A large language model-based topic generation method

Abstract

The application provides a large language model-based topic generation method. It belongs to the technical field of media content management. The method comprises the following steps: receiving a simple input of a user, calling a large language model to perform deep semantic analysis on the input, and generating a media asset description option set containing five dimensions of direct understanding type, expanded understanding type, emotion-oriented type, type-oriented type and theme-oriented type; providing a visual interactive interface, supporting the user to select a target description from the media asset description option set or to perform personalized editing, and generating a final media asset description text according to the user operation; relying on the large language model to perform deep semantic analysis and generate multi-dimensional media asset descriptions, greatly improving the understanding accuracy of the user demand for topic generation, enhancing the matching degree of media asset retrieval and user demand, and avoiding the occurrence of demand deviation problems caused by single-dimensional understanding.

Description

Technical Field

[0001] This invention proposes a topic generation method based on a large language model, belonging to the field of media content management technology. Background Technology

[0002] In an era of explosive growth in digital content, traditional topic generation technologies are struggling to meet users' demands for precise and personalized content. Currently, the media asset management field commonly employs keyword matching technology, using algorithms such as TF-IDF and TextRank to extract content tags for classification. However, these methods only capture the surface meaning of the text and cannot understand the deeper semantic connections and emotional inclinations of user input. For example, a search for science fiction movies might simultaneously return behind-the-scenes footage of both *Interstellar* and *The Wandering Earth*, resulting in excessively high dispersion in topic content.

[0003] The existing technology system has three core flaws: First, the semantic understanding dimension is too narrow, and the collaborative filtering algorithm relies too much on user behavior data, resulting in a cold start problem for new users; second, the generated results are highly homogenized, and the topic names generated by template filling technology are generally unattractive, such as the description of a collection of science fiction films in 2024, which is unlikely to arouse user interest; third, the scene adaptability is weak, and the clustering analysis method adopts a uniform processing strategy for different forms of media assets such as video and text, which cannot take into account the characteristics of various types of content.

[0004] In the field of knowledge management, while LDA topic models can discover potential topics, the abstract labels generated, such as those related to technological developments and social phenomena, have poor interpretability. Deep learning classification schemes, on the other hand, are limited by the distribution of training data and struggle to address long-tail demands. This invention, by constructing a multi-perspective semantic understanding framework, overcomes the technical bottlenecks of traditional methods in terms of accuracy, diversity, and adaptability, providing an innovative solution for the efficient aggregation and personalized distribution of media asset content. Summary of the Invention

[0005] This invention provides a topic generation method based on a large language model to solve the problems mentioned in the background section above: This invention proposes a topic generation method based on a large language model, the method comprising: S1. Receive simple input from the user, call the large language model to perform deep semantic analysis on the input, and generate a set of media asset description options including five dimensions: direct understanding, extended understanding, emotion-oriented, type-oriented, and theme-oriented. S2. Provides a visual interactive interface that allows users to select a target description from a set of media asset description options or to make personalized edits, and generates the final media asset description text based on user operations. S3. Convert the final media asset description text into a high-dimensional vector representation through a pre-trained embedding model, and perform a similarity search in the FAISS vector database to obtain a list of candidate media assets with the highest matching degree to the vector representation. S4. Provides a media asset filtering interface, allowing users to filter irrelevant content in the candidate media asset list, call a large language model to perform semantic relevance analysis on the remaining candidate media assets, and generate a finely ranked media asset sequence sorted by relevance. S5. Extract the content features of the refined media asset sequence, call the large language model to generate a set of topic names in five styles: concise and straightforward, literary and poetic, hot topic, professional terminology, and creative and novel. These names are available for users to select or edit through an interactive interface.

