UGC retrieval method and system based on semantic information, electronic equipment and storage medium

Abstract

The invention provides a UGC retrieval method and system based on semantic information, electronic equipment and a storage medium, and relates to the field of search recommendation. The method comprises the steps of obtaining UGC information data and user search terms, and converting the UGC information data into UGC semantic feature data; performing multi-path recall retrieval according to the UGC semantic features and the user search terms to obtain a UGC candidate set related to the user search terms; carrying out reciprocal sorting fusion on a plurality of recall result candidate sets in the UGC candidate set to obtain a sorted recall result candidate set; the sorted recall result candidate set is rearranged through a BGE erank model, a rearrangement recall result candidate set and a rearrangement score of each UGC candidate item in the rearrangement recall result candidate set are obtained, and the rearrangement recall result candidate set and the rearrangement score of each UGC candidate item in the rearrangement recall result candidate set serve as a UGC retrieval result. Through the method, the correlation and the accuracy of retrieval results are improved.

Description

Technical Field

[0001] The present application relates to the field of search recommendation, and in particular to a UGC retrieval method, system, electronic device and storage medium based on semantic information. Background Art

[0002] With the rapid development of mobile Internet and social media, user-generated content (UGC) has shown an explosive growth trend. Massive UGC provides users with rich and diverse information resources, but also brings huge challenges to information retrieval and acquisition. Existing UGC retrieval technology is mainly based on keyword matching and text similarity calculation, which is difficult to fully explore and utilize the semantic information contained in UGC, resulting in the relevance and accuracy of retrieval results being difficult to meet the actual needs of users. Summary of the invention

[0003] The present application provides a UGC retrieval method, system, electronic device and storage medium based on semantic information, so as to improve the relevance and accuracy of retrieval results.

[0004] In a first aspect, the present application provides a UGC retrieval method based on semantic information, the method comprising: Obtain UGC information data and user search terms, and convert the UGC information data into UGC semantic feature data; Perform multi-way recall retrieval based on UGC semantic features and user search terms to obtain a UGC candidate set related to the user search terms; Multiple recall result candidate sets in the UGC candidate set are sorted and merged in reverse order to obtain a sorted recall result candidate set; The sorted recall result candidate set is reranked through the BGE rerank model to obtain the rearranged recall result candidate set and the reranking score of each UGC candidate item in the rearranged recall result candidate set, and the rearranged recall result candidate set and the reranking score of each UGC candidate item in the rearranged recall result candidate set are used as the UGC retrieval result.

[0005] In the above technical solution, by obtaining UGC information data and user search terms and converting the UGC information data into UGC semantic feature data, a solid semantic foundation is laid for subsequent recall retrieval and ranking re-arrangement. Using UGC semantic feature data, this method can deeply mine the semantic relevance of UGC, overcome the limitations of traditional keyword matching, and effectively improve the accuracy and coverage of retrieval.

[0006] In the multi-way recall retrieval stage, the method comprehensively considers the semantic features of UGC and user search terms, adopts multiple recall strategies for parallel retrieval, and generates multiple complementary recall result candidate sets. This fusion of multi-way recall fully utilizes the advantages of different recall algorithms, improves the recall rate and diversity of retrieval, and provides users with more comprehensive and personalized retrieval results.

[0007] In the inverse sorting fusion stage, the method performs inverse fusion of the ranking positions of UGC candidates in multiple recall result candidate sets to obtain a comprehensive sorted candidate set. This innovative inverse sorting fusion strategy cleverly balances the ranking weights of different recall paths, effectively eliminates sorting bias, makes the comprehensive sorting results more reasonable and accurate, and significantly improves the sorting quality and user experience.

[0008] In the BGE rerank stage, this method reranks the sorted recall result candidate set through the pre-trained BGE model, and outputs the reranked candidate set and the reranking score of each UGC candidate. The BGE model is a dynamic optimization sorting model based on business rules and user feedback. By introducing quantitative business reranking features and combining real-time user interaction data, the sorting results are dynamically adjusted and optimized to make them more in line with business needs and user preferences. This reranking mechanism that integrates business rules and user feedback has greatly improved the practicality and user satisfaction of UGC retrieval.

[0009] In a second aspect of the present application, a UGC retrieval system based on semantic information is provided, the system comprising: A data processing module is used to obtain UGC information data and user search terms, and convert the UGC information data into UGC semantic feature data; The recall module is used to perform multi-way recall retrieval based on UGC semantic features and user search terms to obtain a UGC candidate set related to the user search terms; A sorting module is used to perform inverse sorting and fusion of multiple recall result candidate sets in the UGC candidate set to obtain a sorted recall result candidate set; The rearrangement module is used to rearrange the sorted recall result candidate set through the BGE rerank model to obtain the rearranged recall result candidate set and the rearrangement score of each UGC candidate item in the rearranged recall result candidate set, and use the rearranged recall result candidate set and the rearrangement score of each UGC candidate item in the rearranged recall result candidate set as the UGC retrieval result.

[0010] In a third aspect of the present application, a computer storage medium is provided, wherein the computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the above method steps.

[0011] In the fourth aspect of the present application, an electronic device is provided, including a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the above method.

[0012] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. This application obtains UGC information data and user search terms, and converts the UGC information data into UGC semantic feature data, laying a solid semantic foundation for subsequent recall retrieval and ranking re-arrangement. Using UGC semantic feature data, this method can deeply mine the semantic relevance of UGC, overcome the limitations of traditional keyword matching, and effectively improve the accuracy and coverage of retrieval.

[0013] 2. In the multi-way recall retrieval stage, this method comprehensively considers UGC semantic features and user search terms, adopts multiple recall strategies for parallel retrieval, and generates multiple complementary recall result candidate sets. This fusion of multi-way recall fully utilizes the advantages of different recall algorithms, improves the recall rate and diversity of retrieval, and provides users with more comprehensive and personalized retrieval results.

[0014] 3. In the inverse sorting fusion stage of this application, the method performs inverse fusion of the ranking positions of the UGC candidates in multiple recall result candidate sets to obtain a candidate set after comprehensive sorting. This innovative inverse sorting fusion strategy cleverly balances the sorting weights of different recall paths, effectively eliminates sorting bias, makes the comprehensive sorting results more reasonable and accurate, and significantly improves the sorting quality and user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a UGC retrieval method based on semantic information provided in an embodiment of the present application; Figure 2 An architectural diagram of a UGC retrieval system based on semantic information provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of an electronic device provided in this application. DETAILED DESCRIPTION

[0016] In order to enable technicians in this field to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0017] In the description of the embodiments of the present application, words such as "for example" or "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "for example" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "for example" or "for example" is intended to present related concepts in a specific way.

[0018] In the description of the embodiments of the present application, the meaning of the term "multiple" refers to two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "include", "comprise", "have" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0019] In order to facilitate understanding of the method and system provided by the embodiments of the present application, before introducing the embodiments of the present application, the background of the embodiments of the present application is first introduced.

[0020] With the rapid rise of mobile Internet and social media, user-generated content (UGC) has experienced an unprecedented growth spurt. This massive amount of UGC information is like a treasure trove of rich resources, providing users with diversified and personalized information services. However, with the rapid expansion of the amount of UGC, how to accurately and efficiently retrieve and obtain the required information from the vast amount of content has become a technical problem that needs to be solved urgently.

[0021] Looking at the current mainstream UGC retrieval technologies, most of them are still based on traditional methods such as keyword matching and text similarity calculation. These methods have achieved the retrieval and sorting of UGC to a certain extent, but it is difficult to deeply explore the rich semantic information and internal connections contained in UGC. As content spontaneously created and shared by users, UGC often presents characteristics such as diverse vocabulary, complex semantics, and a wide range of topics. Relying solely on literal matching and similarity comparison cannot accurately grasp the user's true intentions and semantic preferences, resulting in a significant deviation between the retrieval results and the user's actual needs, which seriously affects the user's retrieval experience and satisfaction.

[0022] After the background introduction of the above content, those skilled in the art can understand the problems existing in the prior art. The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments.

[0023] Based on the above background technology, please refer to Figure 1 , Figure 1 A flowchart of a UGC retrieval method based on semantic information provided in an embodiment of the present application. The system can be implemented by a computer program or can be run as an independent tool application. Specifically, in an embodiment of the present application, the method can be applied on a server, but can also be applied on electronic devices such as a server. A UGC retrieval method based on semantic information includes the following steps: S101, obtaining UGC information data and user search terms, and converting the UGC information data into UGC semantic feature data; Specifically, in one embodiment of the present application, UGC information data and user search terms are first obtained. Among them, UGC information data can be collected from major UGC platforms and social media websites through crawler technology, or it can be obtained through data cooperation with UGC platforms. The acquired UGC information data usually includes multi-dimensional information such as UGC text content, author information, release time, interactive data, etc. At the same time, user search terms come from the query keywords entered by the user in the retrieval system. Obtaining these two types of data is the basis and prerequisite for subsequent UGC retrieval.

