using a Conditional Random Field (CRF) model. Given an input sequence... (Word sequences in the text) and their corresponding output sequences (Part-of-speech or named entity tag sequence), the CRF model makes predictions by calculating the conditional probability P(y|x), the formula is: in, It is a normalization factor. These are the parameters of the model. It is a characteristic function.

[0052] Image feature extraction algorithms (SIFT) are used to extract key feature points from images, and feature matching algorithms (FLANN matchers) are used to analyze key information such as objects and scenes in the images. For speech and audio, in addition to the previously mentioned speech recognition and text conversion for analysis, acoustic features (Mel frequency cepstral coefficients, MFCCs) can be extracted from the audio. Machine learning algorithms (Support Vector Machine, SVM) are then used to classify the audio and identify key events (such as laughter, applause, etc.). The calculation process of MFCCs includes steps such as pre-emphasis, framing, Hamming windowing, Fast Fourier Transform (FFT), Mel filter bank filtering, logarithmic operations, and Discrete Cosine Transform (DCT).

[0053] S2 uses a general sentiment dictionary combined with domain-specific corpora of document topics to filter and add domain-related sentiment words, quantifying the feature types of AI response content; it also uses sentiment qualitative descriptive words to locate their positions in the text and extract text topic keywords.

[0054] Specifically, a general sentiment lexicon is collected as a foundation. Domain-specific corpora (if the AI ​​response content has a particular domain bias, such as food, travel, etc.) are combined to filter and add domain-related sentiment words, enriching the sentiment lexicon. For example, in the food domain, "delicious" and "signature dish" are positive sentiment words, while "not recommended" and "salty" are negative sentiment words. The parsed text is then segmented, using existing segmentation tools (such as jieba segmentation). Stop words are removed; common Chinese stop word lists, such as the Harbin Institute of Technology stop word list, can be used to reduce interference from irrelevant words in sentiment analysis. The segmented text is then traversed, and each word is matched against the constructed sentiment lexicon. If a match is successful, the word is labeled as a sentiment word, and its sentiment polarity (positive, negative, or neutral) is recorded.

[0055] Based on the characteristics of the AI ​​response content, identify the types of quantitative features that may appear, such as numbers, percentages, ratios, and time intervals. For example, in AI responses to technology product reviews, quantitative statements such as "battery life up to 12 hours" or "performance improved by 30%" may appear. Write corresponding regular expressions for different types of quantitative features. For numbers, the regular expression can be written as `\d+`; for percentages, it can be written as `\d+%`; for ratios, such as "1:2", it can be written as `\d+:\d+`; for time intervals, such as "3 days" or "5 hours", corresponding regular expressions can be written for matching. Use regular expressions to search and match in the text to extract the quantitative features that meet the requirements. For the extracted quantitative features, analyze their context to determine their specific meaning. For example, "price reduced by 20%" needs to be clarified as a 20% reduction relative to the original price. Analyzing the surrounding vocabulary and sentence structure can aid in understanding.

[0056] Based on the sentiment descriptors identified by S2, their positions in the text are located, and text paragraphs containing these sentiment descriptors and their surrounding related semantics are defined as the sentiment description regions. Topic keywords are extracted from these regions using the TF-IDF algorithm. The TF-IDF value of each word in the sentiment description region is calculated, and words with higher values ​​are selected as topic keywords.

[0057] For example, in calculating the sentiment tendency of qualitative descriptive words (quantifying the intensity of sentiment tendency), suppose text D contains n words, among which... A positive emotional word, One negative emotion word. Intensity of positive emotion tendency. and intensity of negative emotional tendency They can be calculated separately as follows:

[0058]

[0059] The overall emotional intensity S of the text can be calculated as S= If S>0, the overall sentiment of the text tends to be positive; if S<0, the overall sentiment of the text tends to be negative; if S=0, the sentiment of the text tends to be neutral.

[0060] Different sentiment words may have different sentiment intensities, and a weight can be assigned to each sentiment word. (This can be determined through manual annotation or corpus-based statistical methods). Intensity of positive sentiment tendency. ,in This represents the weight of the i-th positive sentiment word. The intensity of negative sentiment tendency. ,in This represents the weight of the j-th negative sentiment word. Overall sentiment intensity. The judgment rules are the same as those for the word frequency-based method described above.

