A method, system, electronic device and storage medium for recommending live broadcast material

Abstract

A method, system, electronic device and storage medium for recommending live broadcast materials, relating to the field of live broadcast. In the method, the host's voice is obtained, the voice is converted into text, and the text is segmented to obtain multiple first keywords to form a keyword set; the multiple first keywords are scored according to the similarity score with the live broadcast theme, the word frequency score and the score of matching with preset hot words, and the first preset number of the first keywords with the highest score are selected as the second keywords; according to the second keyword, the material library is searched to obtain the material information to be selected; according to the context of the live broadcast and the live broadcast theme, the material with the highest matching degree is selected from the material information to be selected for recommendation. The implementation of the technical solution provided by the present application improves the accuracy and timeliness of live broadcast material recommendation.

Description

Technical Field

[0001] The present application relates to the technical field of live broadcasting, and in particular to a method, system, electronic device and storage medium for recommending live broadcasting materials. Background Art

[0002] With the rapid development of the live broadcast industry, more and more content creators and businesses are beginning to use live broadcast platforms for product promotion, brand publicity and interactive communication. In order to improve the quality of live broadcasts and audience participation, live broadcast assistants are often needed to support the various needs of anchors, such as real-time interaction, content management, data analysis, etc. These functions can not only improve the effect of live broadcasts, but also help anchors better understand the needs of the audience, thereby realizing more accurate marketing strategies.

[0003] In existing live broadcast assistants, common material recommendation methods mainly rely on manual selection or simple keyword matching. Specifically, a common way is for the anchor to prepare some materials in advance, and then manually select suitable materials for display during the live broadcast. Another way is to use simple keyword matching to compare the pre-set keywords with the materials in the material library, and select materials with higher matching degrees for recommendation. However, manual selection of materials is time-consuming and labor-intensive, and it is easy to cause material selection to be untimely and inaccurate. Simple keyword matching methods lack intelligence and it is difficult to fully consider the theme and contextual information of the live broadcast, which may cause the recommended materials to be inconsistent with the live broadcast content, affecting the audience's viewing experience.

[0004] Therefore, how to recommend live broadcast materials more intelligently has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The present application provides a method, system, electronic device and storage medium for recommending live broadcast materials. By capturing the host's voice content in real time, accurately extracting keywords, efficiently screening materials and flexibly recommending matching materials, accurate recommendation of live broadcast materials is achieved, and the accuracy, efficiency, interactivity and adaptability of live broadcast are improved, thereby optimizing the audience's viewing experience and the host's live broadcast effect.

[0006] In a first aspect of the present application, a method for recommending live broadcast materials is provided, characterized in that the method is applied to a live broadcast assistant, and comprises:

[0007] Acquire the host's voice, convert the voice into text, and perform word segmentation on the text to obtain a plurality of first keywords to form a keyword set;

[0008] Scoring the plurality of first keywords according to the similarity score with the live broadcast theme, the word frequency score, and the score matching the preset hot words, and selecting a first preset number of the first keywords with the highest scores as the second keywords;

[0009] Searching the material library according to the second keyword to obtain information of the material to be selected;

[0010] According to the live broadcast context and the live broadcast theme, the material with the highest matching degree is selected from the to-be-selected material information for recommendation.

[0011] Optionally, scoring the first keywords according to the similarity score with the live broadcast theme, the word frequency score and the matching score with the preset hot words includes:

[0012] Performing word segmentation processing on the live broadcast topic to obtain topic segmentations, determining an intersection and a union between the first keyword and the topic segmentations, and calculating a first ratio of the intersection and the union as a similarity score of the first keyword;

[0013] Calculating the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, assigning an importance factor to the first keyword according to semantics, position and context, and calculating the word frequency score of the first keyword according to the word frequency, inverse document frequency and importance factor;

[0014] Searching for a hot word matching the first keyword in a preset knowledge base, and calculating a matching score of the first keyword according to the matching result;

[0015] A weighted sum is performed on the similarity score, the word frequency score and the matching score to obtain a score for the first keyword.

[0016] Optionally, the calculating the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, and assigning an importance factor to the first keyword according to semantics, position, and context includes:

[0017] Traversing the text to determine the first number of times the first keyword appears in the text, calculating a second ratio of the first number of times to the second number of times as the word frequency of the first keyword, wherein the second number of times is the total number of times all keywords in the keyword set appear in the text;

[0018] Traversing a preset document set to determine the number of first documents containing the first keyword, and calculating a third ratio of the first document number to the second document number as the document frequency of the first keyword, where the second document number is the total number of all documents in the preset document set;

[0019] Calculating the inverse document frequency according to the document frequency by a preset formula;

[0020] The semantic importance is determined according to the semantic features of the first keyword, the positional importance is determined according to the position of the first keyword in the text, the contextual importance is determined according to the association between the first keyword and other keywords, and the importance factor is determined according to the semantic importance, the positional importance and the contextual importance, wherein the other keywords are keywords in the keyword set other than the first keyword.

[0021] Optionally, searching in the material library according to the second keyword to obtain the information of the material to be selected includes:

[0022] Applying a Bloom filter to filter a data storage structure, removing a column of the data storage structure that does not contain the second keyword to obtain a first target column, the data storage structure comprising a common column and an uncommon column, the common column comprising a field whose search times are greater than a threshold, and the uncommon column comprising a field whose search times are less than or equal to the threshold;

[0023] Retrieving a list of record IDs containing the second keyword on the index node of the first target column by using an inverted index;

[0024] Merging search results from different index nodes, merging all record ID lists to obtain a merged list, and sorting the search results according to relevance;

[0025] The material information to be selected is determined according to the merged list.

[0026] Optionally, the applying a Bloom filter to filter the data storage structure to remove columns of the data storage structure that do not contain the second keyword to obtain the first target column includes:

[0027] creates a bit vector and initializes all bits of said bit vector to zero,

[0028] Traversing the data storage structure, for a second target column, using the same second preset number of hash functions to perform hash calculation on the identifier or content of the second target column, and checking whether the corresponding second preset number of bits in the bit vector are all ones;

[0029] If all are ones, the second target column is retained as part of the target column;

[0030] All the columns reserved as target columns are merged to form the first target column.

[0031] Optionally, merging the search results from different index nodes, merging all record ID lists to obtain a merged list, and sorting the search results according to relevance includes:

[0032] During the merging process, duplicate record IDs are merged based on their occurrence frequency and weight on different index nodes, so as to retain the duplicate record IDs with the highest occurrence frequency and weight.

[0033] Calculate the relevance score of the target record ID according to the material content corresponding to the target record ID, the relevance between the material and the live broadcast theme, and the popularity of the material, where the target record ID is any record ID in the merged list;

[0034] The record IDs in the merged list are sorted in descending order according to the relevance scores.

[0035] Optionally, selecting the material with the highest matching degree from the to-be-selected material information for recommendation according to the live broadcast context and the live broadcast theme includes:

[0036] Analyze the context information of the live broadcast and extract target information points in the context information, wherein the target information points include people, places and events;

[0037] Combining the target information point with the live broadcast theme to form a comprehensive query condition, and filtering and selecting the to-be-selected material information according to the comprehensive query condition, and selecting the to-be-selected material that meets both the live broadcast theme and the target information point;

[0038] Weights are set for timeliness, quality, and user feedback, respectively, and comprehensive scores of the candidate materials are calculated according to the weights. The selected materials are sorted in descending order according to the comprehensive scores, and the candidate materials with the highest comprehensive scores are pushed as recommended materials.