[0024] In order to better mine and utilize the semantic information contained in UGC information data, this embodiment converts the acquired UGC information data into UGC semantic feature data. Specifically, the UGC text content is first preprocessed, such as word segmentation, stop word removal, part-of-speech tagging, etc., and then the text is converted into a low-dimensional dense semantic vector representation using a pre-trained semantic model (such as Word2Vec, BERT, etc.). These semantic vectors can effectively capture the semantic features and relevance of UGC texts, and provide richer and more accurate semantic information for subsequent retrieval and sorting.

[0025] By converting UGC information data into semantic feature data, this embodiment overcomes the limitations of traditional keyword matching and achieves a deep understanding and representation of UGC semantics. Compared with the original text information, semantic feature data can more accurately depict the semantic connotation of UGC, eliminate the semantic gap caused by literal differences, and lay a solid semantic foundation for subsequent recall retrieval and ranking optimization.

[0026] For example, for a UGC text "What to do if your skin is dry and dehydrated, I urgently need a hydrating product", traditional keyword matching methods may find it difficult to accurately understand its semantic intent. However, by converting it into a semantic feature vector, it can be found that it is highly correlated with topics such as "hydrating", "moisturizing", and "skin care", thereby providing users with more accurate and comprehensive recall results for searching "hydrating skin care products".

[0027] Based on the above embodiment, as an optional embodiment, obtaining UGC information data includes: S201, extracting UGC basic information data from a relational database; Specifically, the first step to obtain UGC information data is to extract UGC basic information data from a relational database. Specifically, this embodiment first establishes a structured relational database for storing and managing UGC data from various UGC platforms and social media websites. This relational database usually adopts mainstream database management systems such as MySQL and Oracle, and designs a reasonable data table structure and fields to organize and store UGC data in a standardized and structured manner. UGC basic information data usually includes structured metadata fields such as UGC text content, author information, release time, number of likes, number of comments, etc. These data fields correspond to different data tables and fields in the relational database, respectively, and efficient data storage and query are achieved through foreign key association and index optimization.

[0028] The main purpose of extracting UGC basic information data from a relational database is to use the structured storage and efficient query capabilities of the relational database to quickly obtain the required UGC metadata information. Through SQL query statements, it is easy to filter and extract the required UGC data subset according to different conditions (such as time range, topic tags, authors, etc.), and use it as input for subsequent processing and analysis.

[0029] Specifically, extracting UGC basic information data from a relational database usually requires the following steps: first, write the corresponding SQL query statement based on the specific data requirements and screening conditions; then, use the database connection tool or API (such as JDBC, ODBC, etc.) to establish a connection with the relational database; then, execute the written SQL statement to query and extract the UGC basic information data that meets the conditions from the database; finally, store and transmit the extracted data in a structured format (such as DataFrame, JSON, etc.) for subsequent processing and analysis.

[0030] S202, extracting user and UGC interaction behavior data from the big data platform; Specifically, the second step of obtaining UGC information data is to extract user and UGC interaction behavior data from the big data platform. Specifically, this embodiment builds a big data processing platform based on a distributed computing framework (such as Hadoop, Spark, etc.) for storing and analyzing massive user behavior log data. These log data are usually stored in a distributed file system (such as HDFS) or a NoSQL database (such as HBase, Cassandra, etc.) in an unstructured or semi-structured form (such as text, JSON, etc.).

[0031] The interactive behavior data between users and UGC includes records of various interactive operations such as browsing, liking, collecting, commenting, and forwarding of UGC. These interactive behavior data contain rich user interest preferences and social network information, which is crucial for UGC retrieval and recommendation. However, due to its huge data volume, diverse formats, and rapid growth, traditional relational databases are difficult to effectively store and process these interactive behavior data. Therefore, it is necessary to use the distributed computing and storage capabilities of the big data platform to achieve efficient extraction and analysis of massive interactive behavior data.

[0032] The main purpose of extracting user and UGC interaction behavior data from the big data platform is to mine the valuable information contained in the user behavior data and provide data support for UGC retrieval and recommendation. By analyzing the user's historical interaction behavior, we can understand important characteristics such as user interests, social relationships, and influence, thereby achieving personalized UGC recommendations and socialized UGC retrieval. In addition, interaction behavior data can also be used to train and optimize various machine learning models (such as collaborative filtering, deep learning, etc.) to improve the accuracy of UGC retrieval and recommendation and user experience.

[0033] Specifically, extracting user and UGC interaction behavior data from a big data platform usually requires the following steps: first, using the data access and preprocessing components of the big data platform (such as Flume, Kafka, Hive, etc.) to import the original user interaction behavior log data into a distributed storage system; then, using big data query and analysis tools (such as Hive, Spark SQL, Pig, etc.), write the corresponding data extraction and conversion logic to extract structured interaction behavior data from the original log data; then, perform preprocessing operations such as cleaning, deduplication, and normalization on the extracted interaction behavior data to eliminate data noise and inconsistency; finally, store the processed interaction behavior data in an appropriate format (such as Parquet, ORC, etc.) in a distributed file system or data warehouse for subsequent analysis and application.

[0034] The advantage of extracting user and UGC interaction behavior data from the big data platform is that it fully utilizes the distributed computing and storage capabilities of the big data platform to achieve efficient processing and analysis of massive unstructured data. Through distributed data extraction and conversion, structured interaction behavior data can be quickly extracted from the original log data, and necessary data cleaning and preprocessing can be performed. At the same time, the big data platform provides a wealth of data analysis and mining tools, such as machine learning algorithm libraries, graph computing engines, etc., making the in-depth analysis and application of interaction behavior data more convenient and efficient.

[0035] S203: Associating and matching the UGC basic information data with the UGC interactive behavior data to form UGC information data.

[0036] Specifically, the last step of obtaining UGC information data is to associate and match the UGC basic information data with the UGC interactive behavior data to form complete UGC information data. Specifically, this embodiment uses the unique identifier of UGC (such as ugc_id) as the association key to associate and integrate the UGC basic information data extracted from the relational database and the UGC interactive behavior data extracted from the big data platform. The main purpose of associating and matching UGC basic information data and interactive behavior data is to comprehensively and three-dimensionally characterize the characteristics and attributes of each UGC, and provide a richer and more effective data foundation for subsequent UGC retrieval, sorting, recommendation and other tasks. By combining the content information of UGC with user interactive behavior, it is possible to more comprehensively evaluate important indicators such as the quality, popularity, and influence of UGC, and to dig out valuable information such as hot topics and influential users, thereby achieving more intelligent and accurate UGC services.

[0037] Specifically, associating and matching UGC basic information data with UGC interactive behavior data usually requires the following steps: first, load the extracted UGC basic information data and interactive behavior data into the same data processing framework (such as Spark, Flink, etc.); then, use the data association and connection operations (such as join, leftOuterJoin, etc.) provided by the data processing framework to connect and merge the two data sets with ugc_id as the association key; then, perform necessary field selection, renaming, type conversion and other operations on the merged data to form a standardized UGC information data table; finally, persist the integrated UGC information data in an appropriate data storage system (such as Hive, HBase, Elasticsearch, etc.) for subsequent retrieval, analysis and application.

[0038] In the process of association and matching, some additional data processing and aggregation operations are usually required, such as calculating the number of likes, comments, and reposts of each UGC, extracting semantic features such as keywords and topic tags from the UGC text content, and building a social relationship network for UGC. These additional data processing and aggregation operations can further enrich the dimensions and connotations of UGC information data, and provide more valuable features and signals for subsequent UGC retrieval and recommendation.

[0039] The advantage of associating and matching UGC basic information data with UGC interactive behavior data is that it makes full use of the multi-source heterogeneous characteristics of UGC data, organically integrates structured basic information data and unstructured interactive behavior data, and forms a comprehensive, three-dimensional, and rich UGC information data view. This data view not only includes the content characteristics of UGC, but also includes user feedback and interaction with UGC, which can more comprehensively and objectively evaluate the value and influence of UGC. At the same time, through data association and matching, the inherent connections and correlations between UGC can be discovered, and high-level semantic information such as topics, events, and communities can be excavated, providing important data support for intelligent retrieval and personalized recommendations of UGC.