[0061] In calculations related to quantitative features, if the text mentions a change in a certain quantitative feature, such as "production increased from 100 tons to 150 tons," then the rate of change R can be calculated as follows:

[0062]

[0063] in This is the original quantitative value (100 tons). This is the changed quantitative value (150 tons).

[0064] S3: Retrieve relevant topic content from multimodal source content using topic keywords, and after filtering, select the most matching content based on the semantics and sentiment of the qualitative description area and add it to the area; convert the text into vector representation, and use a concatenation method to fuse the feature vectors of different modalities to form optimized content.

[0065] Specifically, the extracted topic keywords are used to search through multimodal source content, and relevant topic content is obtained by ranking the keywords according to their relevance. For example, if the topic keyword is "mobile phone appearance," then content related to mobile phone appearance design, materials, and colors will be retrieved. The retrieved relevant topic content is then filtered to remove duplicate, irrelevant, or low-quality content. Based on the semantic and sentiment tendencies of the qualitative description area, the most matching content is selected, integrated, and added to the qualitative description area.

[0066] In addition to extracting textual features such as keywords, a word vector model (Word2Vec) is used to convert the text into vector representations to capture its semantic information. If the qualitative description region involves image content, a convolutional neural network (CNN) is used to extract feature vectors from the image. For example, a pre-trained VGG16 model is used, with the last fully connected layer removed, and the image is input into the model to obtain feature vectors. If audio content is present, features such as Mel-frequency cepstral coefficients (MFCCs) are extracted.

[0067] A concatenation method is used to fuse feature vectors from different modalities. Assume the text feature vector is... The image feature vector is The audio feature vector is The fused feature vector for: .

[0068] Data such as quantity, proportion, and frequency from the added related thematic content are collected and organized. The added related thematic content, the fused multimodal feature information, and the data statistics are integrated into the original qualitative description area to form the optimized content.

[0069] S4. Identify entity concepts and relationships from the preprocessed multimodal source content to construct an initial knowledge graph. Map the entity concepts in the optimized content to the graph and mine potential associations in the content. Extract entity relationships based on the semantics of user queries, query relevant information in the knowledge graph, and calculate the correlation between the content and the query.

[0070] Specifically, entity concepts are extracted from preprocessed multimodal data, and the same entity identified by different modalities is standardized. Relation extraction rules are formulated based on domain knowledge and language conventions, and a convolutional neural network is used to train a relation extraction model to extract entity relations. A graph database is chosen for storage, and entity concepts are used as nodes, while relations are used as edges to construct a knowledge graph.

[0071] Entity recognition is performed on the optimized content, and the identified entities are matched with nodes in the knowledge graph. Entity name similarity is calculated using string matching and Levenshtein distance algorithms. If the similarity exceeds a threshold, the entities are considered the same and mapped accordingly. For entities that cannot be directly matched, mapping is performed using contextual information and semantic similarity calculations (using Word2Vec to calculate entity vectors and cosine similarity).

[0072] In the knowledge graph, depth-first search (DFS) or breadth-first search (BFS) is used to query and optimize the paths between entities in the content. The Apriori association rule mining algorithm is used to mine frequent co-occurrence patterns and association rules between entities, and support and confidence are calculated. The LouvAIn community detection algorithm is used to divide the knowledge graph into different communities, and the entity relationships within the communities are analyzed to mine potential associations.

[0073] The system performs word segmentation on user queries, breaking the sentence down into individual words and labeling each word with its part of speech (e.g., noun, verb, adjective). Syntactic analysis determines the grammatical relationships between the sentence's components, such as subject-verb-object relationships. Further analysis of the semantic role of each word clarifies the semantic information such as the actor and the recipient of the action.

[0074] Entity relations are extracted from the semantic analysis results. A convolutional neural network (CNN) is used as input to the semantically analyzed sentence features (such as word vectors, part-of-speech vectors, and syntactic features), and the output is the entity relation category. The extracted entities are matched with nodes in the knowledge graph. For entities that cannot be directly matched, mapping is performed through contextual information and semantic similarity calculation. Based on the extracted entity relations, a graph query language is used to perform queries in the knowledge graph.

[0075] This function analyzes the co-occurrence of entities in the query results with those in the user query. It calculates the proportion of co-occurring entities to the total number of entities in the query and the total number of entities in the query results. For example, if the user query contains 3 entities and the query results contain 5 entities, and 2 of these entities co-occur, then the co-occurrence ratio is [percentage missing]. This can be used as a reference indicator for the degree of correlation.