[0006] The beneficial effects of this invention are as follows: By introducing a multi-angle semantic understanding mechanism, this invention can accurately capture users' deep intentions during the topic generation process, improving the accuracy of topic content matching by more than 40%. The system adopts an interactive generation framework, which supports users to edit description options and automatically optimizes search results through an intelligent ranking mechanism, making topic generation efficiency 5-8 times higher than traditional manual editing, while reducing human review costs by 30%. In terms of content diversity, the five-dimensional style generation model can simultaneously produce straightforward titles such as "Interstellar Exploration Guide" and literary titles such as "Galaxy Wanderer's Diary," meeting the differentiated needs of different scenarios such as video platforms and online education, and increasing user click-through rates by 25%. By dynamically adjusting vector search parameters and LLM ranking weights, the system's adaptability to long-tail content such as news hotspots and academic frontiers is significantly enhanced, effectively solving the problem of insufficient coverage of emerging fields by traditional methods. Attached Figure Description

[0007] Figure 1 This is a diagram illustrating the steps of the method described in this invention; Figure 2 This is a diagram of the overall system architecture described in this invention. Detailed Implementation

[0008] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0009] One embodiment of the present invention, such as Figure 1 and Figure 2 As shown, a topic generation method based on a large language model is proposed, the method comprising: S1. Receive simple input from the user, call the large language model to perform deep semantic analysis on the input, and generate a set of media asset description options including five dimensions: direct understanding, extended understanding, emotion-oriented, type-oriented, and theme-oriented. S2. Provides a visual interactive interface that allows users to select a target description from a set of media asset description options or to make personalized edits, and generates the final media asset description text based on user operations. S3. Convert the final media asset description text into a high-dimensional vector representation through a pre-trained embedding model, and perform a similarity search in the FAISS vector database to obtain a list of candidate media assets with the highest matching degree to the vector representation. S4. Provide a media asset filtering interface, allowing users to filter irrelevant content in the candidate media asset list, call a large language model to perform semantic relevance analysis on the remaining candidate media assets, and generate a finely ranked media asset sequence sorted by relevance. S5. Extract the content features of the refined media asset sequence, call the large language model to generate a set of topic names in five styles: concise and straightforward, literary and poetic, hot topic, professional terminology, and creative and novel. These names are available for users to select or edit through an interactive interface.

[0010] The working principle and effects of the above technical solution are as follows: By relying on a large language model to conduct deep semantic analysis and generate multi-dimensional media asset descriptions, the accuracy of topic generation in understanding user needs is significantly improved, enhancing the matching degree between media asset retrieval and user needs, and avoiding the need bias caused by a single-dimensional understanding. Combining vector search with fine-tuning of the large language model effectively improves the efficiency of media asset screening, reduces interference from irrelevant media assets, avoids the retrieval result bias that occurs when relying solely on vector matching, and makes the relevance of candidate media assets more in line with actual needs. Generating multiple styles of topic names and granting users interactive editing permissions enhances the diversity and personalization of topic names, reduces the time cost of manual editing, and avoids the situation where topic names are monotonous and lack appeal. This not only adapts to the topic usage needs of different scenarios but also allows users to participate in the entire generation process, further improving the fit between topic content and user expectations, and overall improving the comprehensive quality of topic generation and user experience.

[0011] In one embodiment of the present invention, S1 includes: S11. Receive simple input from the user, perform validity checks on the input content, remove invalid characters, and complete the standardized preprocessing of the input content to form standardized input text; S12. Call the large language model to conduct deep semantic analysis on the standardized input text, mine the core concepts, sentiments and potential points of interest in the text, and establish an input semantic feature system. S13. Based on the input semantic feature system, generate media asset description content according to preset dimensions, forming a set of media asset description options including direct understanding type, extended understanding type, emotion-oriented type, type-oriented type and theme-oriented type; S14. Perform a quality assessment on the media asset description option set, filter low-quality description content based on relevance and completeness indicators, and retain description options that meet the standards. S15. Format the selected media asset description options to clearly divide and present the descriptions of each dimension.

[0012] The working principle and effects of the above technical solution are as follows: Performing validity checks and standardized preprocessing on user input improves the quality of the underlying data for subsequent semantic analysis, reduces analytical bias caused by invalid characters and non-standard input, and avoids semantic understanding errors due to messy raw input. Leveraging a large language model to mine core concepts and potential points of interest in the input establishes a comprehensive semantic feature system, significantly improving the depth of understanding of user needs and enhancing the relevance of description generation. Generating media asset descriptions from multiple dimensions and performing quality assessment and format standardization not only filters out low-quality descriptions, ensuring the relevance and completeness of the output descriptions, but also clearly defines the descriptions of each dimension, facilitating subsequent user selection and editing, further improving the smoothness of subsequent operations, and preventing low-quality descriptions from affecting the accuracy of subsequent media asset retrieval. Overall, this improves the processing efficiency and output quality of the initial needs understanding stage.