[0040] S102, performing multi-way recall retrieval according to UGC semantic features and user search terms to obtain a UGC candidate set related to the user search terms; Specifically, after obtaining the UGC candidate set of multi-path recall retrieval, multiple recall result candidate sets in the candidate set are fused in reverse order to obtain a sorted recall result candidate set. Specifically, this embodiment designs an innovative inverse order fusion strategy for fusing and optimizing the candidate set sorting results generated by different recall paths.

[0041] For each recall result candidate set, first sort the candidate UGC items in descending order according to the relevance score to obtain an ordered candidate set sequence. Then, for each candidate UGC item, calculate the sum of the reciprocals of its ranking position in each candidate set as its fusion score. Finally, sort all candidate UGC items in descending order according to the fusion score to obtain a sorted recall result candidate set.

[0042] The core idea of ​​this inverse position fusion strategy is to give priority to UGC candidates that rank high in multiple recall paths, believing that they have a higher comprehensive relevance to the user's search terms. By summing the inverse positions, the contribution of different recall paths is cleverly balanced, taking into account both the ranking advantage of the candidate in a single recall path and its comprehensive performance in multiple recall paths.

[0043] Compared with simple score-weighted fusion, the advantages of inverse position fusion are: first, it is insensitive to the differences in score distribution of different recall paths, which can effectively avoid the problem that the score of a certain recall path is too high and dominates the entire fusion result; second, it is more sensitive to differences in ranking positions, which can highlight the candidates with higher comprehensive rankings and weaken the influence of the candidates with lower rankings.

[0044] For example, assuming that a UGC candidate is ranked 1st, 2nd, and 5th in the three recall result candidate sets, its inverted position fusion score is 1 / 1+1 / 2+1 / 5=1.7. The ranking positions of another candidate are 2nd, 3rd, and 3rd, and its fusion score is 1 / 2+1 / 3+1 / 3≈1.17. It can be seen that although the absolute ranking of the latter candidate is not bad, the former candidate has a higher comprehensive ranking in multiple recall paths, so it obtains a higher fusion score in the inverted position fusion, and thus ranks higher in the final fusion ranking.

[0045] Based on the above embodiment, as an optional embodiment, a multi-way recall search is performed according to UGC semantic features and user search terms to obtain a UGC candidate set related to the user search terms, including: S301, calculating the similarity between the text content of the user search term and the text content in the UGC semantic feature data based on the BM25 algorithm to obtain a first recall result candidate set; Specifically, the first step of performing multi-way recall retrieval based on UGC semantic features and user search terms is to calculate the similarity between the text content of the user search terms and the text content in the UGC semantic feature data based on the BM25 algorithm to obtain the first recall result candidate set. Specifically, this embodiment uses the classic text retrieval algorithm BM25 to calculate the correlation score between the user search terms and the UGC text content, and generates the first recall result candidate set based on the score.

[0046] The main purpose of using the BM25 algorithm for text similarity calculation is to quickly and efficiently retrieve candidate results related to the user's search terms from massive UGC data. The BM25 algorithm is a text retrieval algorithm based on probability statistics. It comprehensively considers factors such as term frequency (TF) and inverse document frequency (IDF), and can better measure the relevance between query terms and documents. Compared with simple keyword matching, the BM25 algorithm can more intelligently understand the semantics of query terms, identify semantically related but differently expressed texts, and improve the accuracy and comprehensiveness of recall results.

[0047] Specifically, text similarity calculation based on the BM25 algorithm usually requires the following steps: first, perform preprocessing operations such as word segmentation and stop word removal on the user search terms and UGC text content, and convert them into a normalized Bag-of-Words representation; then, calculate the IDF value of each word based on the pre-trained IDF dictionary, that is, the scarcity of the word in the entire UGC corpus; then, for each UGC text, calculate the BM25 correlation score between each word and the user search term, and weightedly sum the scores of each word to obtain the overall correlation score between the UGC text and the search term; finally, sort all UGC texts in descending order according to the correlation score, and select the top K UGCs as the first recall result candidate set.

[0048] The advantage of calculating text similarity based on the BM25 algorithm is that it makes full use of text statistics and corpus knowledge, and can quickly and accurately retrieve candidate results related to the user's search terms in massive UGC data. By considering factors such as word frequency and inverse document frequency, the BM25 algorithm can identify UGC texts that are semantically related to the search terms but expressed differently, thereby improving the comprehensiveness and diversity of the recall results. At the same time, the computational complexity of the BM25 algorithm is relatively low, and it can achieve real-time retrieval response in milliseconds, meeting the high concurrency and low latency requirements in UGC search scenarios.

[0049] For example, suppose the user's search term is "summer skin care", and there are the following three texts in the UGC corpus: 1. "Summer skin care tips: sun protection, hydration, and cleansing are all essential!" 2. "In the hot summer, your skin needs this kind of care" 3. "Skin care product recommendations: top 10 must-have sunscreens for summer" For the first text, since it contains the complete search term "summer skin care", the BM25 relevance score is high; for the second text, although it does not explicitly contain the search term, the words "summer", "skin", "care" and so on are semantically related to the search term, so it can also obtain a high BM25 score; and the third text, although it contains the words "summer" and "skin care", mainly discusses skin care product recommendations, and its relevance to the search term is relatively weak, so the BM25 score is relatively low. Through the relevance calculation and sorting of the BM25 algorithm, the first and second texts can be selected into the first recall result candidate set, providing high-quality candidate data for subsequent refinement and re-ranking.

[0050] S302, calculating the similarity between the semantic feature vector of the user search term and the semantic feature vector in the UGC semantic feature data based on the embedding semantic similarity, to obtain a second recall result candidate set; Specifically, the second step of performing multi-way recall retrieval based on UGC semantic features and user search terms is to calculate the similarity between the semantic feature vector of the user search term and the semantic feature vector in the UGC semantic feature data based on the embedding semantic similarity, and obtain the second recall result candidate set. Specifically, this embodiment utilizes the word embedding technology in deep learning to map the user search term and the UGC text to the same low-dimensional dense vector space, and calculates the semantic similarity between them through the distance measurement between the vectors, thereby generating the second recall result candidate set.

[0051] The main purpose of using Embedding semantic similarity for recall retrieval is to overcome the limitations of traditional keyword matching methods and achieve a deep understanding and matching of user search intent and UGC semantic content. The semantic similarity calculation based on Embedding not only considers the surface form of the vocabulary, but also the contextual relationship and semantic connection between words. It can discover UGC content with different word expressions but similar semantics, greatly improving the accuracy and diversity of recall results. Compared with methods based on word frequency statistics such as BM25, Embedding semantic similarity can better capture the deep semantic information of user search terms and UGC texts, and achieve a more intelligent and humanized search experience.

[0052] Specifically, recall retrieval based on Embedding semantic similarity usually requires the following steps: first, use a pre-trained word embedding model (such as Word2Vec, GloVe, etc.) to map the user search term and each word in the UGC text into a low-dimensional dense vector; then, for the user search term and each UGC text, obtain the corresponding semantic feature vector by averaging all the word vectors (or other aggregation methods); then, calculate the similarity between the semantic feature vector of the user search term and the semantic feature vector of each UGC (such as cosine similarity, Euclidean distance, etc.) to obtain the semantic similarity score between each UGC and the search term; finally, sort all UGCs in descending order according to the semantic similarity score, and select the top N UGCs as the second recall result candidate set.

[0053] The advantage of recall retrieval based on Embedding semantic similarity is that it makes full use of deep learning and natural language processing technology, and can match and understand user search terms and UGC content at the semantic level. By mapping words to a low-dimensional dense vector space, the Embedding method can mine the semantic relationships and contextual connections between words, discover UGC content with different expressions but similar meanings, and greatly improve the semantic relevance and diversity of the recall results. At the same time, thanks to the rapid development of deep learning technology and the improvement of computing power, Embedding semantic similarity calculation can also achieve a sub-second real-time response speed, meeting the high concurrency and low latency requirements in UGC search scenarios.

[0054] For example, for the user search term "summer skin care", the traditional keyword matching method can only retrieve UGC texts containing the words "summer" and "skin care". Based on the embedding semantic similarity, the following UGC texts can also be retrieved: 1. "How should girls take good care of their skin in the hot summer?" 2. "Teach you how to protect your skin from the sun and whiten your skin in hot weather!" 3. "Skin problems are common in the hot summer, so it is necessary to do this!" Although these UGC texts do not explicitly contain the words "summer skin care", they are closely related to topics such as summer skin care and sun protection, and contain rich knowledge and experience related to summer skin care, which has a high reference value for users' search needs. Through the Embedding semantic similarity calculation, these semantically related high-quality UGC content can be successfully recalled and enter the second recall result candidate set, thereby greatly enriching the coverage and diversity of the candidate results and providing better candidate data for subsequent sorting and recommendation.