[0076] The text information in user queries and query results is vectorized using the BERT language model. The cosine similarity between the two vectors is then calculated as part of the semantic similarity. For entities in the query and query results, a word vector model is used to calculate their semantic similarity. The relevance based on entity co-occurrence and semantic similarity is then combined. Different weights can be assigned to them (e.g., entity co-occurrence relevance weight is 0.3, semantic similarity relevance weight is 0.7), and a weighted average is calculated as the final relevance between the content and the query. The formula is:

[0077]

[0078] S5 continuously monitors and analyzes real-time, multi-round user interaction flows, extracts key dynamic anchors from the interaction flows, compares the dynamic anchors with the content of the initial knowledge graph, calculates the semantic distance between the current user focus and the original content topic, and converts the numerical value into interpretable logical assertions.

[0079] Dynamic anchors are the smallest semantic units that represent a user's current focus, such as "the user just asked about the side effects of the drug," "the user skipped all theoretical overview content," and "the user spent far more time on the 'carbon neutrality' topic page than on other topics." The output is a set of dynamic anchors with timestamps and weights (determined by the intensity of the behavior), outlining the evolution of the user's real-time interests.

[0080] For example, if a significant increase in "semantic distance" is detected, the logical expression INTENT_DRIFT(CurrentFocus, "side_effects", OriginalFocus, "efficacy") is generated (the intent has drifted from "efficacy" to "side effects"). The output is a formalized, machine-readable logical rule and assertion describing the drift of the current user state relative to the initial state and the changes in the characteristics of the content itself.

[0081] S6, based on a deep learning model and using dynamic anchor points as input, predicts optimized content that meets the user's current intent, generates a preliminary neural ranking score, and presents the results.

[0082] Specifically, based on a deep learning model (Transformer), and using dynamic anchor points as input, it quickly predicts which optimized content is most likely to satisfy the user's current intent, generating a preliminary "neural ranking score".

[0083] Based on the topic rules (such as "medical information should prioritize the display of authoritative content"), business rules (such as "advertisement content should be placed at the top"), and semantic distance in the knowledge graph, symbolic reasoning is performed to verify whether the ranking results of the neural prediction channel are reasonable.

[0084] If the results from the two channels are consistent, the result will be adopted directly and an explanation will be generated, such as "This content is displayed for you first because it comes from [authoritative source] and accurately matches your latest concern about 'side effects'"; if the results from the two channels conflict, an arbitration mechanism will be initiated to prioritize the rule constraints and generate a counterfactual explanation, such as "Although you may be interested in the opinions of online celebrities, the system has already pinned authoritative information from [National Medical Products Administration] to ensure your safety."

[0085] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0086] By collecting and deeply analyzing multimodal content from multiple channels, we can comprehensively and accurately grasp information; identify emotional and quantitative features to provide strong support for optimization; integrate multimodal features to optimize content, making it rich and attractive; construct knowledge graphs to mine potential connections and expand information dimensions; and dynamically adjust the display order and impression score based on relevance and user queries to achieve personalized and accurate display, greatly improving user experience and information acquisition efficiency. Through dynamic anchor points and semantic distance calculation, we can achieve real-time tracking of user interests and accurate judgment of intent drift, thereby providing users with a more personalized and intelligent content service experience.

[0087] Example 2: In Example 1, although multimodal source content of AI responses was collected through multiple channels and processed to optimize the content and construct a knowledge graph to improve the matching degree between content and user needs, it was found in practical application that Example 1 mainly focused on the static analysis and processing of AI response content, and did not adequately consider the dynamic changes during real-time user interaction. Different users have different behavioral patterns and value orientations during the interaction process, and the needs of the same user may also change in different scenarios. If content matching and recommendation are based solely on static analysis results, it is difficult to accurately capture changes in users' real-time needs, which may lead to a low degree of relevance between recommended content and users' current intentions. In order to better adapt to dynamic user interaction scenarios and accurately grasp users' real-time needs, it is necessary to introduce operations such as in-depth analysis of user interaction behavior and extraction of core value orientations. Example 2 is an optimization and improvement aimed at addressing this problem.

[0088] Therefore, the embodiments of this application are optimized based on the above embodiments.