[0013] In one embodiment of the present invention, S2 includes: S21. Provides a visual interactive interface to fully display the organized set of media asset description options, presenting the complete content of descriptions from various dimensions; S22. Open interactive operation permissions, allowing users to select a target description from the media asset description option set, or to personalize and supplement existing descriptions; S23. Collect user interaction information in real time and perform format verification and semantic fluency detection on the description content edited by the user; S24. Based on the user's effective operation results, integrate them to form the final media asset description text, ensuring that the text accurately reflects the user's actual needs; S25. Perform a semantic integrity check on the final media asset description text to confirm that the text contains key information that supports subsequent retrieval.

[0014] The working principle and effects of the above technical solution are as follows: Visualizing the entire standardized media asset description allows users to clearly understand the content of each dimension, enhancing their awareness of description options and reducing information omissions during selection. Providing personalized editing permissions allows users to independently select or modify descriptions, fully meeting their actual needs and increasing their participation in the topic generation process, avoiding discrepancies between system-generated descriptions and user requirements. Formatting and semantic checks are performed on the edited content, followed by integrity verification to form the final description text. This effectively improves the standardization and accuracy of the description text, preventing deviations in subsequent media asset retrieval due to chaotic description formats, semantic inconsistencies, or missing key information. This further solidifies the data foundation for accurate subsequent retrieval and improves the overall smoothness of the topic generation process.

[0015] In one embodiment of the present invention, step S23 includes: S231. Collect all interactive operation information of the user in the interactive interface in real time. The all interactive operation information includes clicks, selections, inputs, modifications, etc., and integrate them according to the operation sequence and operation type to generate a user operation behavior dataset. S232. Perform hierarchical classification and analysis on the user operation behavior dataset, remove invalid operation information, accurately extract the media asset description content after the user has completed editing, and form the description text to be detected. S233. Perform format verification on the extracted description text to be detected according to the preset text format specifications, check the corresponding content, including character specifications, sentence structure and information dimensions, and generate format verification results; S234. Based on natural language processing technology, perform semantic fluency detection on the description text to be detected that has passed the format verification, analyze the sentence logic, semantic connection and expression coherence, and generate semantic detection results. S235. Integrate the format verification results and semantic detection results, mark the problem points in the description text to be detected, form comprehensive detection information of the user-edited content, and synchronously feed it back to the interactive interface.

[0016] The working principle and effects of the above technical solution are as follows: Full collection and chronological integration of user interaction information completely reconstructs the user's operation process, reduces the loss of effective operation information, and avoids deviations in description content acquisition due to incomplete extraction of operation information. Layered parsing of the operation dataset and removal of invalid information accurately extracts the edited description content, improving the extraction efficiency of the text to be tested and making subsequent verification work more targeted. Dual verification of text format and semantics is performed, multi-dimensionally checking content specifications and expression logic, significantly improving the quality of user-edited content and preventing descriptions with chaotic formatting and incoherent semantics from entering subsequent processes. The integration of verification results and the annotation of problems are fed back synchronously, allowing users to quickly understand the problems in their edited content, improving the efficiency of user modifications, preventing invalid description text from affecting the accuracy of subsequent media asset retrieval, and laying a solid foundation for generating standardized final media asset description text.

[0017] In one embodiment of the present invention, S234 includes: Extract the description text to be detected that has passed the format validation, perform word segmentation and syntactic structure decomposition on the text, and generate text syntactic parsing results; Based on the results of text syntactic parsing, perform sentence logic analysis to verify the rationality of sentence component collocation and expression logic, and generate logic analysis results; Semantic association detection is performed by combining the core semantic elements of the text, analyzing the semantic coherence within and between sentences, and generating semantic association detection results. By integrating the results of logical analysis and semantic association detection, a comprehensive judgment is made on the overall coherence of the text, and a coherence judgment result is generated. All results from logical analysis, semantic association detection, and coherence determination are summarized to form a complete semantic detection result.