[0055] S303: Merge the first recall result candidate set and the second recall result candidate set to obtain a UGC candidate set.

[0056] Specifically, after obtaining the first recall result candidate set and the second recall result candidate set, the first recall result candidate set and the second recall result candidate set are merged to obtain the UGC candidate set. The purpose of this step is to summarize the results obtained by text similarity recall and semantic similarity recall to form a candidate set covering the two recall paths. The merge operation can simply splice the two result sets together without the need for complex processing such as deduplication. This is because even if the same UGC appears in both result sets, their relevance and ranking may be different. Retaining duplicate UGC can provide more information and possibilities for subsequent sorting.

[0057] S103, performing inverse sorting and merging of multiple recall result candidate sets in the UGC candidate set to obtain a sorted recall result candidate set; Specifically, after obtaining the UGC candidate set, it needs to be further sorted and optimized to improve the quality of the final retrieval results. Since the UGC candidate set is formed by merging multiple recall result candidate sets, the candidate set rankings obtained by different recall paths may be inconsistent, and the ranking effect after direct merging may not be good. In order to better utilize the ranking information of different recall paths, the multiple recall result candidate sets in the UGC candidate set are reciprocal ranked fusion (RRF) to obtain the sorted recall result candidate set.

[0058] RRF is a method that combines multiple result sets with different relevance indicators into a single result set. It does not rely on specific relevance scores, but calculates the fused ranking based on the ranking of candidates in different result sets. The specific process is as follows: First, each recall result candidate set in the UGC candidate set is sorted independently, and a sorted list of each recall result candidate set is obtained based on indicators such as the relevance score and quality score of the candidate UGC items.

[0059] Then, for each candidate UGC item, calculate its reciprocal ranking in the sorted list of each recall result candidate set. For example, if a candidate UGC item ranks first in the sorted list of a recall result candidate set, then its reciprocal ranking is 1; if it ranks second, its reciprocal ranking is 1 / 2=0.5, and so on.

[0060] Next, for each candidate UGC item, add up its inverse ranking in the ranking list of each recall result candidate set to obtain the fusion ranking score of the candidate UGC item. The higher the fusion ranking score, the higher the comprehensive ranking of the candidate UGC item under multiple recall paths.

[0061] Finally, all candidate UGC items are sorted according to the fusion ranking score of each candidate UGC item to obtain the sorted recall result candidate set. The sorted candidate set comprehensively considers the sorting information of multiple recall result candidate sets and can more accurately and comprehensively reflect the relevance of candidate UGC items.

[0062] Based on the above embodiment, as an optional embodiment, multiple recall result candidate sets in the UGC candidate set are sorted and merged in reverse order to obtain a sorted recall result candidate set, including: S401, independently sorting each recall result candidate set in the UGC candidate set to obtain a sorted list of each recall result candidate set, wherein each candidate UGC item in the recall result candidate set corresponds to a sorted position; Specifically, after obtaining the UGC candidate set, it needs to be further sorted and optimized to improve the quality of the final retrieval results. Since the UGC candidate set is formed by merging multiple recall result candidate sets, the candidate set rankings obtained by different recall paths may be inconsistent, and the sorting effect after direct merging may not be good. In order to better utilize the sorting information of different recall paths, multiple recall result candidate sets in the UGC candidate set are sorted and merged in reverse order to obtain a sorted recall result candidate set.

[0063] The first step of the inverse sort fusion is to independently sort each recall result candidate set in the UGC candidate set to obtain a sorted list of each recall result candidate set, where each candidate UGC item in the recall result candidate set corresponds to a sorted position.

[0064] Specifically, for each recall result candidate set in the UGC candidate set, the candidate UGC items it contains need to be sorted according to certain sorting rules. Common sorting rules can be based on indicators such as the relevance score, quality score, user interaction data, etc. of the candidate UGC items, or they can comprehensively consider multiple indicators to generate a comprehensive ranking score. Regardless of the sorting rule used, the purpose is to prioritize the candidate UGC items based on the degree of match between the candidate UGC items and the user's search intent, so that the candidate UGC items that are most relevant to the search intent and have the highest quality are ranked first.

[0065] After sorting, each candidate set of recall results will get a sorted list. In the sorted list, each candidate UGC item corresponds to a clear sorting position. The higher the sorting position (the smaller the value), the higher the relevance of the candidate UGC item to the user's search intent and the better the quality. For example, if a candidate UGC item ranks third in the sorted list, then its sorting position is 3.

[0066] The purpose of independently sorting each recall result candidate set is that different recall paths may use different sorting rules and algorithms, and directly merging their original scores may lead to deviations. By independently sorting, we can find the best candidate UGC items within the same recall path and use their ranking positions to represent their priorities, avoiding the problem of inconsistent sorting scales between different recall paths and providing comparable sorting information for subsequent reciprocal sorting fusion.

[0067] S402, for each candidate UGC item, calculating the reciprocal ranking position value of the candidate UGC item in the ranking list of each recall result candidate set, where the reciprocal ranking position value is equal to the difference between the ranking position of the candidate UGC item and the length of the corresponding ranking list; Specifically, after independently sorting each recall result candidate set in the UGC candidate set to obtain the sorted list of each recall result candidate set, it is necessary to further calculate the reciprocal sort position value of each candidate UGC item in different sorted lists to prepare for the subsequent reciprocal sort fusion.

[0068] Specifically, for each candidate UGC item, the reciprocal ranking position value of the candidate UGC item in the ranking list of each recall result candidate set is calculated, and the reciprocal ranking position value is equal to the difference between the ranking position of the candidate UGC item and the length of the corresponding ranking list.

[0069] The purpose of this step is to normalize the ranking positions of candidate UGC items in different ranking lists to obtain an indicator that reflects their relative position in the ranking list. The original ranking position is a positive indicator. The smaller the value, the higher the priority of the candidate UGC item. The reciprocal ranking position value is a reverse indicator. The larger the value, the later the candidate UGC item is in the corresponding ranking list and the lower its priority.

[0070] S403, for each candidate UGC item, based on the inverse ranking position value of the candidate UGC item in the ranking list of each recall result candidate set, calculate the comprehensive inverse ranking value of each candidate UGC item, and sort each candidate UGC item based on the comprehensive inverse ranking value to obtain a sorted recall result candidate set.

[0071] Specifically, after calculating the reciprocal ranking position value of each candidate UGC item in the ranking list of each recall result candidate set, the last step of reciprocal ranking fusion can be performed.

[0072] Specifically, for each candidate UGC item, the comprehensive inverse ranking value of each candidate UGC item is calculated based on the inverse ranking position value of the candidate UGC item in the sorted list of each recall result candidate set, and each candidate UGC item is sorted based on the comprehensive inverse ranking value to obtain a sorted recall result candidate set.

[0073] The purpose of this step is to comprehensively consider the ranking performance of candidate UGC items under different recall paths, calculate an indicator that can reflect its comprehensive priority, and use this as a basis to globally sort all candidate UGC items to obtain the final ranking result.

[0074] There are many specific methods for calculating the comprehensive inverse ranking value. A common approach is to sum or average the inverse ranking position values ​​of the candidate UGC items in each ranking list. For example, the inverse ranking position values ​​of the candidate UGC items in each ranking list can be directly added to obtain its comprehensive inverse ranking value; the inverse ranking position values ​​can also be weighted and summed, and the weights of different ranking lists can be set according to their importance or credibility. The smaller the comprehensive inverse ranking value, the higher the comprehensive ranking of the candidate UGC item under multiple recall paths, and the higher the priority.

[0075] After obtaining the comprehensive inverted ranking value of each candidate UGC item, all candidate UGC items are sorted in ascending order according to the comprehensive inverted ranking value to obtain the sorted recall result candidate set. The candidate UGC items at the top have smaller comprehensive inverted ranking values, indicating that they have better ranking performance under multiple recall paths, have higher overall relevance to the user's search intent, and should be displayed to the user first.

[0076] S104, re-ranking the sorted recall result candidate set through the BGE rerank model to obtain the re-ranked recall result candidate set and the re-ranking score of each UGC candidate item in the re-ranked recall result candidate set, and using the re-ranked recall result candidate set and the re-ranking score of each UGC candidate item in the re-ranked recall result candidate set as the UGC retrieval result.