[0039] In a second aspect of the present application, a recommendation system for live broadcast materials is provided, including a text module, a rating module, a retrieval module and a recommendation module, wherein:

[0040] A text module is configured to obtain the host's voice, convert the voice into text, and perform word segmentation on the text to obtain a plurality of first keywords to form a keyword set;

[0041] A scoring module configured to score the plurality of first keywords according to a similarity score with a live broadcast theme, a word frequency score, and a score matching a preset hot word, and select a first preset number of the first keywords with the highest scores as second keywords;

[0042] A retrieval module, configured to search the material library according to the second keyword to obtain the material information to be selected;

[0043] The recommendation module is configured to select the material with the highest matching degree from the material information to be selected for recommendation according to the live broadcast context and the live broadcast theme.

[0044] In the third 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 both 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 any one of the methods described above.

[0045] In a fourth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores instructions, and when the instructions are executed, any of the methods described above is executed.

[0046] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0047] 1. By converting the host's voice into text and performing word segmentation, the real-time content of the host in the live broadcast can be accurately captured, thereby extracting keywords closely related to the live broadcast content. By scoring the keywords, the keyword with the highest score is selected as the second keyword, and the materials most relevant to the live broadcast theme are further screened, thereby improving the accuracy of the recommendation;

[0048] 2. It can select the most matching material from the material library for recommendation based on the context and theme of the live broadcast, ensuring that the recommended material is closely related to the live broadcast content, thereby enhancing the audience's viewing experience. Real-time and accurate material recommendation can maintain the coherence and attractiveness of the live broadcast and reduce the possibility of audience loss due to irrelevant content;

[0049] 3. The anchor does not need to manually search for suitable materials among a large number of materials. The live broadcast assistant can automatically complete this process, which greatly saves the anchor's time and energy. The efficient material recommendation mechanism allows the anchor to focus more on the creation of live content and interaction with the audience, thereby improving the overall quality of the live broadcast. By capturing the anchor's voice content in real time and recommending relevant materials, the interactivity and real-time nature of the live broadcast can be enhanced. The audience can feel the anchor's precise control of the live content and timely response to the audience's needs, thereby enhancing the audience's sense of participation and loyalty;

[0050] 4. It can flexibly recommend materials based on the themes and contexts of different live broadcasts. It is suitable for live broadcast scenarios of various types and styles. Through mechanisms such as preset high-hot word matching scores, it can meet the live broadcast needs of different fields and topics, and improve the adaptability and breadth of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flowchart of a method for recommending live broadcast materials disclosed in an embodiment of the present application;

[0052] Figure 2 It is a schematic diagram of the overall platform planning disclosed in the embodiment of the present application;

[0053] Figure 3 It is a schematic diagram of the architecture of the live broadcast material management system disclosed in the embodiment of the present application;

[0054] Figure 4 It is a flowchart of the live broadcast assistant system disclosed in the embodiment of the present application;

[0055] Figure 5 It is a functional schematic diagram of the live broadcast assistant system disclosed in the embodiment of the present application;

[0056] Figure 6 It is a module diagram of a recommendation system for live broadcast materials disclosed in an embodiment of the present application;

[0057] Figure 7 It is a structural schematic diagram of an electronic device disclosed in an embodiment of the present application.

[0058] Explanation of the reference numerals: 601, text module; 602, scoring module; 603, retrieval module; 604, recommendation module; 701, processor; 702, communication bus; 703, user interface; 704, network interface; 705, memory. DETAILED DESCRIPTION

[0059] 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.

[0060] 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.

[0061] 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.

[0062] This embodiment discloses a method for recommending live broadcast materials, which is applied to a live broadcast assistant system. Figure 1 is a flow chart of a method for recommending live broadcast materials disclosed in an embodiment of the present application, such as Figure 1 As shown, the method comprises the following steps:

[0063] S110, acquiring the host's voice, converting the voice into text, and performing word segmentation on the text to obtain a plurality of first keywords to form a keyword set;

[0064] Figure 2 is a schematic diagram of the overall platform planning disclosed in the embodiment of this application, such as Figure 2 As shown in the figure, the overall platform includes self-developed parts and third-party interface parts. The self-developed parts include: live broadcast assistant, material management, operation management, KOL (Key Opinion Leader) management, data statistics, and account authority management. The third-party interface parts include: Baidu Qianfan, Baidu AI (Artificial Intelligence), Alibaba Cloud API (Application Programming Interface), and Alibaba Cloud CDN (Content Delivery Network). Figure 3 Schematic diagram of the architecture of the live broadcast material management system disclosed in the embodiment of the present application, such as Figure 3 As shown, the live broadcast material management system includes car series data, vehicle model data, material upload, vehicle model comparison, material approval, image acceleration, usage statistics and log query.

[0065] Figure 4 is a flowchart of the live broadcast assistant system disclosed in the embodiment of the present application, such as Figure 4As shown, the editor logs in to the material management system, creates vehicle model data, and uploads vehicle model materials. The operation staff reviews the materials, publishes the vehicle model data after the review, accelerates images or videos through Alibaba CDN, and KOLs screen, search, compare, and download vehicle model data. The management staff records and counts KOLs' use of materials and optimizes material sources based on material usage analysis.

[0066] Figure 5 is a functional schematic diagram of the live broadcast assistant system disclosed in the embodiment of the present application, such as Figure 5 As shown, the live broadcast assistant system includes six major functions: statistical reports, permission management, system management, material management, anchor management and live broadcast management.

[0067] The following combination Figure 2-Figure 5 The embodiments of the present application are described.

[0068] The live voice of the host is captured from a live stream or recording device. This is usually done using an audio capture device (such as a microphone) and corresponding audio processing software or hardware. After the host's voice is captured, the voice is converted to text. This is usually done through speech recognition technology, which can analyze audio signals and convert them into corresponding text representations. The accuracy of speech recognition technology is crucial to the success of subsequent steps because it directly affects the accuracy and relevance of keywords extracted from the text. After the voice is converted to text, the text is segmented. Word segmentation is the process of splitting a continuous text string into independent words or phrases, which is crucial for subsequent keyword extraction. In Chinese, since there is no obvious space between words, the word segmentation process is usually implemented with the help of natural language processing (NLP) technology. After the word segmentation process is completed, keywords are extracted from the text. These keywords are usually the most important or most frequently occurring words or phrases in the text, which can summarize the main content of the text. By extracting keywords, it is easier to understand the theme and main points of the text, thereby providing useful information for subsequent content analysis, recommendations or other operations. The extracted keywords are collected into a keyword set for further processing and analysis.