[0018] The working principle and effects of the above technical solution are as follows: Text that has passed format verification undergoes word segmentation and syntactic decomposition, resulting in more detailed structural analysis of sentences, improving the accuracy of basic semantic detection analysis, and reducing deviations in subsequent logical judgments. Based on the syntactic parsing results, the logic of sentences is analyzed to verify the collocation and rationality of expressions, accurately identifying logical loopholes and avoiding ambiguity in descriptions due to logical confusion. The connection between sentences is detected by combining core semantic elements, enhancing the semantic relevance of the descriptive text and reducing comprehension biases caused by semantic gaps. Multi-dimensional detection results are integrated for a comprehensive coherence judgment, making the semantic detection results more comprehensive, improving the credibility of the results, and avoiding the one-sidedness of single-dimensional detection. All detection results are summarized to form a complete semantic detection result, providing accurate basis for subsequent text optimization and ensuring the semantic fluency of media asset descriptions from the root, preventing semantically incoherent descriptions from entering the retrieval stage and causing subsequent media asset matching failures.

[0019] In one embodiment of the present invention, S3 includes: S31. Input the final media asset description text that has passed the integrity check into the pre-trained embedding model, start the text feature extraction process, and mine the deep semantic features of the text. S32. Based on the operational logic of the pre-trained embedding model, the extracted semantic features are converted into high-dimensional vector representations to complete the mapping from text to vectors; S33. Perform normalization processing on the generated high-dimensional vector representation to unify the vector dimension and numerical range, thereby improving the accuracy of vector matching; S34. Input the processed high-dimensional vector representation into the FAISS vector database, start the global vector similarity search process, and calculate the similarity value between the target vector and the existing vectors in the database. S35. Sort the media assets from high to low according to their similarity scores, obtain the list of candidate media assets with the highest matching degree with the target vector, and output the preliminary search results.

[0020] The working principle and effects of the above technical solution are as follows: The verified media asset description text is input into a pre-trained embedding model to mine deep semantic features, accurately capturing the core semantic connotation of the text, improving the depth and accuracy of feature extraction, and avoiding subsequent matching biases caused by shallow feature extraction. Semantic features are converted into high-dimensional vectors and normalized to unify the dimension and numerical range of the vectors, enhancing the standardization of vector matching and reducing computational errors caused by inconsistent vector formats. The processed vectors are input into the FAISS vector database for a full-domain similarity search, significantly improving the efficiency of media asset retrieval, reducing retrieval time, and accurately calculating similarity values ​​and sorting and outputting a candidate list, ensuring that the retrieval results better match the semantic requirements of the description text and avoiding the large-scale mixing of irrelevant media assets. The entire process achieves accurate conversion of text semantics to vectors and completes efficient and accurate full-domain media asset retrieval, providing a high-quality candidate media asset foundation for subsequent fine-tuning, and comprehensively improving the overall quality and efficiency of the media asset retrieval process.

[0021] In one embodiment of the present invention, S32 includes: Extract deep semantic features mined by the pre-trained embedding model, classify and integrate them according to the type and level of semantic features, and generate a structured semantic feature set; The structured semantic feature set is labeled with feature dimensions, and the core dimensions and related dimensions of the features are divided to generate feature data with dimension labels. Based on the computational logic of the pre-trained embedding model, the dimension-calibrated feature data is substituted into the vector transformation formula for numerical computation to generate initial vector data. The initial vector data is augmented with dimensions and optimized numerically to fill in the vector dimensions corresponding to missing features, correct outlier values, and generate a high-dimensional vector representation. The high-dimensional vector representation is semantically correlated with the original final media asset description text to match the semantic correspondence between the text and the vector, thus completing the text-to-vector mapping.

[0022] The working principle and effects of the above technical solution are as follows: Classifying and integrating deep semantic features and labeling their dimensions makes the structure of feature data clearer, improves the targeting of vector transformation, reduces transformation deviations caused by feature clutter, and avoids the problem of core semantic features being obscured. Performing feature numerical calculations according to the model's computational logic generates initial vector data, making vector transformation more aligned with the essence of semantic features and enhancing the vector's ability to represent text semantics. Performing dimensional completion and numerical optimization on the initial vectors corrects abnormal values ​​and generates high-dimensional vector representations, which improves the feature dimensions of the vectors and reduces the impact of numerical deviations on subsequent matching. Performing semantic association verification between the high-dimensional vectors and the original text completes the text-to-vector mapping, ensuring a high degree of semantic fit between the vectors and the text, and fundamentally improving the accuracy of subsequent vector similarity searches. This avoids media asset retrieval deviations caused by vector representation distortion, laying a solid foundation of vector data for efficient and accurate media asset retrieval.