[0077] Specifically, the sorted candidate set of recall results is further rearranged and optimized. Specifically, this embodiment introduces a BERT-based sorting model BGE rerank for fine-tuning the candidate set of recall results.

[0078] The BGE rerank model is a pre-trained transformer language model. It is pre-trained with massive UGC corpus and user interaction data, and has learned rich semantic representations and user preference information. In the reranking stage, each recall candidate UGC item is concatenated with the user search term into a sequence, which is input into the BGE rerank model for feature extraction and correlation calculation to obtain the reranking score of each candidate item. Then, the recall candidate set is sorted in descending order according to the reranking score to obtain the reranked recall result candidate set.

[0079] The main purpose of introducing the BGE rerank model for re-ranking is to further improve the relevance and ranking quality of the recall results. Compared with inverted position fusion, BGE rerank can more comprehensively and deeply explore the semantic association and user preference matching between UGC candidates and user search terms. It makes full use of the powerful semantic representation capabilities of large-scale pre-trained language models and the user preference information contained in massive user interaction data, so as to more accurately evaluate and score the relevance of candidates.

[0080] In addition, the BGE rerank model can also flexibly incorporate various additional feature information, such as the quality score of UGC, historical interaction information between users and candidates, etc., to comprehensively evaluate the candidates from multiple dimensions, so that the reranking results can better meet the personalized needs of users.

[0081] For example, for the user search term "summer skin care", the BGE rerank model can not only capture the semantic relevance between the candidate UGC and the search term, but also consider the quality factors of the candidate UGC (such as the number of likes, the number of favorites, etc.), as well as the user's historical browsing preferences for the topic, and comprehensively calculate the reranking score of each candidate. In this way, it can ensure the relevance of the reranking results to the search term, while taking into account the quality of UGC and the user's personalized preferences, making the ranking results more comprehensive and accurate.

[0082] Finally, the candidate set of rearranged recall results is returned to the user as the UGC search result, and the rearrangement score of each candidate is also presented as its relevance score. This approach not only provides users with a highly relevant and high-quality UGC search result ranking, but also intuitively presents the degree of match between each result and the search term to users in the form of relevance scores, improving the interpretability of the results and user experience.

[0083] Based on the above embodiment, as an optional embodiment, the sorted recall result candidate set is re-ranked by a preset BGE rerank model to obtain a re-ranked recall result candidate set and a re-ranking score of each UGC candidate item in the re-ranked recall result candidate set, including: S501, extracting features from the candidate UGC items in the sorted recall result candidate set to obtain a multi-dimensional feature vector of the candidate UGC items; Specifically, the first step of re-ranking the sorted recall result candidate set through the preset BGE rerank model is to extract features from the candidate UGC items in the sorted recall result candidate set to obtain a multi-dimensional feature vector of the candidate UGC items. Specifically, this embodiment uses natural language processing, image recognition and other technical means for the sorted candidate set obtained in the previous step, and comprehensively describes the internal features and external performance of the candidate UGC items from multiple perspectives such as text content, visual information, and user interaction, so as to provide rich and high-quality feature support for subsequent refined re-ranking.

[0084] The main purpose of extracting multi-dimensional features from candidate UGC items is to make full use of the effective information contained in the recall results after sorting, to deeply understand and characterize the candidate UGC items from multiple dimensions such as semantics, emotions, topics, and user feedback, and to explore their internal laws and key influencing factors in different aspects. Compared with the original UGC text, the extracted multi-dimensional feature vector can characterize the characteristics of the candidate UGC items in a more structured and comprehensive manner, covering both the semantic information of the text content itself and the evaluation feedback of external users on the content, reflecting the quality, popularity, and attractiveness of UGC from multiple perspectives. High-quality features are the key to optimizing the rerank model and the basic guarantee for quality improvement in UGC search and recommendation scenarios.

[0085] Specifically, extracting multi-dimensional features from candidate UGC items usually requires the following steps: first, for each candidate UGC item in the sorted recall result set, obtain its complete text content, visual information (such as pictures, videos, etc.), user interaction data (such as likes, comments, favorites, etc.) and other original materials; then, design corresponding feature extractors for different types of original materials, such as text feature extraction can be based on common algorithms such as TF-IDF and Word2Vec, visual feature extraction can be based on classic models such as CNN and SIFT, and user interaction features can directly count the original data; then, use the constructed feature extractor to process the original materials to obtain the structured feature representation of the candidate UGC items in different dimensions; finally, concatenate and combine the structured features of each dimension to form a multi-dimensional feature vector of the candidate UGC item as the input of the downstream rerank model.

[0086] The advantage of multi-dimensional feature extraction is that it can automatically learn and extract concise, comprehensive, and informative feature representations from massive, heterogeneous UGC data, mapping the widely varying UGC content into a common, computable feature space, which not only greatly compresses the data size, but also greatly promotes the efficiency and effectiveness of subsequent sorting optimization. By considering all aspects of UGC in a multi-pronged manner, this method can more realistically and three-dimensionally portray the intrinsic value of each candidate, so that subsequent sorting is no longer limited to simple text matching, but can be refined from multiple dimensions to consider the quality, popularity, and interaction of UGC, thereby obtaining a more comprehensive and accurate re-ranking result, significantly improving the overall effect of UGC recommendations. At the same time, multi-dimensional features also provide a good data foundation for subsequent applications such as higher-level user portraits and UGC clustering.

[0087] For example, suppose that a candidate UGC item A is selected from the sorted recall result candidate set. For the text content of A, its TF-IDF text vector, sentiment tendency features, topic distribution features, etc. can be extracted; for the visual information of A, its color histogram features, texture features, key object features, etc. can be extracted; for the user interaction data of A, its number of likes, comments, reposts, favorites, etc. can be extracted. The above-mentioned dimensional features are spliced ​​in a certain way to obtain the multi-dimensional feature vector of the candidate UGC item A: A:{TF-IDF vector, sentiment orientation, topic distribution, color histogram, texture features, key objects, number of likes, number of comments, number of reposts, number of favorites} This feature vector covers multiple key attributes of UGC, such as semantics, emotions, topics, vision, and user feedback, and comprehensively characterizes its content characteristics and external performance. Once the multi-dimensional feature vectors of all candidate UGC items are extracted, they can be input into the preset rerank model for sorting optimization.

[0088] S502, input the multi-dimensional feature vector into a preset BGE rerank model to obtain a reranking score of each UGC candidate item, and sort the sorted recall result candidate set according to the reranking score of each UGC candidate item to obtain a reranked recall result candidate set.

[0089] Specifically, the second step of re-ranking the sorted recall result candidate set through the preset BGE rerank model is to input the multi-dimensional feature vector into the preset BGE rerank model to obtain the re-ranking score of each UGC candidate item, and sort the sorted recall result candidate set according to the re-ranking score of each UGC candidate item, and finally obtain the re-ranked recall result candidate set. Specifically, this embodiment uses the pre-trained BGE rerank model to learn and comprehensively evaluate the feature weights based on the multi-dimensional features of the candidate UGC items generated in the previous step, makes a quantitative judgment on the quality of the candidate items, and generates a re-ranking score based on this as a direct basis for fine-grained sorting optimization. The addition of the BGE rerank model further improves the sorting quality, so that the final re-ranked recall result can be more accurate and more in line with user needs.

[0090] The main purpose of inputting the multi-dimensional feature vectors of candidate UGC items into the BGE rerank model and generating reranking scores is to make full use of the advantages of machine learning sorting technology, automatically learn feature combination weights from massive user behavior data, and make more accurate and personalized judgments on the comprehensive quality of candidate items. Compared with simple strategies such as weighted average and summation used in fusion sorting, the BGE rerank model can better explore and utilize the correlation between multi-dimensional features through end-to-end training optimization, find key feature combinations that are highly correlated with UGC quality and user interests, and give them higher weight coefficients. The addition of the model has transformed the reranking process from experience-driven to data-driven, incorporating the wisdom of the crowd from a global perspective, and significantly improving the quality and objectivity of the sorting results.

[0091] Specifically, the process of generating the re-ranking scores of candidate UGC items using the BGE rerank model is as follows: first, according to the goal of ranking optimization, select a suitable machine learning ranking model (such as RankNet, LambdaRank, etc.) as the basic framework of BGE rerank; then, adapt the model structure to the objective function of ranking learning, and train and optimize the model based on large-scale implicit feedback data such as user clicks and dwell time; then, use the multi-dimensional feature vector of the candidate UGC item obtained in the previous step as input and feed it into the trained BGE rerank model. After a series of calculations and transformations, output the re-ranking score of each candidate item as a quantitative representation of its comprehensive quality and priority; finally, sort the candidate UGC items in descending order according to their re-ranking scores to obtain a candidate set of re-ranked recall results as the final recommended candidate items generated in this stage.