[0069] S120, scoring the plurality of first keywords according to the similarity score with the live broadcast theme, the word frequency score, and the score of matching with the preset hot words, and selecting a first preset number of the first keywords with the highest scores as the second keywords;

[0070] Similarity score: Using the Jaccard similarity method, calculate the ratio of the intersection and union between the keywords extracted from the speech text and the live broadcast theme and script segmentation. This ratio reflects the similarity between the keywords and the live broadcast theme. The keyword frequency of the entire live broadcast is the number of times each keyword appears in the live broadcast text. Frequency score: Referring to the TF-IDF (term frequency-inverse document frequency) method, calculate the frequency score of each keyword and multiply it by an importance factor (Importance_Factor_w) to reflect the relative importance of the keyword in the live broadcast text. The high-heat word library in the knowledge base is a preset word library that contains keywords that frequently appear in related fields or topics and have high heat. High-heat word matching score: Check whether each keyword matches the high-heat word in the knowledge base, and give a corresponding score based on the degree of matching. Use a scoring formula (Total_Score=M*A +N*B+P*C) to combine the above three scores to get the final score for each keyword. Among them, A represents the similarity score, B represents the word frequency score, C represents the high-heat word matching score, M, N, and P are the weights of each part, and their sum is equal to 1. After obtaining the scores of all keywords, they are sorted in order from high to low, and then the first preset number (for example, 5-8) of keywords with the highest scores are selected as the second keywords. For example, the voice text is "It adopts a double horizontal line daytime running light and a hidden side turn signal design; and the headlight group is also full of "black technology"! It adopts a split design, and the high and low beams and fog lights are integrated together. It is worth mentioning that the headlights also have special features such as adaptive high and low beams, front fog lights, automatic headlights and geometric multi-beams." The system calculates and obtains a list of keywords from high to low scores, such as "headlights", "headlights", "daytime running lights", "side turn signals", etc. These keywords reflect the important content and theme focus in the live text. The selected second keyword can be used to retrieve relevant pictures, videos or text materials in the material library to support the display and interpretation of the live content. These second keywords can also be used to analyze the theme, focus and audience interests of the live broadcast content, providing data support for subsequent live broadcast planning and marketing.

[0071] Optionally, scoring the first keywords according to the similarity score with the live broadcast theme, the word frequency score and the matching score with the preset hot words includes:

[0072] Performing word segmentation processing on the live broadcast topic to obtain topic segmentations, determining an intersection and a union between the first keyword and the topic segmentations, and calculating a first ratio of the intersection and the union as a similarity score of the first keyword;

[0073] Calculating the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, assigning an importance factor to the first keyword according to semantics, position and context, and calculating the word frequency score of the first keyword according to the word frequency, inverse document frequency and importance factor;

[0074] Searching for a hot word matching the first keyword in a preset knowledge base, and calculating a matching score of the first keyword according to the matching result;

[0075] A weighted sum is performed on the similarity score, the word frequency score and the matching score to obtain a score for the first keyword.

[0076] The topic of the live broadcast is segmented to obtain a series of topic segmentations. These segmentations represent the core vocabulary of the live broadcast topic. Determine the intersection (i.e., vocabulary that appears in both the keyword and the topic segmentation) and the union (i.e., all non-repeated vocabulary in the keyword and the topic segmentation) between each first keyword and the topic segmentation. Calculate the first ratio of the intersection and the union, which reflects the similarity between the keyword and the live broadcast topic. The higher the ratio, the stronger the correlation between the keyword and the live broadcast topic. This ratio is used as the similarity score of the keyword. For each first keyword, calculate its word frequency (i.e., the number of times it appears) in the live broadcast text and its document frequency (i.e., the number of documents containing the keyword) in the entire document collection. Calculate the inverse document frequency (IDF) based on the document frequency, which reflects the prevalence of the word. The more documents a word appears in, the lower its inverse document frequency, and vice versa. Assign an importance factor to each keyword based on semantics, position, and context information. This factor takes into account the position of the word in a sentence or paragraph (such as the beginning, end, or middle), the degree of correlation with other words, and whether it contains specific semantic information. Calculate the word frequency score for each keyword based on word frequency, inverse document frequency, and importance factor. This score takes into account the frequency, prevalence, and importance of the keyword in the text. Search for high-hot words that match each first keyword in the preset knowledge base. These high-hot words represent frequently appearing and high-hot words in related fields or topics. Calculate the matching score for each keyword based on the matching results. The matching score can be determined based on the number of matched high-hot words, the degree of matching (such as full match, partial match, or synonym match), and the weight of the high-hot words. The similarity score, word frequency score, and matching score are weighted and summed according to the predetermined weights to obtain the final score for each first keyword. These weights can be adjusted according to actual needs to reflect the relative importance of different scores in the final score.

[0077] By segmenting the live broadcast topic and calculating the intersection and union with the first keyword, the correlation between the keyword and the live broadcast topic can be accurately measured. This calculation method based on text similarity effectively improves the accuracy of keyword scoring. Not only the word frequency of the keyword in the live broadcast text is considered, but also the inverse document frequency (IDF) and importance factor are combined to more comprehensively evaluate the weight of the keyword. The inverse document frequency reflects the prevalence of the keyword in the document collection, while the importance factor can be dynamically adjusted according to the semantics, position and context of the keyword, further improving the comprehensiveness of the scoring. Searching for high-hot words matching the first keyword in the preset knowledge base and calculating the matching score based on the matching results can capture the popularity of the keyword in related fields or topics, thereby increasing the reference value of the scoring. By scoring multiple first keywords and selecting the keyword with the highest score as the basis for subsequent operations, the recommendation and search of live broadcast content can be optimized. For example, when retrieving materials related to the live broadcast topic in the material library, keywords with higher scores can be given priority, thereby improving the accuracy and efficiency of the retrieval. Combined with the user's historical viewing history, interest preferences and other information, the keyword score can be further personalized to achieve more accurate personalized recommendations. During the live broadcast, the host's voice content can be analyzed in real time, and keywords can be extracted for scoring. This helps the host to understand the audience's focus and interests in a timely manner, so as to adjust the live broadcast content and rhythm, and improve the interactivity and attractiveness of the live broadcast.

[0078] Optionally, the calculating the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, and assigning an importance factor to the first keyword according to semantics, position, and context includes:

[0079] Traversing the text to determine the first number of times the first keyword appears in the text, calculating a second ratio of the first number of times to the second number of times as the word frequency of the first keyword, wherein the second number of times is the total number of times all keywords in the keyword set appear in the text;

[0080] Traversing a preset document set to determine the number of first documents containing the first keyword, and calculating a third ratio of the first document number to the second document number as the document frequency of the first keyword, where the second document number is the total number of all documents in the preset document set;

[0081] Calculating the inverse document frequency according to the document frequency by a preset formula;

[0082] The semantic importance is determined according to the semantic features of the first keyword, the positional importance is determined according to the position of the first keyword in the text, the contextual importance is determined according to the association between the first keyword and other keywords, and the importance factor is determined according to the semantic importance, the positional importance and the contextual importance, wherein the other keywords are keywords in the keyword set other than the first keyword.

[0083] Traverse the text to determine the number of times the first keyword appears in the text, which is called the first count. Calculate the total number of times all keywords in the keyword set (i.e., the set of all keywords in the text) appear in the text, which is called the second count. Calculate the ratio of the first count to the second count, which is the word frequency of the first keyword. Word frequency reflects the relative importance of a keyword in the text, that is, the more frequently a keyword appears, the greater its importance may be. Traverse a preset document set (i.e., a set containing multiple documents) to determine the number of documents containing the first keyword, which is called the first document number. Calculate the total number of all documents in the preset document set, which is called the second document number. Calculate the ratio of the first document number to the second document number, which is the document frequency of the first keyword. Document frequency reflects the prevalence of a keyword in a document set, that is, the more documents a keyword appears in, the higher its prevalence may be. Use a preset formula (usually IDF=log(total number of documents / number of documents containing the word), where IDF stands for inverse document frequency) to calculate the inverse document frequency based on the document frequency. The inverse document frequency reflects the rarity of a keyword, that is, the fewer documents a keyword appears in, the higher its rarity may be, and thus its importance may be greater. The semantic importance of the first keyword is determined based on its semantic features (such as part of speech, meaning, etc.). For example, content words such as nouns and verbs may be more important than function words such as conjunctions and prepositions. The positional importance of the first keyword is determined based on its position in the text (such as the title, the beginning of a paragraph, the end, etc.). Generally, keywords that appear in the title or the beginning of a paragraph may be more important. The contextual importance of the first keyword is determined based on its correlation with other keywords (such as co-occurrence relationship, semantic relationship, etc.). Keywords with high correlation with other keywords may be more important. The importance factor of the first keyword is determined by combining semantic importance, positional importance, and contextual importance. This factor is a comprehensive indicator used to reflect the overall importance of a keyword in a text.