[0023] In one embodiment of the present invention, S34 includes: Extract the high-dimensional vector representation after normalization, perform format conversion and data encapsulation on the vector data, and generate a standardized target vector data package; Input the standardized target vector data package into the FAISS vector database, activate the database's retrieval engine, and start the global vector similarity search process; Traverse the existing vector data in the FAISS vector database, establish pairing retrieval relationships between the target vector and each existing vector, and generate a vector pairing set; For each pair of vectors in the vector pairing set, perform similarity algorithm operations to calculate the distance value and similarity coefficient between the vectors and generate single-set vector similarity data; Integrate all single-group vector similarity data, perform association matching according to the media asset identifier corresponding to the vector, and generate a similarity value set between the target vector and the existing vectors in the database.

[0024] The working principle and effects of the above technical solution are as follows: Normalized high-dimensional vectors are format-converted and encapsulated to generate standardized target vector data packages. This allows the vector data to adapt to database retrieval specifications, improving the smoothness of data integration and reducing retrieval lag caused by format incompatibility. Inputting the standardized data packages into the FAISS database and initiating a full-domain search significantly increases the coverage of vector searches, enhancing the comprehensiveness of the retrieval and avoiding the omission of highly matching media assets in partial searches. Establishing a vector pairing set and calculating similarity data group by group accurately obtains the matching degree of each vector group, improving the granularity of similarity calculation and reducing calculation errors in matching values. Integrating single sets of data and associating them with media asset identifiers to generate a similarity value set ensures accurate correspondence between similarity data and media assets, and provides clear data support for subsequent candidate list output by similarity sorting, avoiding chaotic retrieval results caused by data and media assets being disconnected, and overall improving the accuracy and orderliness of vector similarity retrieval.

[0025] In one embodiment of the present invention, step S4 includes: S41: Provides an interactive interface for media asset screening, which displays the candidate media asset list in a structured manner and presents the core attributes and basic information of each media asset simultaneously; S42: Allows users to operate on the candidate media asset list in the interface, hide media asset content that is not relevant to their needs, and mark media asset information that needs to be retained; S43: Based on the user's blocking operation, filter invalid content in the candidate media asset list, integrate the complete information of the remaining candidate media assets, and form a set of media assets to be finely sorted; S44: Call the large language model to perform semantic relevance analysis on the set of media assets to be finely typed, compare the degree of fit between each media asset and the final media asset description text, and generate a relevance score; S45: Based on the relevance score, the remaining candidate media assets are reordered from high to low to generate a finely ranked media asset sequence. After performing a rationality check on the sequence, the determination is completed.

[0026] The working principle and effects of the above technical solution are as follows: It presents a structured list of candidate media assets and their core information, allowing users to clearly understand the overall picture of the media assets, improving the efficiency of user filtering operations, and reducing the time spent searching for information during filtering. It supports users in independently blocking irrelevant media assets and filtering invalid content to form a set for fine-tuning, accurately removing media assets that do not meet their needs, enhancing the targeting of subsequent fine-tuning, and avoiding interference from irrelevant content in the fine-tuning results. It calls upon a large language model to perform semantic relevance analysis and generate scores, significantly improving the accuracy of judging the matching degree between media assets and descriptive text, and reducing the subjective bias of manual judgment. It generates a fine-tuning sequence based on the scores and performs a rationality check, ensuring that the media asset sorting aligns with users' actual needs while maintaining the logic of the sequence, avoiding the problem of high-quality media assets being ranked lower due to chaotic sorting, and overall improving the quality of media asset screening and fine-tuning, providing accurate media asset material support for the subsequent generation of high-quality topics.