[0092] The advantage of introducing the BGE rerank model for fine-grained re-ranking is that it can make full use of the feature information of multiple dimensions of UGC and automatically learn the optimal feature combination weights from the data, greatly improving the accuracy, relevance and personalization of the ranking results. Through end-to-end optimization training, the BGE rerank model has mastered the distribution patterns and association patterns of candidate UGC items in different feature dimensions, and can evaluate their intrinsic value more finely and comprehensively, and identify high-quality content that best matches users' long-term and short-term interests. At the same time, the addition of the model also makes the re-ranking process more flexible. It can achieve continuous self-improvement as data accumulates, continuously learn new ranking rules, and adaptively present the best recommendation results, thereby significantly improving users' subjective satisfaction and interactive stickiness.

[0093] For example, the multi-dimensional feature vector of candidate UGC item A obtained in the previous step is input into the preset BGE rerank model: A:{TF-IDF vector, sentiment orientation, topic distribution, color histogram, texture features, key objects, number of likes, number of comments, number of reposts, number of favorites} Assume that after calculation, the BGE rerank model outputs a reranking score of 0.85 for candidate UGC item A, which means that after comprehensively considering factors such as semantics, emotions, vision, and social influence of A, the model believes that it has a high overall quality and priority and is a high-scoring, high-quality candidate item. After the reranking scores of all candidate UGC items are calculated, they are arranged in descending order according to the reranking scores, and the candidate set of reranking recall results is obtained, as shown below: C (0.92)>B (0.89)>A (0.85)>F (0.80)>D (0.76)>E (0.74) It can be seen that the BGE rerank model significantly changes the order of candidate items after fusion sorting, so that the originally parallel candidate items are further subdivided and rearranged according to the reranking scores. This shows that the model has learned multi-dimensional feature weights and incorporated feedback from global behavioral data, making the judgment of the quality of candidate UGC items more comprehensive and accurate, and can better capture the user's real interest preferences.

[0094] Based on the above embodiment, as an optional embodiment, before the multi-dimensional feature vector is input into the preset BGE rerank model to obtain the reranking score of each UGC candidate item, the method further includes: S601, obtaining business scenario data and a reordering training data set, and defining a business reordering rule according to the business scenario data; Specifically, before inputting the multi-dimensional feature vector into the preset BGE rerank model and obtaining the reranking score of each UGC candidate, an important preparatory step is also included, namely, obtaining business scenario data and reranking training data sets, and defining business reranking rules based on the business scenario data. Specifically, before officially launching BGE rerank, this embodiment first collects and extracts various types of original data closely related to sorting optimization from actual business scenarios and historical data, and summarizes a set of business reranking rules that meet application requirements and guide model training, thereby laying a solid data and knowledge foundation for subsequent feature learning and model optimization. The prior acquisition of data and definition of rules helps the BGE rerank model to better adapt to specific business environments, learn domain-specific sorting patterns, and ultimately generate more accurate and commercially valuable reranking results.

[0095] The main purpose of obtaining business scenario data, re-arranging training data sets, and defining business re-ranking rules is to make full use of prior knowledge and historical data in application scenarios to provide targeted training guidance and evaluation criteria for the BGE rerank model. Compared with training from scratch, with the support of data in specific fields, the BGE model can more efficiently and accurately grasp the sorting rules in the field and learn the feature combination patterns that best reflect the business value. In addition, the definition of business re-ranking rules also helps to incorporate external factors such as business demands and operational strategies into model optimization considerations, so that it can take into account more diverse interests while pursuing sorting quality. The dual guidance of data and rules will ultimately make the BGE rerank model more domain-adaptable and goal-driven, and its output results will be closer to actual business needs.

[0096] Specifically, the process of obtaining business scenario data, re-ranking the training data set and defining the re-ranking rules is as follows: First, according to the application scenarios of UGC search and recommendation systems, various data sources that can reflect business characteristics and user preferences are selected, including user attributes, historical clicks, interactive feedback, UGC metadata, etc.; then, the original data is collected, cleaned, sampled and other pre-processed, and organized into a re-ranked training data set according to certain format and scale requirements, which serves as a direct sample for the BGE model to learn the sorting rules; then, business experts and data analysts deeply explore the business scenario data, and mine key information that can guide sorting optimization, such as hot topics, popular elements, high-quality features, etc., and summarize a series of re-ranking rules that meet business needs based on this; finally, the business re-ranking rules are embedded in the training objectives and evaluation indicators of the BGE model to guide the model to optimize in a specific direction, in order to obtain the best sorting effect that fits the business scenario.

[0097] Compared with pure data-driven ranking learning, the advantages of BGE rerank supported by business data and guided by rules are mainly reflected in the following points: First, by introducing business scenario data, the model can understand the multiple characteristics of UGC, users, and scenarios more comprehensively and deeply, and learn the key attribute combination that best reflects the value of content from a domain perspective; second, the construction of the re-ranking training data set provides direct learning samples for model optimization, enabling it to grasp the ranking rules in the field more efficiently and accurately; third, the business re-ranking rules convert expert experience and business strategies into computable training targets, which not only indicates the optimization direction for the model, but also enables its output results to take into account more diverse interests. In general, the introduction of data and rules has greatly enhanced the domain adaptability and goal-driven nature of the BGE model, so that the re-ranking sequences it generates can better match specific business scenarios, presenting recommendation results with more accurate content and more commercially valuable rankings.

[0098] For example, suppose you need to develop a BGE rerank model for a short video search and recommendation system. First, you need to obtain multi-source heterogeneous data that can reflect the business characteristics and user behavior of the system, including video meta information (such as title, category, release time, etc.), user portraits (such as age, gender, interest tags, etc.), and interactive feedback data (such as clicks, likes, comments, and stay time). Then, clean and sample the collected raw data to build a moderately sized and comprehensive reranking training set, such as: {"video_id":"v1001","title":"Funny and Cute Pets","category":"Pets","duration":32,"click":10000,"like":500,"comment_num":200,"stay_time":25} Next, business experts and data analysts conducted in-depth research on the scenario data and summarized the following re-arrangement business rules: 1. For short video content, videos with moderate length (30-90 seconds) and concise titles should be preferred; 2. Among the interactive feedback features, we should focus on active feedback indicators such as likes and comments as the core reference for video quality and user preferences; 3. The primary goal of sorting optimization is to increase the user's playback completion rate and interaction rate, which will be used as evaluation indicators for the rearrangement model.

[0099] Finally, based on the re-ranked training set, the optimization training of the BGE rerank model is carried out based on business rules, and the model effect is evaluated based on the evaluation indicators defined by the rules. Once the BGE rerank model learns a relatively stable ranking mode, it can be put into actual business scenarios to support and strengthen the ranking of UGC search and recommendation systems.

[0100] S602, quantifying the business reordering rules into computable business rule indicators, and processing the reordering training data set according to the business reordering rules to generate corresponding business reordering features; Specifically, after obtaining business scenario data, re-arranging the training data set and defining the business re-arrangement rules, it is necessary to further quantify the business re-arrangement rules into computable business rule indicators, and process the re-arrangement training data set according to the business re-arrangement rules, so as to generate corresponding business re-arrangement features. Specifically, after obtaining the business re-arrangement rules in text form, this embodiment converts them into a series of measurable and computable numerical business rule indicators through feature engineering means, so that they can directly participate in the training and optimization process of the model. At the same time, based on the business re-arrangement rules, the original samples in the re-arrangement training data set are screened, combined, transformed, and processed to extract the re-arrangement features closely related to the business rules, providing solid support for the rule-oriented learning of the BGE model. Embedding business rules into the data foundation of BGE rerank helps to strengthen its domain adaptability and goal-drivenness from a multi-dimensional and fine-grained perspective, so that the model outputs a ranking result that is more in line with business expectations.

[0101] The main purpose of quantifying the business re-ranking rules into business rule indicators and processing the training data to generate business re-ranking features is to give full play to the guiding role of business rules in BGE model training, so that its optimization process can include more refined and comprehensive domain priors and target constraints. Through the quantification of rule indicators, the BGE model can transform the experience insights of business experts into learnable and measurable optimization goals, and integrate rule-driven on the basis of data-driven, so as to form a more accurate and targeted grasp of the sorting mode. At the same time, relying on business rules to enhance the original samples in the re-ranking training data set, on the one hand, it can highlight the rule-related features and guide the model to focus on the information dimension that best reflects the business value; on the other hand, it also expands the scale and diversity of samples, further improving the effect of model learning. The quantitative embedding of business rules and data enhancement will eventually give BGE rerank stronger domain cognition and purposefulness, so that the learned sorting mode can fully reflect the characteristics of the industry and match the business demands.