[0089] In some embodiments, step S5, which involves continuously monitoring and parsing real-time, multi-round user interaction streams, further includes:

[0090] S51 collects all user interaction data across different sessions and uses deep behavioral analysis technology to deeply analyze user interaction data and behavioral patterns.

[0091] Specifically, this involves collecting various interaction data, such as text, clicks, and dwell time, from user interaction records across different sessions to construct a user interaction dataset. Where n represents the total amount of interactive data in the i-th session. This represents the j-th interaction data in the i-th session; construct a user interaction data set based on the user interaction dataset. Where m represents the total number of users, Let D represent the interaction dataset of the i-th session; each user's dialog window is a dataset, and each interaction within a dialog window is a data point. A recurrent neural network (RNN) is used to train the user interaction dataset D to mine user behavior patterns. Let the model be M, and the user behavior pattern P = M(D) is obtained from the input dataset D.

[0092] S52 uses a temporal neural network model to encode the identified user behavior patterns, transform the behavioral features in different scenarios into a unified feature representation for fusion, and extract the feature vector of the user's core value orientation.

[0093] Among them, the core value orientation is to extract the direction that users value when obtaining information, such as focusing on data accuracy, data integrity, time span, real-time information, or opinion information.

[0094] Specifically, a temporal neural network model is used to encode the identified user behavior pattern P. Let the encoding function be E, and the encoded behavior pattern representation E(P) is obtained after encoding.

[0095] The behavioral features from different scenarios (included in E(P)) are transformed into a unified feature representation and then fused. Assume the feature vectors from different scenarios are as follows: We use a weighted average method for fusion, with weights set as follows: And satisfy The fused feature vector .

[0096] From the fused feature vector F, extract the feature vector V that represents the user's core value orientation. Dimensionality reduction and feature extraction can be performed using methods such as Principal Component Analysis (PCA). Let A be the PCA transformation matrix, then V = AF.

[0097] S53 constructs a unique value map for each user based on extracted user value characteristics.

[0098] The value graph uses abstract value concepts as nodes, such as "environmental protection", "privacy awareness", and "deep thinking". Each node is connected by weighted edges. The weight of the edge represents the strength of the association between value concepts. Changes in the edge reflect the evolution of value concepts. The graph structure is updated in real time with each user interaction.

[0099] Specifically, abstract values ​​are identified as nodes in the value map, such as "environmental protection," "privacy awareness," and "deep thinking," and the set of nodes is denoted as . , where k is the number of nodes. Through statistical analysis of a small number of samples, the weights of the edges between nodes are initialized. (represents a node) and nodes (The strength of the correlation between them), constructing the initial weight matrix. When a user generates new interaction data, the value map is updated based on the feature vector V extracted from the new interaction data. Let the update function be U, and the updated weight matrix be... The specific update rules can be designed according to the actual situation. For example, if the value orientation reflected by the new interaction data is strengthened with the association between two nodes, the weight of the edge between these two nodes should be increased accordingly. At the same time, the rationality and effectiveness of the weight matrix should be ensured, such as that the weights are non-negative.

[0100] S54: Each time content visibility is optimized, value semantic analysis is performed on the question and answer content of user interaction, the parsed value orientation is matched with the value graph, and the value fit is calculated.

[0101] Specifically, during each content visibility optimization, value semantic analysis is performed on the user's question-and-answer content. Natural Language Processing (NLP) techniques such as word vector models (Word2Vec or GloVe) can be used to convert the user's interactive question-and-answer content into word vector representations, and then a deep learning model (convolutional neural network) is used to extract the value-oriented feature vector Q of the question-and-answer content.

[0102] The extracted value orientation feature vector Q is matched with the node features in the value map to calculate the value fit degree S. Let the value map contain nodes... The feature vector is The value fit S is calculated by comparing Q with each... The cosine similarity is obtained, that is Take the maximum value As the final value alignment, an alignment threshold θ can be set, with a reference range of [0.6, 0.9]. When S ≥ θ, the user's question and answer content is considered to have a high degree of alignment with the value graph.

[0103] S55 dynamically adjusts semantic distance based on value alignment, thereby updating the content visibility ranking in a floating manner.