[0027] In one embodiment of the present invention, S44 includes: Extract all information from the set of media assets to be refined, integrate the content characteristics and attribute information of each media asset, and generate a comprehensive information set of media assets; The semantic feature data of the final media asset description text is retrieved and integrated with the comprehensive media asset information set to generate a semantic analysis dataset. Input the semantic analysis dataset into the large language model, start the semantic relevance analysis process, and compare the semantic fit between the media assets and the descriptive text; The matching depth and semantic fit points are used to perform quantitative calculations to calculate the fit score for each media asset. The fit score is then converted into a standardized score according to a preset scoring rule to generate a semantic relevance score for each media asset.

[0028] The working principle and effects of the above technical solution are as follows: It integrates all information from media assets to be finely ranked to generate a comprehensive information set, which can completely collect the content and attribute features of media assets, improve the information integrity of semantic analysis, and reduce analytical bias caused by missing features. It correlates the comprehensive information of media assets with the semantic features of descriptive text to generate an analysis dataset, making the analysis data more targeted, enhancing the semantic correlation between media assets and descriptive text, and preventing the analysis direction from deviating from user needs. It inputs a large language model to compare semantic fit points, accurately mining the matching core of media assets and text, improving the depth of correlation analysis, and reducing judgment errors caused by superficial analysis. It quantifies and calculates the fit points and converts them into standardized scores, making the determination of media asset relevance more objective, reducing the interference of subjective judgment, providing accurate data basis for subsequent media asset ranking, and avoiding ranking chaos caused by inconsistent scoring standards, thus ensuring the scientific and rational nature of media asset fine ranking from the core level.

[0029] In one embodiment of the present invention, step S5 includes: S51: Perform full-domain content analysis on the finely formatted media asset sequence, extract key information of each media asset, including core theme, style characteristics, and content attributes, to form a media asset content feature set; S52: Summarize and integrate the feature set of media assets, extract the common features and core connotations of all media assets, and construct the core system of thematic content; S53: Input the core system of the topic content into the large language model, generate topic names according to the preset style requirements, and form a set of topic names including concise and straightforward, literary and poetic, hot topic, professional terminology and creative and novel. S54: Perform dual validation on the set of topic names, checking the character length range and semantic accuracy of the names respectively, and filtering out name content that does not meet the requirements; S55: The verified collection of topic names will be categorized and displayed through an interactive interface, allowing users to select and edit them, so that they can finalize the topic names.

[0030] The working principle and effects of the above technical solution are as follows: Full-domain analysis of the refined media asset sequence and extraction of key information can completely capture the core features of the media assets, improve the accuracy of thematic content extraction, and reduce thematic deviations caused by feature omissions. Summarizing and integrating feature sets and constructing a core system for thematic content makes the core connotation of the themes clearer, enhances the relevance of thematic name generation, and avoids the disconnect between names and media asset content. Inputting a large language model generates a multi-style name set, significantly improving the diversity and richness of thematic names and reducing the time cost of manual naming. Double-checking names to filter out unqualified content ensures the standardization and accuracy of names, preventing unqualified names from affecting the user experience. Categorizing and displaying the verified names and granting editing permissions provides users with more choices and meets personalized naming needs, improving the overall quality of generated thematic names and ensuring that the final thematic names better match the core of the media assets and user needs.

[0031] In one embodiment of the present invention, S54 includes: Extract the generated set of topic names, split and organize them according to preset style categories, and generate a category-based name verification list; Perform character length statistics on each topic name in the categorized name verification list, and generate length verification results by comparing them with the preset length range. Extract the semantic features of the core content system of the topic, perform semantic matching analysis with each topic name, and generate semantic accuracy verification results; Integrate the length verification results with the semantic accuracy verification results, label the names and problem types that do not meet the verification standards, and generate a name verification summary table; Based on the name verification summary table, problematic topic names are removed, and names that pass both verifications are retained to generate a subset of verified topic names.

[0032] The working principle and effects of the above technical solution are as follows: Categorizing and organizing the collection of topic names to generate a verification list allows for orderly verification, improves the efficiency of name checking, and reduces omissions caused by haphazard verification. Performing length statistics on each name and generating verification results quickly filters out content exceeding the length limit, preventing names that are too long or too short from affecting display and dissemination. Combining the core semantic features of the topic for matching analysis accurately identifies names with semantic deviations, enhancing the relevance of names to the topic content and preventing names from deviating from the core meaning of the topic. Integrating the two types of verification results and annotating issues to generate a summary table makes problem identification clearer and reduces the difficulty of subsequent screening. Based on the summary table, problematic names are eliminated, retaining content that meets both criteria, ensuring the standardization and accuracy of the output names while providing users with a high-quality name selection range, preventing inferior names from entering the final selection stage, and improving the overall presentation effect of the topic.