[0084] By traversing the text, the number of times the first keyword appears in the text is determined, and the ratio of the number of times the first keyword appears in the text to the total number of times all the keywords in the keyword set appear in the text is calculated, that is, the word frequency. This step can accurately measure the frequency of occurrence of keywords in a specific text and is the basis for evaluating the weight of keywords. The number of documents containing the first keyword is determined by traversing the preset document set, and the ratio of the number of documents containing the first keyword to the total number of document sets is calculated, that is, the document frequency. Further, the inverse document frequency is calculated according to the document frequency by a preset formula, and the inverse document frequency reflects the prevalence of the keyword in the document set. These two indicators together constitute an important basis for evaluating the importance of keywords. According to the semantic features of the first keyword, its position in the text, and its correlation with other keywords, the semantic importance, position importance, and context importance are determined respectively. These three dimensions together constitute the comprehensive importance of the keyword in a specific context. By comprehensively considering these dimensions, the weight and influence of the keyword can be more comprehensively evaluated. After calculating the word frequency, document frequency, inverse document frequency, and importance factor of the first keyword, it can be used as the basis for subsequent text processing and information retrieval. For example, in tasks such as text classification, information extraction, and sentiment analysis, keywords with higher weights can be given priority to improve processing accuracy and efficiency.

[0085] S130, searching in the material library according to the second keyword to obtain information of the material to be selected;

[0086] The data in the material library is divided into "common columns" and "uncommon columns". Common columns contain frequently searched fields such as brand, model, product name and label, which are of special interest due to their high search frequency. Uncommon columns contain less frequently searched fields such as detailed description and feature details. The database uses sharding technology, which means that the data is stored in multiple nodes in a dispersed manner, and each node is responsible for storing a part of the database data. This design improves the efficiency and scalability of data access. For the text data in the common columns, the system has established an inverted index. An inverted index is a special data structure that maps each word or phrase in the text to a list of documents (i.e., record IDs) containing the word or phrase. When searching based on the second keyword, the system sends the query to the nodes containing the common column index. These nodes use the inverted index to quickly find the list of record IDs containing the keyword. Since the data is stored in shards, the query results may come from multiple nodes. The system needs to merge the query results from these different nodes. The merging process mainly summarizes the list of record IDs returned by each node and sorts the results according to relevance or other sorting criteria. In this way, a list of material information to be selected sorted by relevance can be obtained.

[0087] Optionally, searching in the material library according to the second keyword to obtain the information of the material to be selected includes:

[0088] Applying a Bloom filter to filter a data storage structure, removing a column of the data storage structure that does not contain the second keyword to obtain a first target column, the data storage structure comprising a common column and an uncommon column, the common column comprising a field whose search times are greater than a threshold, and the uncommon column comprising a field whose search times are less than or equal to the threshold;

[0089] Retrieving a list of record IDs containing the second keyword on the index node of the first target column by using an inverted index;

[0090] Merging search results from different index nodes, merging all record ID lists to obtain a merged list, and sorting the search results according to relevance;

[0091] The material information to be selected is determined according to the merged list.

[0092] In order to improve the retrieval efficiency and reduce unnecessary calculations, the system first applies a Bloom filter to filter the data storage structure. The Bloom filter maps elements to different positions of the bitmap through multiple hash functions and sets the values ​​at these positions to 1. During the retrieval process, the system performs the same hash function calculation on the second keyword and checks whether the values ​​at the corresponding positions are all 1. If the values ​​at all corresponding positions are 1, the second keyword may exist in some columns of the data storage structure (but there is a false positive rate); if the value at any position is not 1, it can be determined that the second keyword does not exist in these columns. In this way, the system can remove the columns that do not contain the second keyword and obtain the first target column. In this scenario, the Bloom filter is mainly used to quickly determine whether the second keyword may exist in certain fields in the common or uncommon columns. Due to the characteristics of the Bloom filter (i.e., there may be false positives, but there are no missed positives), it is suitable for fast screening in large-scale data sets. Based on the first target column, the system uses the inverted index to quickly find the list of record IDs containing the second keyword. The inverted index is a data structure that maps each word or phrase in the text to a list of documents (i.e., record IDs) containing the word or phrase. The system performs a search operation on the index node of the first target column and uses the inverted index to quickly locate the record ID containing the second keyword. Since the data storage structure uses sharding technology, the search results may come from multiple index nodes. The system needs to merge these results into a unified list. The system merges all the record ID lists from different index nodes to obtain a merged list. Then, the merged list is sorted according to relevance or other sorting criteria (such as timestamp, score, etc.) to obtain a merged list sorted by relevance. The merged list contains the record ID containing the second keyword. Based on the record ID in the merged list, the system extracts the corresponding material information (such as brand, model, product name, label, etc.) from the data storage structure as the material information to be selected.

[0093] By filtering the data storage structure through the Bloom filter, the columns that do not contain the second keyword can be quickly removed, thereby greatly reducing the scope of subsequent retrieval. This filtering mechanism not only improves the retrieval efficiency, but also reduces unnecessary computing resource consumption. The data storage structure is divided into common columns and uncommon columns, and thresholds are set based on the number of searches to distinguish them, so that the system can more accurately locate high-frequency search fields. This design optimizes the data storage and retrieval strategy and improves the overall performance of the system. An inverted index is established on the common columns to achieve a fast mapping of keywords to record IDs. This index structure enables the system to find all records containing the second keyword in a short time, improving the accuracy and efficiency of retrieval. Using the distributed system architecture, the inverted index is distributed on multiple index nodes. This design not only improves the scalability of the system, but also enables the retrieval process to be processed in parallel, further shortening the retrieval time. The retrieval results from different index nodes are merged to obtain a complete list of record IDs. This step ensures that all records containing the second keyword can be retrieved, avoiding the omission of information. The retrieval results are sorted according to relevance, so that the material information that best meets user needs can be displayed first. This sorting mechanism improves user satisfaction and efficiency in retrieval.

[0094] Optionally, the applying a Bloom filter to filter the data storage structure to remove columns of the data storage structure that do not contain the second keyword to obtain the first target column includes:

[0095] creates a bit vector and initializes all bits of said bit vector to zero,

[0096] Traversing the data storage structure, for a second target column, using the same second preset number of hash functions to perform hash calculation on the identifier or content of the second target column, and checking whether the corresponding second preset number of bits in the bit vector are all ones;

[0097] If all are ones, the second target column is retained as part of the target column;

[0098] All the columns reserved as target columns are merged to form the first target column.

[0099] Create a large enough bit vector and initialize all bits to zero. This bit vector will be used as the core data structure of the Bloom filter to store the hash value of the keyword. Traverse each column (second target column) in the data storage structure. These columns may contain fields such as brand, model, product name, etc., which are the data sources to be retrieved. For each second target column, use a preset number of hash functions (here is the second preset number) to hash its identifier or content. These hash functions map the content of the column to different positions of the bit vector. Check whether the values ​​at these corresponding positions in the bit vector are all one. If all are one, it means that the column may contain the keyword to be retrieved (that is, there is a possibility of misjudgment, but we will temporarily retain it at this stage). If a second target column passes the hash calculation and check, that is, the bit vector values ​​at its corresponding positions are all one, then the column is retained as part of the target column. After the traversal is completed, all retained columns are merged to form a new column set, namely the first target column. This set contains all columns that may contain the second keyword.