[0104] Specifically, semantic distance is dynamically adjusted based on value fit S. Let the adjustment function be T, then A linear adjustment function is used. Where α and β are adjustment parameters that can be adjusted according to the actual situation so that when S is high, Smaller, and vice versa, while ensuring The range of values ​​is reasonable. Based on the adjusted semantic distance... The content visibility sort is updated using a floating method. Let the original content sort be... The updated sorting is , where G is the sorting update function, which can adjust the original sorting based on semantic distance, for example, placing content with smaller semantic distances at the beginning.

[0105] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0106] By deeply analyzing user interaction patterns and combining them with temporal neural networks to extract core value orientation features and construct a unique value graph, we can accurately grasp users' intrinsic value orientations. When optimizing content visibility, we match the value orientation of Q&A content with the value graph to calculate the fit, and dynamically adjust the content ranking accordingly. This ensures that the content presented to users better aligns with their core value orientations, effectively improving the match between content and user needs, enhancing user experience, and increasing the efficiency and accuracy of information acquisition.

[0107] Example 3: While Example 2 improved the matching degree between content and user needs to some extent by deeply analyzing user interaction behavior patterns, extracting core value orientation features, and constructing a value graph, it primarily focused on value analysis for a single user. However, in real-world applications, multiple users with different value graphs often use a single AI response system for decision-making, such as family travel planning or project teams choosing technology solutions. In such cases, the system struggles to handle the conflict and integration of multiple value systems and cannot optimize the collective visibility that is most beneficial to the group.

[0108] Therefore, the embodiments of this application are optimized based on the above embodiments.

[0109] In some embodiments, step S51, collecting all user interaction data across different sessions, further includes:

[0110] S511 records and preprocesses each user's interaction behavior data during the user's use of the AI ​​response system.

[0111] Preprocessing includes, but is not limited to, cleaning the collected data to remove noise and invalid information. For example, removing duplicate click records and correcting erroneous text input. Natural language processing operations such as word segmentation, part-of-speech tagging, and named entity recognition are performed on the text data to facilitate subsequent analysis.

[0112] S512 extracts features that reflect an individual's value orientation from the preprocessed data, and uses a variant of the TF-IDF algorithm to calculate the individual's value orientation weight for each feature based on the frequency and importance of the user's interaction with specific features.

[0113] S513 divides users into different subgroups based on the relationships and interactions between them.

[0114] Users are further divided into different subgroups based on their relationships and interactions, such as shared conversation topics and frequency of interaction. For example, in a family setting, they are divided according to factors such as age and interests; in a project team, they are divided according to factors such as members' responsibilities and areas of expertise.

[0115] S514, taking into account the value orientation of individuals within a subgroup, calculate the value orientation of the subgroup for each characteristic; then integrate the value orientations of each subgroup to calculate the value orientation of the entire group for each characteristic.

[0116] S515 integrates the value orientations of individuals, subgroups, and the entire group to form a comprehensive, multi-level value assessment system.

[0117] S516, based on the integrated multi-level value assessment system, sorts and filters the content in the AI ​​response system to optimize content visibility.

[0118] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0119] By constructing a multi-layered value assessment system that comprehensively considers the value orientations of individuals, subgroups, and the entire group, the needs of users at different levels can be grasped more comprehensively and accurately. When dealing with multi-user scenarios, the system effectively integrates multiple value systems, avoids value conflicts, and makes the content recommended by the AI ​​response system more aligned with the overall needs of the group. This significantly improves the matching degree between content and user needs, optimizes content visibility, and provides users with a higher-quality and more targeted service experience.

[0120] Example 4: In Example 3, when using existing data collection and preliminary analysis methods to obtain characteristics related to group value orientations and to determine value conflicts and formulate strategies, it was found that different groups exhibited significant differences in value orientation characteristics across different time dimensions, and the manifestations of individual and group value differences were complex and diverse. Existing judgment methods rely on relatively single and fixed indicators and models, making it difficult to comprehensively and accurately capture these dynamic changes and complex differences. Faced with the diverse problems arising from different group characteristics and varying degrees of individual and group value differences, using a unified judgment and processing strategy inevitably leads to inaccurate identification of value conflicts, insufficient targeting and effectiveness in strategy formulation, and an inability to achieve a good balance between individual and group values ​​and accurate adaptation of system content recommendations.

[0121] Therefore, the embodiments of this application are optimized based on the above embodiments.

[0122] In some embodiments, step S515, which forms a comprehensive multi-level value assessment system, further includes:

[0123] S5151 continuously collects various interaction data of users in the AI ​​response system, extracts features that can reflect the group's value orientation, and arranges the extracted group value orientation features in chronological order to form time series data.

