Live broadcast comment analysis method and device, storage medium and computer equipment

Abstract

According to the live broadcast comment analysis method and device, the storage medium and the computer equipment provided by the invention, after the original comments generated during live broadcast of the live broadcast platform are obtained, the original comments are preprocessed, noise interference is eliminated, and the subsequent analysis quality and accuracy are improved. And after the preprocessed comment information is obtained, classifying the comment information according to themes, extracting representative keywords from a clustering result, then classifying the keywords, inputting a classification result and the comment information in a clustering cluster into a preset large language model for further analysis, and outputting a detailed analysis result. In this way, on one hand, emerging topic categories can be captured and recognized in real time through topic clustering, and on the other hand, multi-dimensional correlation analysis is achieved by means of a large language model. Besides, the LLM and clustering two-way analysis architecture is combined, the analysis efficiency, depth and breadth can be considered, the manual operation cost is reduced, the live broadcast operation efficiency is improved, data support is provided for live broadcast effect evaluation, and refined and scientific operation decision making is assisted.

Description

Technical Field

[0001] The present application relates to the field of live broadcast operation technology, and in particular to a live broadcast comment analysis method, device, storage medium and computer equipment. Background Art

[0002] Currently, in the live broadcasting process, in order to effectively help the broadcaster and the entire operations team grasp the needs and feedback of the majority of users in real time and accurately, and then optimize the content presentation, product promotion strategies, and interactive user experience, the operations team generally manually collects the massive amount of comment data generated during the live broadcast. Subsequently, this massive amount of comment data is carefully analyzed to extract valuable information and formulate corresponding and feasible operations strategies based on it.

[0003] However, this process currently relies primarily on manual operations and analysis. This approach not only leads to relatively low overall efficiency in live streaming operations and makes it difficult to adapt to the rapidly changing live streaming environment and user needs, but also inevitably increases manual operation costs and increases the burden on enterprises. More importantly, because manual analysis is often simple and rough, lacking depth and breadth, it is difficult to provide solid data support for a comprehensive and scientific evaluation of live streaming results, which to a certain extent restricts the optimization and improvement of live streaming operations. Summary of the Invention

[0004] The purpose of this application is to solve at least one of the above-mentioned technical defects, especially the technical defects in that the existing technology mainly relies on manual operation and analysis when analyzing live broadcast comments, resulting in low live broadcast operation efficiency, increased manual operation costs, and inability to provide solid data support for the live broadcast effect.

[0005] This application provides a live broadcast comment analysis method, which includes:

[0006] Obtaining original comments generated by the live broadcast platform during the live broadcast process, and obtaining comment information after pre-processing the original comments;

[0007] After subject-clustering the review information, a plurality of clustering results are obtained, keywords in each clustering result are extracted, and each keyword is classified to obtain a classification result;

[0008] The classification results of each keyword and the comment information in the cluster corresponding to each classification result are respectively input into a preset large language model to obtain analysis results corresponding to each classification result output by the large language model.

[0009] Optionally, the comment information obtained after preprocessing the original comment includes:

[0010] Performing text cleaning on the original comments, and performing word segmentation processing on the cleaned original comments to obtain word segmentation results;

[0011] After removing invalid characters from the word segmentation results, the word segmentation results after removing the invalid characters are normalized to obtain final comment information.

[0012] Optionally, after performing topic clustering on the comment information, multiple clustering results are obtained, including:

[0013] Convert the review information into a semantic vector;

[0014] Clustering the semantic vectors, and filtering noise data of the semantic vectors during the clustering process to obtain a plurality of clusters;

[0015] After generating topic labels corresponding to each cluster, multiple clustering results are obtained.

[0016] Optionally, before clustering the semantic vectors, the method further includes:

[0017] Obtaining user basic information, live broadcast metadata, and comment interaction data corresponding to the original comment;

[0018] Converting the user basic information, the live broadcast metadata, and the comment interaction data into a structured vector;

[0019] The structured vector is fused with the semantic vector, and the fused semantic vector is clustered.

[0020] Optionally, after generating the topic labels corresponding to each cluster, multiple clustering results are obtained, including:

[0021] For each cluster:

[0022] Segment the semantic vectors in the cluster, and extract representative words from the cluster based on the segmentation results;

[0023] After generating topic labels corresponding to the representative words, a clustering result of the cluster is obtained.

[0024] Optionally, extracting keywords from each clustering result separately includes:

[0025] For each clustering result:

[0026] Use TF-IDF to calculate the frequency weight of each word in the clustering result and obtain the weight corresponding to each word;

[0027] Stop word filtering is performed on each word in the clustering result, and the word with the highest weight is selected from the filtered words as the keyword.

[0028] Optionally, after classifying each keyword, a classification result is obtained, including:

[0029] Get the preset rule library;

[0030] Each keyword is matched with the preset rule library respectively, and a classification result corresponding to each keyword is determined according to the matching result.