[0100] By creating a bit vector and initializing all bits to zero, a basis is provided for subsequent hash calculation and bitmap marking. This step ensures the clarity and consistency of the data structure and facilitates subsequent operations. The system traverses the data storage structure and performs hash calculation and bit vector check on each column. This traversal method ensures the comprehensiveness and integrity of the data and avoids missing any columns that may contain the second keyword. The same second preset number of hash functions are used to perform hash calculation on the identifier or content of the second target column, and check whether the corresponding bits in the bit vector are all one. This step utilizes the principle of Bloom filter, and maps elements to different positions of the bitmap through multiple hash functions, thereby realizing the function of quickly judging whether the element may exist in the set. If the corresponding bits in the bit vector are all one, the second target column is retained as part of the target column. This step ensures that only columns that may contain the second keyword are retained, thereby narrowing the scope of subsequent retrieval. Finally, all retained columns are merged to form the first target column, which provides a basis for subsequent inverted index retrieval. Through the filtering mechanism of Bloom filter, the system can quickly remove columns that do not contain the second keyword, thereby reducing the amount of data that needs to be processed in the subsequent retrieval process. This step not only improves the retrieval efficiency, but also reduces the resource consumption of the system.

[0101] Optionally, merging the search results from different index nodes, merging all record ID lists to obtain a merged list, and sorting the search results according to relevance includes:

[0102] During the merging process, duplicate record IDs are merged based on their occurrence frequency and weight on different index nodes, so as to retain the duplicate record IDs with the highest occurrence frequency and weight.

[0103] Calculate the relevance score of the target record ID according to the material content corresponding to the target record ID, the relevance between the material and the live broadcast theme, and the popularity of the material, where the target record ID is any record ID in the merged list;

[0104] The record IDs in the merged list are sorted in descending order according to the relevance scores.

[0105] Duplicate record IDs are merged based on their frequency of occurrence and weight on different index nodes. Specifically, if a record ID is retrieved on multiple nodes, the system will count the frequency of occurrence of the record ID and may adjust its final frequency of occurrence based on the weight of the node (for example, the weight set based on factors such as the node's processing power, data quality, etc.). When processing duplicate record IDs, the system will retain the record ID with the highest frequency of occurrence and weight to ensure that the record IDs in the merged list are unique and important. After the above processing, the system will generate a merged list that contains all record IDs retrieved during the retrieval process (duplicates are removed, and the record IDs with the highest frequency of occurrence and weight are retained). In order to further improve the accuracy and practicality of the retrieval results, the system will calculate its relevance score based on the content of the material corresponding to the target record ID, the correlation between the material and the live broadcast theme, and the popularity of the material. The specific calculation method is as follows: the system will calculate the relevance score of the material content based on the degree of match between the material content corresponding to the target record ID and the second keyword (i.e. the search term entered by the user). The higher the degree of match, the higher the score; the relevance between the material and the live broadcast theme is considered. The higher the relevance, the more suitable the material is for live broadcast, so the score will be higher; the relevance score will also be calculated based on the popularity of the material (for example, the number of clicks, shares, comments, etc.). The more popular the material, the more popular it is, so the score will be higher accordingly.

[0106] When merging the search results from different index nodes, the system will merge the duplicate record IDs according to their frequency and weight on different index nodes. This step ensures that only the most representative record IDs are retained, avoiding the interference of duplicate data. By comprehensively considering the frequency and weight of duplicate record IDs, the system can more accurately evaluate the importance and relevance of record IDs. This consideration method not only improves the accuracy of the merged results, but also optimizes the performance of the system. The system calculates the relevance score based on multiple dimensions such as the content of the material corresponding to the target record ID, the relevance of the material to the live broadcast theme, and the popularity of the material. This step ensures the objectivity and accuracy of the sorting results, so that the material that best meets the user's needs can be displayed first. By comprehensively considering multiple factors such as material content, relevance and popularity, the system can more comprehensively evaluate the relevance and value of the material. This multi-dimensional consideration method improves the diversity and practicality of the sorting results. The record IDs in the merged list are sorted in descending order according to the relevance score, so that the record ID with the highest relevance is ranked first. This step provides users with a more intuitive and convenient way to select materials, improving the efficiency and satisfaction of user retrieval. Through the merging and deduplication steps, the system can reduce the interference of redundant data and improve the clarity and readability of the search results. Through precise relevance calculation and sorting steps, the system can more accurately locate the materials needed by users and improve the accuracy and practicality of the search results.

[0107] S140: According to the live broadcast context and the live broadcast theme, select the material with the highest matching degree from the to-be-selected material information for recommendation.

[0108] Before the broadcast, the anchor or live broadcast platform needs to clarify the theme of the live broadcast, which is usually the core and focus of the live broadcast content. The live broadcast theme may involve a specific event, topic, festival, product, etc., and is an important reference for selecting materials. The live broadcast context includes real-time content of the live broadcast, audience interaction, anchor style and other aspects. By analyzing the live broadcast context, you can understand the current atmosphere of the live broadcast, the audience's interests and concerns, and the personality and style of the anchor. According to the material library collected and sorted in the early stage, filter out the information of the materials to be selected that are related to the live broadcast theme. These material information should include key information such as the type, content, source, and usage rights of the material. For each material to be selected, evaluate it according to the relevance of its content to the live broadcast theme and context. The evaluation can be based on a variety of technical means such as keyword matching, semantic analysis, and sentiment analysis. The evaluation result can be a quantitative score or grade, which is used to indicate the degree of match between the material and the live broadcast. Among all the materials to be selected, select the material with the highest match and the most consistent with the live broadcast theme and context for recommendation. If there are multiple highly matched materials, you can further consider the novelty, quality, audience preferences and other factors of the material for screening. The selected materials will be recommended and displayed during the live broadcast to attract the audience's attention and improve the live broadcast effect. The display methods can include scrolling playback, pop-up prompts, anchor explanations and other forms. When selecting materials, the legality and compliance of the materials should be ensured to avoid using infringing or inappropriate content. The recommendation and display of materials should respect the audience's experience and feelings, and avoid excessive interference or impact on the normal progress of the live broadcast. Anchors or live broadcast platforms should regularly update and optimize the material library to maintain the freshness and diversity of the materials.

[0109] Optionally, selecting the material with the highest matching degree from the to-be-selected material information for recommendation according to the live broadcast context and the live broadcast theme includes:

[0110] Analyze the context information of the live broadcast and extract target information points in the context information, wherein the target information points include people, places and events;

[0111] Combining the target information point with the live broadcast theme to form a comprehensive query condition, and filtering and selecting the to-be-selected material information according to the comprehensive query condition, and selecting the to-be-selected material that meets both the live broadcast theme and the target information point;

[0112] Weights are set for timeliness, quality, and user feedback, respectively, and comprehensive scores of the candidate materials are calculated according to the weights. The selected materials are sorted in descending order according to the comprehensive scores, and the candidate materials with the highest comprehensive scores are pushed as recommended materials.