[0102] Specifically, the process of quantifying business reordering rules into business rule indicators and generating business reordering features is as follows: First, for each business reordering rule in text form, a corresponding measurable and calculable numerical business rule indicator is designed to characterize the quantitative performance of the rule. For example, for the rule of "preferring short videos with moderate length", a "duration score" can be defined as a business indicator, and the calculation formula of the indicator is given; then, the business rule indicator is used as an auxiliary goal of model training, and is reflected in the sample annotation, loss function design and other links, guiding the BGE model to learn from the perspective of rule constraints. and optimization, such as incorporating the weighted sum of various business indicators into the objective function of ranking learning; then, screening and cleaning the rearrangement training data set according to the business rearrangement rules, filtering out noise samples that do not meet the rule requirements, highlighting the core features related to the rules, and implementing rich and balanced data enhancement on the original samples in combination with the rules to expand the learning materials of the BGE model; finally, integrating the business rule indicators and rearrangement features into the training data of the BGE model to form rule-enhanced learning samples, which support the model to model business scenarios from both knowledge and data dimensions and learn the sorting mode that best meets business needs.

[0103] Compared with simple end-to-end learning, the advantages of BGE rerank after business rule quantification and data enhancement are mainly reflected in the following points: First, the abstract business rules are transformed into a computable indicator system, which provides a clear and measurable optimization direction for sorting optimization, and guides the BGE model to take into account various business considerations while pursuing overall results; second, the training data set is screened and transformed using rules, which not only highlights the samples and features that best reflect business value, but also expands the scale and richness of learning materials, thereby enhancing the quality and effect of model training; third, rule-driven is integrated on the basis of data-driven to form a collaborative learning paradigm of knowledge and data, which not only gives full play to the automatic optimization ability of machine learning, but also uses domain knowledge to effectively guide the model. In general, the quantification of business rules and the regular processing of data will eventually enable the BGE model to form a composite learning model of "data-driven and rule-guided", which will better fit the business scenario while improving the accuracy of sorting, and produce a re-ranked sequence with more accurate content and more commercial value.

[0104] For example, for the business scenario of short video search and recommendation, after obtaining the rearrangement business rule "for short video content, videos with moderate length (30-90 seconds) and concise titles should be preferred", the following two business rule indicators can be extracted from it: 1. Duration score = 1 - | duration - 60 | / 60, where duration is the video duration in seconds; 2. Title complexity = 1 / log(1 + title_length), where title_length is the number of words in the video title.

[0105] The above two indicators quantify the degree of compliance between video content and business rules from the perspectives of duration fit and title simplicity. The higher the value, the more it meets the requirements of the rule. When optimizing the BGE model, the duration score and title complexity can be included as auxiliary training targets to guide the model to learn the corresponding sorting preferences, such as adding the weighted items of duration score and title complexity to the learning objective function: Loss = Ranking_Loss + 0.1 * Duration_Score + 0.05 * Title_Simplicity At the same time, based on business rules, the rearrangement training data set is trimmed and enhanced. On the one hand, video samples with too long / too short duration or complicated titles are filtered out to highlight the core features related to the rules; on the other hand, data enhancement is performed on the original samples based on the rules, such as random trimming, slow-down and fast-forwarding transformations on videos of moderate length, to generate more samples that meet the expectations of the rules.

[0106] After rule indicator quantification and data enhancement, a BGE rerank training set integrating business rule information is formed. Each sample not only contains the multi-dimensional features of the video, but also incorporates business indicators such as duration score and title complexity, such as: {"video_id":"v1001","title":"Funny and Cute Pets","category":"Pets","duration":32,"duration_score": 0.93,"title_simplicity": 0.85,"click": 10000,"stay_time": 25} By carrying out BGE model learning on this training set, data-driven and rule-guided collaborative optimization can be achieved, and a sorting mode that can accurately grasp the content quality and fit business factors such as video length and title simplicity can be learned. This composite learning paradigm of knowledge and data will ultimately make the BGE model more domain-adaptable and goal-driven, and output a rearranged sequence that is highly matched with the expected scenario.

[0107] Based on the above embodiment, as an optional embodiment, the reordering training data set is processed according to the service reordering rule to generate corresponding service reordering features, including: Feature data is extracted from the re-arrangement training data set to obtain re-arrangement training feature data, and the re-arrangement training feature data is merged to obtain business re-arrangement features.

[0108] S603, constructing an initial BGE rerank model based on the original BGE rerank model and the service reranking characteristics, and training the initial BGE rerank model based on the reranking training data set, and obtaining a preset BGE rerank model when the training result reaches a preset accuracy value.

[0109] Specifically, after obtaining the business re-ranking features, it is necessary to further construct an initial BGE rerank model based on the original BGE rerank model and the business re-ranking features, and use the re-ranking training data set to train and optimize the initial BGE rerank model until the training result reaches the preset accuracy value, that is, to obtain a preset BGE rerank model that meets the business needs. Specifically, this embodiment is based on the original BGE sorting model, and the quantized business re-ranking features are embedded therein to form an initial BGE rerank model that integrates domain knowledge and end-to-end learning paradigm. On this basis, the initial model is trained using the rule-enhanced re-ranking training data set, and through data-driven and rule-guided collaborative optimization, the internal parameters are continuously refined to learn a sorting mode that is highly consistent with the business scenario. When indicators such as sorting accuracy reach the preset threshold, the construction of the preset BGE rerank model is completed. Fully integrating business rules and domain data into the entire process of model training can maximize the advantages of BGE rerank in intelligent sorting and business adaptation, making it a core engine and value tool in UGC search and recommendation scenarios.

[0110] The main purpose of building the initial model based on the original BGE rerank model and business re-ranking features, and conducting training optimization on the re-ranking training data set is to comprehensively strengthen the BGE ranking mechanism on the basis of knowledge fusion and data-driven, so that it can more accurately and comprehensively grasp the business scenarios and business demands while learning intelligent and efficient ranking strategies. On the one hand, by embedding the business re-ranking features in the original BGE model, the initial model can achieve the integration of end-to-end learning and rule research and learning at the architectural level, and flexibly absorb domain priors to form reasonable guidance for model optimization; on the other hand, the rule-enhanced re-ranking training data provides the initial model with rich and high-quality learning materials, enabling it to efficiently learn the inherent laws of business data and business rules from massive cases, forming a knowledge-driven and data-driven collaborative learning situation. With the dual support of algorithm design and data feeding, the initial BGE rerank model will eventually evolve into a preset BGE model with a deep understanding of business scenarios and a precise grasp of content value after sufficient training, providing strong intellectual support and effect guarantee for the healthy development of the UGC ecosystem.

[0111] Specifically, the process of constructing the initial model based on the original BGE rerank model and business re-ranking features and conducting training optimization is as follows: First, on the basis of the original BGE rerank model, the business re-ranking features are embedded as new input parameters to expand the model's ability to receive and understand business information, and at the same time, the network structure, objective function, etc. of the original model are adaptively adjusted to form an initial BGE rerank model; then, the rule-enhanced re-ranking training data set is input into the initial model, and the model learning is triggered by real business data. The training rhythm is controlled by setting reasonable hyperparameters (such as learning rate, regularization coefficient, etc.) to guide the model to balance between global optimization and local exploration; then, during the model training process, feedback signals such as sorting accuracy and quantitative values ​​of business indicators are continuously monitored, and training parameters and learning directions are adjusted in a timely manner according to the feedback, so as to drive the model to continuously optimize in the direction that best fits the business scenario; finally, when evaluation indicators such as model accuracy are stable above the preset threshold (such as 99%) for several consecutive epochs, it can be considered that the initial BGE model has fully learned and internalized the business rules and successfully evolved into the preset BGE rerank model.

[0112] Compared with the simple machine learning sorting, the advantages of the business rule-integrated and data-driven BGE rerank model are mainly reflected in the following points: First, the concept of knowledge integration is adhered to from the beginning of model construction. By embedding business features, the BGE sorting mechanism can directly connect and internalize industry accumulation and expert experience, forming a technical abstraction and algorithm framework that is closer to the essence of the business; second, using real business data as learning materials can not only help the model to fully perceive the information landscape of the business environment, but also fully explore the implicit laws contained in the data, transcending the sample limitations to achieve broader and deeper business insights; third, the domain rules are converted into quantitative indicators, and by setting optimization goals, designing loss functions and other means, a closed loop of rule research and end-to-end training is constructed to achieve all-round business adaptation and effect improvement from local to global, from explicit to implicit. In general, business integration and data-driven enable the BGE model to form a composite sorting intelligence. While sorting massive heterogeneous content in real time and accurately, it also forms a high degree of fit with business logic and value proposition, ultimately bringing multi-dimensional fission of content distribution, user experience and commercial value to the UGC platform.