[0124] S5152 uses historical time series data to train a long-term value evolution model and adjusts the model parameters to accurately fit the historical data.

[0125] S5153 uses a trained long-term value evolution model to predict the future trend of value orientation changes in a group.

[0126] S5154 defines indicators for measuring the difference in value between individuals and groups. These indicators are used to measure the difference between an individual's value orientation and the average value orientation of a group.

[0127] The definition of indicators for measuring the difference in value between individuals and groups includes, but is not limited to, indicators of difference in value orientation and indicators of difference in decision-making style. Indicators of difference in value orientation can be constructed based on factors such as users' attention to different topics and their evaluation tendencies, for example, calculating the semantic similarity between an individual's and a group's evaluation of a certain type of product; indicators of difference in decision-making style can be measured from dimensions such as decision-making speed and risk preference, such as statistically analyzing the average decision-making time of individuals and groups when faced with the same decision-making task.

[0128] S5155 sets an early warning threshold for each indicator. When the indicator value exceeds the threshold, it indicates that there is a value conflict. At this time, the prediction results of the long-term value evolution model and the value conflict early warning information are combined to simulate different value integration and content recommendation strategies.

[0129] For example, if predictions indicate that excessive integration of group preferences leads to a decline in individual innovation capabilities, while appropriately increasing recommendations of individual-related content can alleviate this problem and promote long-term team innovation, then the recommendation strategy is adjusted to increase the proportion of individual content. This chosen balancing strategy is applied to the value integration and content recommendation within the AI ​​response system. During implementation, user feedback and new interaction data are continuously collected, and the strategy is dynamically adjusted based on actual conditions. For instance, if the adjusted strategy is found to be ineffective in the short term, the data is re-analyzed and the strategy parameters are optimized.

[0130] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages:

[0131] By collecting user interaction data and extracting group value orientation characteristics, a long-term value evolution model is used to accurately predict future value change trends within the group. Simultaneously, by defining metrics for individual and group value differences and setting early warning thresholds, value conflicts can be detected promptly. Based on this, different strategies are simulated, and appropriate strategies are applied to the system. Feedback is continuously collected during implementation to adjust strategies, effectively balancing individual and group values, improving the accuracy and adaptability of the system's content recommendations, and promoting the long-term stable development of the system.

[0132] Furthermore, embodiments of the present invention also provide an optimization system for content visibility in AI responses.

[0133] Figure 2 This is a schematic diagram of the structure of the AI ​​response content visibility optimization system in an embodiment of the present invention.

[0134] like Figure 2 As shown, the AI ​​response content visibility optimization system includes: a collection module, an analysis module, a fusion module, a mining module, a calculation module, and a prediction module.

[0135] The data collection module is used to collect multimodal source content of AI responses from multiple channels using web crawling technology, and then sort out the themes of the multimodal source content after preprocessing.

[0136] The analysis module is used to filter and add domain-related sentiment words based on a general sentiment dictionary combined with domain corpus of document topics, quantify the feature types of AI response content, locate the position of sentiment qualitative descriptive words in the text, and extract text topic keywords;

[0137] The fusion module is used to retrieve relevant topic content from multimodal source content using topic keywords, and after filtering, selects the most matching content based on the semantics and sentiment of the qualitative description area and adds it to that area; it converts the text into vector representation and uses a concatenation method to fuse the feature vectors of different modalities to form optimized content;

[0138] The mining module is used to identify entity concepts and relationships from preprocessed multimodal source content to construct an initial knowledge graph, map entity concepts in optimized content to the graph, and mine potential associations in the content; based on the semantics of user queries and extract entity relationships, relevant information is queried in the knowledge graph, and the relevance between the content and the query is calculated accordingly;

[0139] The computation module is used to continuously monitor and parse real-time, multi-round user interaction streams, extract key dynamic anchors from the interaction streams, compare the dynamic anchors with the content of the initial knowledge graph, calculate the semantic distance between the current user focus and the original content topic, and convert the numerical value into interpretable logical assertions.

[0140] The prediction module is used to predict optimized content that meets the user's current intent based on a deep learning model with dynamic anchor points as input, generate a preliminary neural ranking score, and present the results.

[0141] It should be noted that other specific implementations of the AI ​​response content visibility optimization system in this embodiment of the invention can refer to the above-described AI response content visibility optimization method.