[0113] The context information of live broadcast includes the real-time content of the live broadcast, the host's explanation, the audience's interaction, and the live broadcast environment. By analyzing this context information, key information points such as people, places, and events that are closely related to the live broadcast can be extracted. The target information points refer to key elements such as people, places, and events that are closely related to the live broadcast theme extracted from the live broadcast context. These information points will serve as an important basis for subsequent screening of candidate materials. The extracted target information points are combined with the live broadcast theme to form a set of comprehensive query conditions. These conditions will be used to filter and select the candidate material information to find the candidate materials that meet both the live broadcast theme and the target information points. According to the formed comprehensive query conditions, the candidate material information is compared and screened one by one. The candidate materials that meet both the live broadcast theme and the target information points are screened out as candidates for subsequent calculation of the comprehensive score. Weights are set for key indicators such as timeliness, quality, and user feedback. According to these weights and the performance of the candidate materials on the corresponding indicators, the comprehensive score of each candidate material is calculated. The comprehensive score reflects the degree of match between the candidate material and the live broadcast theme and context, as well as its performance in multiple dimensions. The selected materials are sorted in descending order according to the comprehensive scores. The selected materials with the highest comprehensive scores are pushed as recommended materials to attract the audience's attention and improve the live broadcast effect.

[0114] For example, in a live broadcast about cars, the host is introducing a newly launched SUV model. The contextual information of the live broadcast may include:

[0115] Characters: The host (there may also be an interactive session for guests or audience to ask questions);

[0116] Location: car showroom or outdoor test drive venue;

[0117] Events: Detailed introduction of new SUV models, test drive experience, performance comparison, etc.

[0118] During the live broadcast, the anchor may mention the specific configuration, price, appearance, interior, power performance and other aspects of the model. These are target information points and are closely related to the live broadcast theme. The theme of the live broadcast is "Introduction to new SUV models", so the extracted target information points (such as model configuration, price, appearance, etc.) can be combined with the live broadcast theme to form comprehensive query conditions. These conditions will be used to filter and screen the material information to be selected. For example, the following comprehensive query conditions can be set: the material must contain relevant information about the new SUV model; the material should match the target information points such as model configuration, price, appearance, etc. mentioned by the anchor; the material should have high quality and timeliness to attract the audience's attention. According to the comprehensive query conditions, the material information to be selected can be filtered and screened. These materials may come from official promotional materials of automobile brands, evaluation articles on professional automobile websites, live broadcast replays or short videos of other anchors, etc. In the screening process, the focus will be on those materials that meet both the live broadcast theme and the target information points. For example, a short video containing the detailed configuration and test drive experience of the new SUV model can be selected as the material to be selected, because it not only meets the live broadcast theme, but also contains multiple target information points mentioned by the anchor. In order to further improve the matching degree and recommendation effect of the materials, weights can be set for key indicators such as timeliness, quality, and user feedback. These weights can be adjusted according to the specific needs of the live broadcast and the preferences of the audience. For example, the following weights can be set: timeliness: 0.4 (because the introduction of new models needs to keep up with market trends, timeliness is very important); quality: 0.3 (high-quality materials can enhance the audience's viewing experience); user feedback: 0.3 (user feedback can reflect the popularity and practicality of the materials). The comprehensive score of each candidate material can be calculated based on these weights. The comprehensive score can be calculated by weighted average, that is, comprehensive score = timeliness score × timeliness weight + quality score × quality weight + user feedback score × user feedback weight. The selected materials can be sorted in descending order according to the comprehensive score, and the candidate material with the highest comprehensive score can be pushed as the recommended material. In this way, when watching the live broadcast, the audience can see materials that are highly matched with the live broadcast theme and target information points, high in quality and strong in timeliness, thereby improving the viewing experience and interactive effect of the live broadcast. For example, in the above example, if a short video containing the detailed configuration and test drive experience of a new SUV model is selected as the candidate material, and the material performs well in terms of timeliness, quality and user feedback, then it is likely to become the recommended material with the highest comprehensive score and be pushed to the audience for viewing.

[0119] By analyzing the contextual information of the live broadcast, key target information points such as people, places and events are extracted. These information points can accurately reflect the core content of the live broadcast and the focus of the audience's attention. The target information points are combined with the live broadcast theme to form a comprehensive query condition. This step ensures the close relevance and high matching of the selected material information with the live broadcast content. Through the filtering and screening mechanism, the materials that do not match the live broadcast theme and target information points can be quickly eliminated, and the most suitable materials can be retained, which improves the accuracy and efficiency of material selection. The weights such as timeliness, quality and user feedback are set. These weights reflect the importance of different dimensions in material evaluation. Timeliness ensures the freshness and timeliness of the material; quality reflects the production level and content value of the material; user feedback reflects the audience's preference and acceptance of the material. The comprehensive score of the selected material is calculated according to the weight. This step comprehensively considers the influence of multiple dimensions, making the evaluation result more comprehensive and objective. Through the descending sorting mechanism, the selected material with the highest comprehensive score is pushed as the recommended material. This step ensures the quality and attractiveness of the recommended material and improves the user experience. Combining artificial intelligence and big data analysis, it can realize intelligent processing and analysis of live broadcast context and information about materials to be selected. By continuously learning and optimizing the algorithm model, the technology can gradually adapt to user needs and preferences and realize more personalized material recommendations. This intelligent and personalized recommendation method not only improves the accuracy and efficiency of recommendations, but also enhances user participation and satisfaction.

[0120] The embodiment of the present application performs horizontal and vertical slicing of the storage of the original data, thereby avoiding the original direct retrieval from the massive data. Horizontal slicing test: simulate a database of 1 billion records, each record is 1KB in size. It is horizontally sliced ​​into 10 slicings, each slicing stores 100 million records.

[0121] ;

[0122] Vertical cutting test: A table with 100 columns, of which only 20 columns are frequently queried, is cut vertically into two tables, one containing the 20 frequently queried columns and the other containing the remaining 80 columns.

[0123] ;

[0124] Scene prediction logic: The analysis found that one of the characteristics of retrieval is that the number of keywords has a direct impact on the retrieval speed. The host's speech during the live broadcast is monitored in real time, and then the speech is converted into text and word segmentation is performed. When the host's material retrieval instruction is received (in most cases, the host's instruction only has one keyword), the system combines the context and summarizes 5-8 keywords through association with the knowledge base to complete the retrieval instruction. Related tests:

[0125] ;

[0126] The test results of the retrieval system based on inverted index and distributed computing are as follows:

[0127] ;

[0128] Average response time: 38 ms; Maximum response time: 75 ms; Minimum response time: 28 ms.

[0129] Filter algorithm: It mainly consists of two parts: a very long binary vector (bitmap) and a series of random mapping functions (hash functions). When an element is added to a set, it is mapped to K points in the bitmap through K hash functions, and these points are set to 1. When searching, you only need to check whether these points are all 1. If so, the element may exist in the set; if any point is 0, the element must not exist in the set. The test results are as follows:

[0130] ;

[0131] The above data are approximate values ​​under simulated test environment. Actual performance may vary due to environmental differences.

[0132] At the end of each live broadcast, the system will regularly generate a completed data report, including: original voice, converted text, keyword score, material list of keyword search results, usage frequency of each material, usage duration, etc. According to the usage of each material, the system will first evaluate it according to the built-in rules, such as defining a usage duration of less than 5 seconds as an error result. Each material of each live broadcast will be classified and labeled by relevant personnel, which is also a review of the system evaluation, and then the data characteristics will be recorded. The system will summarize the review results of the live broadcast sessions within the time period and publish them to the material library through scheduled tasks.