[0113] On the other hand, the present application also provides a UGC retrieval system based on semantic information, such as Figure 2 The system comprises: Data processing module 1, used to obtain UGC information data and user search terms, and convert the UGC information data into UGC semantic feature data; Recall module 2, used to perform multi-way recall retrieval based on UGC semantic features and user search terms to obtain a UGC candidate set related to the user search terms; Sorting module 3, used to perform inverse sorting and fusion of multiple recall result candidate sets in the UGC candidate set to obtain a sorted recall result candidate set; The rearrangement module 4 is used to rearrange the sorted recall result candidate set through the BGE rerank model to obtain the rearranged recall result candidate set and the rearrangement score of each UGC candidate item in the rearranged recall result candidate set, and use the rearranged recall result candidate set and the rearrangement score of each UGC candidate item in the rearranged recall result candidate set as the UGC retrieval result.

[0114] Please refer to Figure 3 The application also discloses an electronic device. Figure 3 The electronic device 300 may include: at least one processor 301 , at least one network interface 304 , a user interface 303 , a memory 305 , and at least one communication bus 302 .

[0115] The communication bus 302 is used to realize the connection and communication between these components.

[0116] The user interface 303 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0117] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0118] Among them, the processor 301 may include one or more processing cores. The processor 301 uses various interfaces and lines to connect various parts in the entire server, and executes various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and calling data stored in the memory 305. Optionally, the processor 301 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field~Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The processor 301 can integrate one or a combination of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing the content to be displayed on the display screen; the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 301, and it can be implemented separately through a chip.

[0119] Among them, the memory 305 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read~Only Memory). Optionally, the memory 305 includes a non-transitory computer-readable medium. The memory 305 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playback function, an image playback function, etc.), instructions for implementing the above-mentioned various method embodiments, etc.; the data storage area may store data involved in the above-mentioned various method embodiments, etc. The memory 305 may optionally be at least one storage system located away from the aforementioned processor 301. Refer to Figure 3 , the memory 305 as a computer storage medium may include an operating system, a network communication module, a user interface module and an application program of a UGC retrieval method based on semantic information.

[0120] exist Figure 3In the electronic device 300 shown, the user interface 303 is mainly used to provide an input interface for the user and obtain the data input by the user; and the processor 301 can be used to call the application program storing the road assessment method in the memory 305. When executed by one or more processors 301, the electronic device 300 executes one or more methods in the above-mentioned embodiments. It should be noted that for the aforementioned method embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the described order of actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for the present application. In the above embodiments, the description of each embodiment has its own emphasis. For the parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0121] In the several implementations provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are only schematic, such as the division of units, which is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interfaces, and the indirect coupling or communication connection of the system or unit can be electrical or other forms.

[0122] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0123] The present application also provides a computer storage medium that can store multiple instructions, which are suitable for being loaded and executed by a processor as described above. Figure 1 The road assessment method of the embodiment shown in the figure can be found in the specific implementation process. Figure 1 The specific description of the illustrated embodiment will not be repeated here.

[0124] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0125] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, which is stored in a memory and includes several instructions for a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned memory includes: various media that can store program codes, such as USB flash drives, mobile hard drives, magnetic disks or optical disks.

[0126] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure.

[0127] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A UGC retrieval method based on semantic information, characterized in that: Methods include: Acquire UGC information data and user search terms, and convert the UGC information data into UGC semantic feature data; Perform multi-way recall retrieval according to the UGC semantic features and the user search terms to obtain a UGC candidate set related to the user search terms; Performing inverse sorting and merging of multiple recall result candidate sets in the UGC candidate set to obtain a sorted recall result candidate set; The sorted recall result candidate set is reranked through the BGE rerank model to obtain the rearranged recall result candidate set and the reranking score of each UGC candidate item in the rearranged recall result candidate set, and the reranked recall result candidate set and the reranking score of each UGC candidate item in the rearranged recall result candidate set are used as the UGC retrieval result.

2. The method according to claim 1, characterized in that The obtaining of UGC information data includes: Extract UGC basic information data from the relational database; Extract user and UGC interaction behavior data from the big data platform; The UGC basic information data is associated and matched with the UGC interactive behavior data to form the UGC information data.

3. The method according to claim 1, characterized in that The performing multi-channel recall retrieval according to the UGC semantic features and the user search terms to obtain a UGC candidate set related to the user search terms includes: Based on the BM25 algorithm, similarity calculation is performed on the text content of the user search term and the text content in the UGC semantic feature data to obtain a first recall result candidate set; Based on the embedding semantic similarity, similarity is calculated between the semantic feature vector of the user search term and the semantic feature vector in the UGC semantic feature data to obtain a second recall result candidate set; The first recall result candidate set and the second recall result candidate set are merged to obtain the UGC candidate set.

4. The method according to claim 1, characterized in that: The step of performing inverse sorting and fusion on the multiple recall result candidate sets in the UGC candidate set to obtain a sorted recall result candidate set includes: independently sorting each recall result candidate set in the UGC candidate set to obtain a sorted list of each recall result candidate set, wherein each candidate UGC item in the recall result candidate set corresponds to a sorted position; For each of the candidate UGC items, calculate the reciprocal ranking position value of the candidate UGC item in the ranking list of each recall result candidate set, where the reciprocal ranking position value is equal to the difference between the ranking position of the candidate UGC item and the length of the corresponding ranking list; For each of the candidate UGC items, based on the inverse ranking position value of the candidate UGC item in the sorted list of each recall result candidate set, a comprehensive inverse ranking value of each of the candidate UGC items is calculated, and each of the candidate UGC items is sorted based on the comprehensive inverse ranking value to obtain the sorted recall result candidate set.

5. The method according to claim 1, characterized in that The sorted recall result candidate set is re-ranked by using a preset BGE rerank model to obtain a re-ranked recall result candidate set and a re-ranking score of each UGC candidate item in the re-ranked recall result candidate set, including: Performing feature extraction on the candidate UGC items in the sorted recall result candidate set to obtain a multi-dimensional feature vector of the candidate UGC items; The multi-dimensional feature vector is input into the preset BGE rerank model to obtain a reranking score of each UGC candidate item, and the sorted recall result candidate set is sorted according to the reranking score of each UGC candidate item to obtain the reranked recall result candidate set.

6. The method according to claim 5, characterized in that Before inputting the multi-dimensional feature vector into the preset BGE rerank model to obtain the reranking score of each UGC candidate item, the method further includes: Acquire business scenario data and a reordering training data set, and define business reordering rules according to the business scenario data; quantifying the service reordering rule into a computable service rule indicator, and processing the reordering training data set according to the service reordering rule to generate corresponding service reordering features; Based on the original BGE rerank model and the service reranking characteristics, an initial BGE rerank model is constructed, and the initial BGE rerank model is trained based on the reranking training data set. When the training result reaches a preset accuracy value, the preset BGE rerank model is obtained.

7. The method according to claim 6, characterized in that The processing of the reordering training data set according to the service reordering rule to generate corresponding service reordering features includes: Feature data is extracted from the re-arrangement training data set to obtain re-arrangement training feature data, and the re-arrangement training feature data is merged to obtain the service re-arrangement feature.

8. A UGC retrieval system based on semantic information, characterized in that: The system comprises: A data processing module, used to obtain UGC information data and user search terms, and convert the UGC information data into UGC semantic feature data; A recall module, used to perform multi-way recall retrieval according to the UGC semantic features and the user search terms to obtain a UGC candidate set related to the user search terms; A sorting module, used to perform inverse sorting and fusion of multiple recall result candidate sets in the UGC candidate set to obtain a sorted recall result candidate set; A rearrangement module is used to rearrange the sorted recall result candidate set through the BGE rerank model to obtain the rearranged recall result candidate set and the rearrangement score of each UGC candidate item in the rearranged recall result candidate set, and use the rearranged recall result candidate set and the rearrangement score of each UGC candidate item in the rearranged recall result candidate set as the UGC retrieval result.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: It includes a processor, a memory and a transceiver, the memory is used to store instructions, the transceiver is used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device executes the method according to any one of claims 1 to 7.

Images