[0142] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing content visibility in AI responses, characterized in that, The method includes: This research utilizes web crawling technology to collect multimodal source content of AI responses from multiple channels. After preprocessing, the themes of the multimodal source content are sorted. Domain-related sentiment words are filtered and added based on a general sentiment dictionary combined with domain-specific corpora of document themes to quantify the feature types of AI response content. The position of sentiment qualitative descriptive words in the text is located, and the text's theme keywords are extracted. Relevant theme content is retrieved from the multimodal source content using these theme keywords, and after filtering, the most matching content is added to the region based on the semantics and sentiment tendency of the qualitative description area. The text is converted into a vector representation, and the feature vectors of different modalities are fused using a concatenation method to form optimized content. Entity concepts and relationships are identified from the multimodal source content to construct an initial knowledge graph. Entity concepts in the optimized content are mapped to the graph to uncover potential content associations. Based on the semantics of user queries and entity relationships extracted, relevant information is queried in the knowledge graph, and the relevance between the content and the query is calculated. Real-time, multi-round user interaction flows are continuously monitored and analyzed, and key dynamic anchors are extracted from the interaction flows. The dynamic anchors are compared with the content of the initial knowledge graph to calculate the semantic distance between the current user focus and the original content theme, and the numerical value is converted into a solvable value. The logical assertion of the statement; the continuous monitoring and analysis of real-time, multi-round user interaction flows includes: collecting all user interaction data in different sessions, using deep behavioral analysis technology to deeply analyze user interaction data behavior patterns; using a temporal neural network model to encode the identified user behavior patterns, transforming behavioral features in different scenarios into a unified feature representation for fusion, and extracting feature vectors of users' core value orientations; constructing a unique value graph for each user based on the extracted user value features; performing value semantic analysis on the question-and-answer content of user interactions each time content visibility is optimized, matching the parsed value orientation with the value graph, and calculating the value fit; dynamically adjusting the semantic distance according to the value fit, and then updating the content visibility ranking in a floating manner; wherein, the collection of all user interaction data in different sessions also includes dividing users into different subgroups according to the relationships and interactions between users; comprehensively considering the value orientations of individuals within the subgroups, calculating the value orientation of the subgroups for each feature; and then integrating the value orientations of each subgroup to calculate the value orientation of the entire group for each feature. Based on a deep learning model, using dynamic anchor points as input, the system predicts optimized content that satisfies the user's current intent, generates a preliminary neural ranking score, and presents the results.

2. The method for optimizing content visibility in AI responses as described in claim 1, characterized in that, The process of identifying entity concepts and relationships from preprocessed multimodal source content to construct an initial knowledge graph includes: performing word segmentation on user queries, breaking sentences down into individual words and labeling the part of speech of each word; determining the grammatical relationships between the components of the sentence through syntactic analysis; extracting entity relationships from the semantic analysis results, using a convolutional neural network to input the semantically analyzed sentence features as model input, and outputting entity relationship categories; matching the extracted entities with nodes in the knowledge graph, and performing queries in the knowledge graph using a graph query language based on the extracted entity relationships.

3. The method for optimizing content visibility in AI responses as described in claim 2, characterized in that, The use of graph query language in the knowledge graph specifically involves: statistically analyzing the co-occurrence of entities in the query results and entities in the user query, calculating the proportion of co-occurring entities to the total number of entities in the query and the total number of entities in the query results; vectorizing the text information in the user query and query results, and calculating the cosine similarity between the two vectors; calculating the semantic similarity between entities in the query and entities in the query results using a word vector model; and combining the correlation based on entity co-occurrence and the correlation based on semantic similarity, using the following formula: .

4. The method for optimizing content visibility in AI responses as described in claim 1, characterized in that, The deep learning model, using dynamic anchors as input, includes the following steps: A dynamic anchor is the smallest semantic unit representing the user's current focus; the deep learning model, using dynamic anchors as input, quickly predicts which optimized content satisfies the user's current intent, generating a preliminary neural ranking score; based on the knowledge graph, business rules, and semantic distance, symbolic reasoning is performed to verify the reasonableness of the ranking results from the neural prediction channels; if the results from the two channels are consistent, the result is directly adopted and an explanation is generated; if the results from the two channels conflict, an arbitration mechanism is initiated to prioritize rule constraints and generate a counterfactual explanation.