[0133] One of the characteristic requirements of a live broadcast assistant is response speed. The topic switching speed in a live broadcast is very fast. After the host triggers the system once, the system needs to hit the appropriate material data in the shortest time and display it in the live broadcast window. The embodiment of this application makes full use of the HTML5+CSS3 of modern browsers, combined with third-party JavaScript frameworks such as jQuery, Bootstrap, CanvasJS, etc., to support most of the material file types in live broadcast scenes, and divide different playback effects and transition animations according to different types. An essential ability of an artificial intelligence system is to gradually learn, improve and improve during use, which is the core ability of data retrieval and matching.

[0134] The embodiments of the present application are applicable to almost all live broadcast platforms in the existing industry, wherein the material playback is pure front-end technology html+css+js, the playback window is based on the browser's built-in windowsopen, and the combination with the live broadcast window utilizes the window sharing of the standard capability in the live broadcast platform; all third-party interfaces in the embodiments of the present application are configurable, so that one capability can support a mode of switching multiple interfaces; product management has good scalability and is suitable for live broadcasts of different categories; the monitoring mechanism is perfect, and the business closed loop is rigorous and suitable for enterprises with different management processes.

[0135] This embodiment also discloses a recommendation system for live broadcast materials. Figure 6 is a module diagram of a recommendation system for live broadcast materials disclosed in an embodiment of the present application, such as Figure 6 As shown, the system includes a text module 601, a scoring module 602, a retrieval module 603 and a recommendation module 604, wherein:

[0136] The text module 601 is configured to obtain the host's voice, convert the voice into text, and perform word segmentation on the text to obtain a plurality of first keywords to form a keyword set;

[0137] A scoring module 602 is configured to score the plurality of first keywords according to a similarity score with the live broadcast theme, a word frequency score, and a score matching with a preset hot word, and select a first preset number of the first keywords with the highest scores as second keywords;

[0138] A search module 603 is configured to search the material library according to the second keyword to obtain information of the material to be selected;

[0139] The recommendation module 604 is configured to select the material with the highest matching degree from the material information to be selected for recommendation according to the live broadcast context and the live broadcast theme.

[0140] Optionally, the scoring module 602 is configured to:

[0141] Performing word segmentation processing on the live broadcast topic to obtain topic segmentations, determining an intersection and a union between the first keyword and the topic segmentations, and calculating a first ratio of the intersection and the union as a similarity score of the first keyword;

[0142] Calculating the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, assigning an importance factor to the first keyword according to semantics, position and context, and calculating the word frequency score of the first keyword according to the word frequency, inverse document frequency and importance factor;

[0143] Searching for a hot word matching the first keyword in a preset knowledge base, and calculating a matching score of the first keyword according to the matching result;

[0144] A weighted sum is performed on the similarity score, the word frequency score and the matching score to obtain a score for the first keyword.

[0145] Optionally, the scoring module 602 is configured to:

[0146] Traversing the text to determine the first number of times the first keyword appears in the text, calculating a second ratio of the first number of times to the second number of times as the word frequency of the first keyword, wherein the second number of times is the total number of times all keywords in the keyword set appear in the text;

[0147] Traversing a preset document set to determine the number of first documents containing the first keyword, and calculating a third ratio of the first document number to the second document number as the document frequency of the first keyword, where the second document number is the total number of all documents in the preset document set;

[0148] Calculating the inverse document frequency according to the document frequency by a preset formula;

[0149] The semantic importance is determined according to the semantic features of the first keyword, the positional importance is determined according to the position of the first keyword in the text, the contextual importance is determined according to the association between the first keyword and other keywords, and the importance factor is determined according to the semantic importance, the positional importance and the contextual importance, wherein the other keywords are keywords in the keyword set other than the first keyword.

[0150] Optionally, the retrieval module 603 is configured to:

[0151] Applying a Bloom filter to filter a data storage structure, removing a column of the data storage structure that does not contain the second keyword to obtain a first target column, the data storage structure comprising a common column and an uncommon column, the common column comprising a field whose search times are greater than a threshold, and the uncommon column comprising a field whose search times are less than or equal to the threshold;

[0152] Retrieving a list of record IDs containing the second keyword on the index node of the first target column by using an inverted index;

[0153] Merging search results from different index nodes, merging all record ID lists to obtain a merged list, and sorting the search results according to relevance;

[0154] The material information to be selected is determined according to the merged list.

[0155] Optionally, the retrieval module 603 is configured to:

[0156] creates a bit vector and initializes all bits of said bit vector to zero,

[0157] Traversing the data storage structure, for a second target column, using the same second preset number of hash functions to perform hash calculation on the identifier or content of the second target column, and checking whether the corresponding second preset number of bits in the bit vector are all ones;

[0158] If all are ones, the second target column is retained as part of the target column;

[0159] All the columns reserved as target columns are merged to form the first target column.

[0160] Optionally, the retrieval module 603 is configured to:

[0161] During the merging process, duplicate record IDs are merged based on their occurrence frequency and weight on different index nodes, so as to retain the duplicate record IDs with the highest occurrence frequency and weight.

[0162] Calculate the relevance score of the target record ID according to the material content corresponding to the target record ID, the relevance between the material and the live broadcast theme, and the popularity of the material, where the target record ID is any record ID in the merged list;

[0163] The record IDs in the merged list are sorted in descending order according to the relevance scores.

[0164] Optionally, the recommendation module 604 is configured to:

[0165] Analyze the context information of the live broadcast and extract target information points in the context information, wherein the target information points include people, places and events;

[0166] Combining the target information point with the live broadcast theme to form a comprehensive query condition, and filtering and selecting the to-be-selected material information according to the comprehensive query condition, and selecting the to-be-selected material that meets both the live broadcast theme and the target information point;

[0167] Weights are set for timeliness, quality, and user feedback, respectively, and comprehensive scores of the candidate materials are calculated according to the weights. The selected materials are sorted in descending order according to the comprehensive scores, and the candidate materials with the highest comprehensive scores are pushed as recommended materials.

[0168] It should be noted that: when the device provided in the above embodiment realizes its function, only the division of the above functional modules is used as an example. In actual application, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0169] This embodiment also discloses an electronic device, referring to Figure 7 The electronic device may include: at least one processor 701 , at least one communication bus 702 , a user interface 703 , a network interface 704 , and at least one memory 705 .

[0170] The communication bus 702 is used to realize the connection and communication between these components.

[0171] The user interface 703 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 703 may also include a standard wired interface and a wireless interface.

[0172] The network interface 704 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0173] Among them, the processor 701 may include one or more processing cores. The processor 701 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 705, and calling data stored in the memory 705. Optionally, the processor 701 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 701 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 701, and it can be implemented separately through a chip.

[0174] Among them, the memory 705 may include a random access memory (Random Access Memory, RAM) and may also include a read-only memory (Read-Only Memory). Optionally, the memory 705 includes a non-transitory computer-readable storage medium. The memory 705 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 705 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 705 may also be optionally at least one storage device located away from the aforementioned processor 701. As Figure 7 As shown, the memory 705 as a computer storage medium may include an operating system, a network communication module, a user interface module, and an application program for a method of recommending live broadcast materials.

[0175] exist Figure 7 In the electronic device shown, the user interface 703 is mainly used to provide an input interface for the user and obtain data input by the user; and the processor 701 can be used to call the application of the recommended method for storing live broadcast materials in the memory 705. When executed by one or more processors 701, the electronic device executes one or more methods in the above-mentioned embodiments.

[0176] It should be noted that, for the aforementioned method embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the order of the actions described, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware 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.

[0177] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device 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 devices or units can be electrical or other forms.

[0179] 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.