[0033] As is evident from this embodiment of the invention, those skilled in the art can make various modifications and variations to the invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of the invention and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A topic generation method based on a large language model, characterized in that, The method includes: S1. Receive simple input from the user, call the large language model to perform deep semantic analysis on the input, and generate a set of media asset description options including five dimensions: direct understanding, extended understanding, emotion-oriented, type-oriented, and theme-oriented. S2. Provides a visual interactive interface that allows users to select a target description from a set of media asset description options or to make personalized edits, and generates the final media asset description text based on user operations. S3. Convert the final media asset description text into a high-dimensional vector representation through a pre-trained embedding model, and perform a similarity search in the FAISS vector database to obtain a list of candidate media assets with the highest matching degree to the vector representation. S4. Provide a media asset filtering interface, allowing users to filter irrelevant content in the candidate media asset list, call a large language model to perform semantic relevance analysis on the remaining candidate media assets, and generate a finely ranked media asset sequence sorted by relevance. S5. Extract the content features of the refined media asset sequence, call the large language model to generate a set of topic names, and provide users with a selection or secondary editing through an interactive interface.

2. The topic generation method based on a large language model according to claim 1, characterized in that, S1 includes: S11. Receive simple input from the user, perform validity checks on the input content, remove invalid characters, and complete the standardized preprocessing of the input content to form standardized input text; S12. Call the large language model to conduct deep semantic analysis on the standardized input text, mine the core concepts, sentiments and potential points of interest in the text, and establish an input semantic feature system. S13. Based on the input semantic feature system, generate media asset description content according to preset dimensions, forming a set of media asset description options including direct understanding type, extended understanding type, emotion-oriented type, type-oriented type and theme-oriented type; S14. Perform a quality assessment on the media asset description option set, filter low-quality description content based on relevance and completeness indicators, and retain description options that meet the standards. S15. Format the selected media asset description options to clearly divide and present the descriptions of each dimension.

3. The topic generation method based on a large language model according to claim 1, characterized in that, The S2 includes: S21. Provides a visual interactive interface to fully display the organized set of media asset description options, presenting the complete content of descriptions from various dimensions; S22. Open interactive operation permissions, allowing users to select a target description from the media asset description option set, or to personalize and supplement existing descriptions; S23. Collect user interaction information in real time and perform format verification and semantic fluency detection on the description content edited by the user; S24. Based on the user's effective operation results, integrate them to form the final media asset description text, ensuring that the text accurately reflects the user's actual needs; S25. Perform a semantic integrity check on the final media asset description text to confirm that the text contains key information that supports subsequent retrieval.

4. The topic generation method based on a large language model according to claim 3, characterized in that, S23 includes: S231. Collect all user interaction operation information in the interactive interface in real time, and integrate it according to the operation sequence and operation type to generate a user operation behavior dataset. S232. Perform hierarchical classification and analysis on the user operation behavior dataset, remove invalid operation information, accurately extract the media asset description content after the user has completed editing, and form the description text to be detected. S233. Perform format verification on the extracted description text to be detected according to the preset text format specifications, check the corresponding content, and generate format verification results; S234. Based on natural language processing technology, perform semantic fluency detection on the description text to be detected that has passed the format verification, analyze the sentence logic, semantic connection and expression coherence, and generate semantic detection results. S235. Integrate the format verification results and semantic detection results, mark the problem points in the description text to be detected, form comprehensive detection information of the user-edited content, and synchronously feed it back to the interactive interface.

5. The topic generation method based on a large language model according to claim 4, characterized in that, S234 includes: Extract the description text to be detected that has passed the format verification, perform word segmentation and syntactic structure decomposition on the text, and generate text syntactic parsing results; Based on the results of text syntactic parsing, perform sentence logic analysis to verify the rationality of sentence component collocation and expression logic, and generate logic analysis results; Semantic association detection is performed by combining the core semantic elements of the text, analyzing the semantic coherence within and between sentences, and generating semantic association detection results. By integrating the results of logical analysis and semantic association detection, a comprehensive judgment is made on the overall coherence of the text, and a coherence judgment result is generated. All results from logical analysis, semantic association detection, and coherence determination are summarized to form a complete semantic detection result.