5. The method for optimizing content visibility in AI responses as described in claim 1, characterized in that, The construction of a personalized value graph for each user includes: the value graph uses abstract value concepts as nodes, connecting each node through weighted edges, where the weight of the edge represents the strength of the association between value concepts, and changes in the edge reflect the evolution of value concepts; specifically: determining abstract value concepts as nodes in the value graph, assuming the node set is... , where k is the number of nodes; through statistical analysis of a small number of samples, the weights of the edges between each node are initialized. Construct the initial weight matrix When a user generates new interaction data, the value map is updated based on the feature vector V extracted from the new interaction data; let the update function be U, and the updated weight matrix be... .

6. The method for optimizing content visibility in AI responses as described in claim 1, characterized in that, The calculation of value fit includes: performing value semantic analysis on user question-and-answer content each time content visibility is optimized, and extracting the value tendency feature vector Q of the question-and-answer content through a deep learning model; matching the parsed value tendency feature vector Q with the node features in the value graph, and calculating the value fit S; assuming that the value graph contains nodes... The feature vector is The value fit S is calculated by comparing Q with each... The cosine similarity is obtained, that is Take the maximum value As the ultimate value alignment.

7. The method for optimizing content visibility in AI responses as described in claim 1, characterized in that, The method of collecting all user interaction data in different sessions includes: recording and preprocessing each user's interaction behavior data during the use of the AI ​​response system; extracting features that reflect individual value orientations from the preprocessed data; calculating the individual's value orientation weights for each feature based on the frequency and importance of the user's interaction on specific features using a variant of the TF-IDF algorithm; the method also includes: sorting and filtering the content in the AI ​​response system according to the fused multi-level value evaluation system to optimize content visibility; making preliminary modifications to the text to be optimized according to the established authoritative content introduction and fluency optimization strategies; and using a multimodal feature representation model to re-analyze the pre-optimized text to establish a comprehensive evaluation index system for authority and fluency.

8. The method for optimizing content visibility in AI response as described in claim 7, characterized in that, The formation of a comprehensive, multi-level value assessment system specifically includes: continuously collecting various interaction data of users in the AI ​​response system, extracting features that reflect the group's value orientation, arranging the extracted group value orientation features in chronological order to form time series data; training a long-term value evolution model using historical time series data, adjusting the model parameters to accurately fit historical data; using the trained long-term value evolution model to predict the future trend of the group's value orientation; defining indicators to measure the difference between individual and group value orientations, using these indicators to measure the difference between individual value orientations and the group's average value orientation; setting warning thresholds for each indicator, when the indicator value exceeds the threshold, indicating the existence of value conflict, and then combining the prediction results of the long-term value evolution model and the value conflict warning information to simulate different value integration and content recommendation strategies.

9. A system for optimizing content visibility in AI responses, applied to the method for optimizing content visibility in AI responses as described in any one of claims 1 to 8, characterized in that, The system includes: The data collection module is used to collect multimodal source content of AI responses from multiple channels using web crawling technology, and then sort out the themes of the multimodal source content after preprocessing. The analysis module is used to filter and add domain-related sentiment words based on a general sentiment dictionary combined with domain corpus of document topics, quantify the feature types of AI response content, locate the position of sentiment qualitative descriptive words in the text, and extract text topic keywords; The fusion module is used to retrieve relevant topic content from multimodal source content using topic keywords, and after filtering, selects the most matching content based on the semantics and sentiment of the qualitative description area and adds it to that area; it converts the text into vector representation and uses a concatenation method to fuse the feature vectors of different modalities to form optimized content; The mining module is used to identify entity concepts and relationships from preprocessed multimodal source content to construct an initial knowledge graph, map entity concepts in optimized content to the graph, and mine potential associations in the content; based on the semantics of user queries and extract entity relationships, relevant information is queried in the knowledge graph, and the relevance between the content and the query is calculated accordingly. The computation module is used to continuously monitor and parse real-time, multi-round user interaction streams, extract key dynamic anchors from the interaction streams, compare the dynamic anchors with the content of the initial knowledge graph, calculate the semantic distance between the current user focus and the original content topic, and convert the numerical value into interpretable logical assertions. The prediction module is used to predict optimized content that meets the user's current intent based on a deep learning model with dynamic anchor points as input, generate a preliminary neural ranking score, and present the results.

Images