[0180] 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.

[0181] 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 705 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 705 includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk.

[0182] The above is only an exemplary embodiment 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, those skilled in the art will easily think of other embodiments of the present disclosure. This application is intended to cover any modification, 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 technical field that are not recorded in the present disclosure. The description and examples are only regarded as exemplary, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A method for recommending live broadcast material, characterized in that: Applied to a live broadcast assistant system, the method includes: Acquire the host's voice, convert the voice into text, and perform word segmentation on the text to obtain a plurality of first keywords to form a keyword set; Scoring the plurality of first keywords according to the similarity score with the live broadcast theme, the word frequency score, and the score matching the preset hot words, and selecting a first preset number of the first keywords with the highest scores as the second keywords; Searching the material library according to the second keyword to obtain information of the material to be selected; According to the live broadcast context and the live broadcast theme, the material with the highest matching degree is selected from the material information to be selected for recommendation, Scoring the first keywords according to the similarity score with the live broadcast theme, the word frequency score and the matching score with the preset hot words includes: Performing word segmentation processing on the live broadcast topic to obtain topic segmentations, determining an intersection and a union between the first keyword and the topic segmentations, and calculating a first ratio of the intersection and the union as a similarity score of the first keyword; Calculating the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, assigning an importance factor to the first keyword according to semantics, position and context, and calculating the word frequency score of the first keyword according to the word frequency, inverse document frequency and importance factor; Searching for a hot word matching the first keyword in a preset knowledge base, and calculating a matching score of the first keyword according to the matching result; Performing a weighted summation on the similarity score, the word frequency score and the matching score to obtain a score for the first keyword, The calculating of the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, and assigning the importance factor to the first keyword according to semantics, position and context includes: Traversing the text to determine the first number of times the first keyword appears in the text, calculating a second ratio of the first number of times to the second number of times as the word frequency of the first keyword, wherein the second number of times is the total number of times all keywords in the keyword set appear in the text; Traversing a preset document set to determine the number of first documents containing the first keyword, and calculating a third ratio of the first document number to the second document number as the document frequency of the first keyword, where the second document number is the total number of all documents in the preset document set; Calculating the inverse document frequency according to the document frequency by a preset formula; The semantic importance is determined according to the semantic features of the first keyword, the positional importance is determined according to the position of the first keyword in the text, the contextual importance is determined according to the association between the first keyword and other keywords, and the importance factor is determined according to the semantic importance, the positional importance and the contextual importance, wherein the other keywords are keywords in the keyword set other than the first keyword.

2. The method for recommending live broadcast material according to claim 1, characterized in that: The searching in the material library according to the second keyword to obtain the material information to be selected includes: Applying a Bloom filter to filter a data storage structure, removing a column of the data storage structure that does not contain the second keyword to obtain a first target column, the data storage structure comprising a common column and an uncommon column, the common column comprising a field whose search times are greater than a threshold, and the uncommon column comprising a field whose search times are less than or equal to the threshold; Retrieving a list of record IDs containing the second keyword on the index node of the first target column by using an inverted index; Merging search results from different index nodes, merging all record ID lists to obtain a merged list, and sorting the search results according to relevance; The material information to be selected is determined according to the merged list.

3. The method for recommending live broadcast material according to claim 2, characterized in that: The applying a Bloom filter to filter the data storage structure and removing the column of the data storage structure that does not contain the second keyword to obtain the first target column includes: creates a bit vector and initializes all bits of said bit vector to zero, Traversing the data storage structure, for a second target column, using the same second preset number of hash functions to perform hash calculation on the identifier or content of the second target column, and checking whether the corresponding second preset number of bits in the bit vector are all ones; If all are ones, the second target column is retained as part of the target column; All the columns reserved as target columns are merged to form the first target column.

4. The method for recommending live broadcast material according to claim 2, characterized in that: The merging of the search results from different index nodes, merging all the record ID lists to obtain a merged list, and sorting the search results according to relevance includes: During the merging process, duplicate record IDs are merged based on their occurrence frequency and weight on different index nodes, so as to retain the duplicate record IDs with the highest occurrence frequency and weight. Calculate the relevance score of the target record ID according to the material content corresponding to the target record ID, the relevance between the material and the live broadcast theme, and the popularity of the material, where the target record ID is any record ID in the merged list; The record IDs in the merged list are sorted in descending order according to the relevance scores.

5. The method for recommending live broadcast material according to claim 1, characterized in that: The selecting the material with the highest matching degree from the to-be-selected material information for recommendation according to the live broadcast context and the live broadcast theme includes: Analyze the context information of the live broadcast and extract target information points in the context information, wherein the target information points include people, places and events; Combining the target information point with the live broadcast theme to form a comprehensive query condition, and filtering and selecting the to-be-selected material information according to the comprehensive query condition, and selecting the to-be-selected material that meets both the live broadcast theme and the target information point; Weights are set for timeliness, quality, and user feedback, respectively, and comprehensive scores of the candidate materials are calculated according to the weights. The selected materials are sorted in descending order according to the comprehensive scores, and the candidate materials with the highest comprehensive scores are pushed as recommended materials.

6. A recommendation system for live broadcast materials, characterized in that: It includes text module, scoring module, retrieval module and recommendation module, among which: A text module is configured to obtain the host's voice, convert the voice into text, and perform word segmentation on the text to obtain a plurality of first keywords to form a keyword set; A scoring module configured to score the plurality of first keywords according to a similarity score with a live broadcast theme, a word frequency score, and a score matching a preset hot word, and select a first preset number of the first keywords with the highest scores as second keywords; A retrieval module, configured to search the material library according to the second keyword to obtain information of the material to be selected; A recommendation module is configured to select the material with the highest matching degree from the material information to be selected for recommendation according to the live broadcast context and the live broadcast theme, Scoring the first keywords according to the similarity score with the live broadcast theme, the word frequency score and the matching score with the preset hot words includes: Performing word segmentation processing on the live broadcast topic to obtain topic segmentations, determining an intersection and a union between the first keyword and the topic segmentations, and calculating a first ratio of the intersection and the union as a similarity score of the first keyword; Calculating the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, assigning an importance factor to the first keyword according to semantics, position and context, and calculating the word frequency score of the first keyword according to the word frequency, inverse document frequency and importance factor; Searching for a hot word matching the first keyword in a preset knowledge base, and calculating a matching score of the first keyword according to the matching result; Performing a weighted summation on the similarity score, the word frequency score and the matching score to obtain a score for the first keyword, The calculating of the word frequency and document frequency of the first keyword, calculating the inverse document frequency according to the document frequency, and assigning the importance factor to the first keyword according to semantics, position and context includes: Traversing the text to determine the first number of times the first keyword appears in the text, calculating a second ratio of the first number of times to the second number of times as the word frequency of the first keyword, wherein the second number of times is the total number of times all keywords in the keyword set appear in the text; Traversing a preset document set to determine the number of first documents containing the first keyword, and calculating a third ratio of the first document number to the second document number as the document frequency of the first keyword, where the second document number is the total number of all documents in the preset document set; Calculating the inverse document frequency according to the document frequency by a preset formula; The semantic importance is determined according to the semantic features of the first keyword, the positional importance is determined according to the position of the first keyword in the text, the contextual importance is determined according to the association between the first keyword and other keywords, and the importance factor is determined according to the semantic importance, the positional importance and the contextual importance, wherein the other keywords are keywords in the keyword set other than the first keyword.

7. An electronic device, characterized in that: It includes 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 both 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 as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 5 is performed.

Images