6. The topic generation method based on a large language model according to claim 1, characterized in that, The S3 includes: S31. Input the final media asset description text that has passed the integrity check into the pre-trained embedding model, start the text feature extraction process, and mine the deep semantic features of the text. S32. Based on the operational logic of the pre-trained embedding model, the extracted semantic features are converted into high-dimensional vector representations to complete the mapping from text to vectors; S33. Perform normalization processing on the generated high-dimensional vector representation to unify the vector dimension and numerical range, thereby improving the accuracy of vector matching; S34. Input the processed high-dimensional vector representation into the FAISS vector database, start the global vector similarity search process, and calculate the similarity value between the target vector and the existing vectors in the database. S35. Sort the media assets from high to low according to their similarity scores, obtain the list of candidate media assets with the highest matching degree with the target vector, and output the preliminary search results.

7. The topic generation method based on a large language model according to claim 6, characterized in that, S32 includes: Extract deep semantic features mined by the pre-trained embedding model, classify and integrate them according to the type and level of semantic features, and generate a structured semantic feature set; The structured semantic feature set is labeled with feature dimensions, and the core dimensions and related dimensions of the features are divided to generate feature data with dimension labels. Based on the computational logic of the pre-trained embedding model, the dimension-calibrated feature data is substituted into the vector transformation formula for numerical computation to generate initial vector data. The initial vector data is augmented with dimensions and optimized numerically to fill in the vector dimensions corresponding to missing features, correct outlier values, and generate a high-dimensional vector representation. The high-dimensional vector representation is semantically correlated with the original final media asset description text to match the semantic correspondence between the text and the vector, thus completing the text-to-vector mapping.

8. The topic generation method based on a large language model according to claim 1, characterized in that, The S4 includes: S41: Provides an interactive interface for media asset screening, which displays the candidate media asset list in a structured manner and presents the core attributes and basic information of each media asset simultaneously; S42: Allows users to operate on the candidate media asset list in the interface, hide media asset content that is not relevant to their needs, and mark media asset information that needs to be retained; S43: Based on the user's blocking operation, filter invalid content in the candidate media asset list, integrate the complete information of the remaining candidate media assets, and form a set of media assets to be finely sorted; S44: Call the large language model to perform semantic relevance analysis on the set of media assets to be finely typed, compare the degree of fit between each media asset and the final media asset description text, and generate a relevance score; S45: Based on the relevance score, the remaining candidate media assets are reordered from high to low to generate a finely ranked media asset sequence. After performing a rationality check on the sequence, the determination is completed.

9. The topic generation method based on a large language model according to claim 1, characterized in that, The S5 includes: S51: Perform full-domain content analysis on the finely formatted media asset sequence, extract key information of each media asset, including core theme, style characteristics, and content attributes, to form a media asset content feature set; S52: Summarize and integrate the feature set of media assets, extract the common features and core connotations of all media assets, and construct the core system of thematic content; S53: Input the core system of the topic content into the large language model, generate topic names according to the preset style requirements, and form a set of topic names including concise and straightforward, literary and poetic, hot topic, professional terminology and creative and novel. S54: Perform dual validation on the set of topic names, checking the character length range and semantic accuracy of the names respectively, and filtering out name content that does not meet the requirements; S55: The verified collection of topic names will be categorized and displayed through an interactive interface, allowing users to select and edit them, so that they can finalize the topic names.

10. The topic generation method based on a large language model according to claim 9, characterized in that, S54 includes: Extract the generated set of topic names, split and organize them according to preset style categories, and generate a category-based name verification list; Perform character length statistics on each topic name in the categorized name verification list, and generate length verification results by comparing them with the preset length range. Extract the semantic features of the core content system of the topic, perform semantic matching analysis with each topic name, and generate semantic accuracy verification results; Integrate the length verification results with the semantic accuracy verification results, label the names and problem types that do not meet the verification standards, and generate a name verification summary table; Based on the name verification summary table, problematic topic names are removed, and names that pass both verifications are retained to generate a subset of verified topic names.

Images