[0031] This application also provides a live comment analysis device, including:

[0032] An information acquisition module is used to obtain original comments generated by the live broadcast platform during the live broadcast process, and obtain comment information after pre-processing the original comments;

[0033] A topic clustering module is used to perform topic clustering on the review information to obtain multiple clustering results, extract keywords from each clustering result, and classify each keyword to obtain a classification result;

[0034] The intelligent analysis module is used to input the classification results of each keyword and the comment information in the cluster corresponding to each classification result into a preset large language model, and obtain the analysis results corresponding to each classification result output by the large language model.

[0035] The present application also provides a computer-readable storage medium, which stores computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the live commentary analysis method as described in any of the above embodiments.

[0036] The present application also provides a computer device, comprising: one or more processors, and a memory;

[0037] The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the live comment analysis method as described in any one of the above embodiments are performed.

[0038] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0039] The live comment analysis method, device, storage medium and computer equipment provided by the present application can perform a series of preprocessing operations on the original comments generated by the live broadcast platform during the live broadcast process after obtaining the original comments. The purpose of this step is to effectively eliminate various noise interference factors that may exist in the original comments, thereby significantly improving the quality and accuracy of subsequent analysis work; after successfully obtaining the preprocessed comment information, the present application can classify the comment information according to different topics, and accurately extract representative keywords in each clustering result. Subsequently, the present application can further classify these keywords, and input the classification results corresponding to each keyword and the comment information in the cluster cluster corresponding to each classification result into a preset large language model for further analysis and processing. Through the powerful processing capabilities of the large language model, detailed analysis results corresponding to each classification result are output. The advantage of doing this is that, on the one hand, the method of topic clustering can be used to capture in real time and automatically discover and identify emerging topic categories based on the dynamic changes in live broadcast comments. On the other hand, with the help of the comprehensive analysis capabilities of the large language model, this application can realize the correlation analysis of multiple dimensions such as sentiment tendencies, product problems, and improvement suggestions. In addition, this application uses a dual-path analysis architecture that combines the LLM model and the clustering model to ensure the efficiency of analysis while taking into account the depth and breadth of analysis. Through this intelligent analysis method, the cost of manual operation can be effectively reduced, the overall efficiency of live broadcast operations can be significantly improved, and solid data support can be provided for the comprehensive evaluation of live broadcast effects, ultimately helping to achieve more refined and scientific operational decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0041] Figure 1 A flowchart of a live comment analysis method provided in an embodiment of the present application;

[0042] Figure 2 A schematic diagram of the process of subject-clustering review information provided in an embodiment of the present application;

[0043] Figure 3 A schematic diagram of the process of classifying keywords provided in an embodiment of the present application;

[0044] Figure 4 A schematic diagram of the structure of a live comment analysis device provided in an embodiment of the present application;

[0045] Figure 5 A schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0047] In one embodiment, Figure 1 As shown, Figure 1 A flowchart of a live comment analysis method provided in an embodiment of the present application is provided. The present application provides a live comment analysis method, which may include:

[0048] S110: Obtain original comments generated by the live broadcast platform during the live broadcast process, and obtain comment information after pre-processing the original comments.

[0049] In this step, during the live broadcast operation, the API service module can be used to connect with the live broadcast platform and obtain the original comments generated by the live broadcast platform during the live broadcast. Then, in order to improve the accuracy of the analysis results, this application can also perform pre-processing operations on the obtained original comments to eliminate various noise interference factors that may exist in the original comments, thereby improving the subsequent analysis quality and accuracy.

[0050] The API service module of this application can provide live comment analysis services through RESTful API interfaces (such as / api / v1 / analyze, / api / v1 / analyze_questions, etc.). Of course, this application can also provide corresponding live comment analysis services through other interfaces. The specific settings can be set according to actual circumstances and are not restricted here. Because this type of interface supports real-time comment analysis and batch analysis of comments for the entire live broadcast, it can be integrated with live broadcast platforms or other systems, lowering the threshold for service invocation, improving the reusability of analysis capabilities, and supporting external systems to obtain analysis results on demand.

[0051] After obtaining the original comments generated during the live broadcast, this application can pre-process the original comments. The pre-processing process includes but is not limited to text cleaning, word segmentation, character filtering, and standardization. Among them, text cleaning can remove irrelevant information such as garbled characters, advertisements, special symbols, etc. in the original comments to improve the accuracy and efficiency of subsequent word segmentation; word segmentation can split the comments into independent words to facilitate subsequent processing and analysis; after filtering invalid characters in the word segmentation results, standardization can be performed more efficiently, such as converting words to lowercase, removing duplicate words, etc., so as to obtain the final comment information and provide basic data for subsequent topic clustering, keyword extraction and other steps.

[0052] S120: After performing topic clustering on the comment information, a plurality of clustering results are obtained, keywords in each clustering result are extracted respectively, and each keyword is classified to obtain a classification result.

[0053] In this step, the original comments generated by the live broadcast platform during the live broadcast are obtained through S110, and the original comments are pre-processed. After obtaining the comment information, this application can also perform topic clustering on the comment information, and after extracting the keywords in each clustering result, classify each keyword so as to perform intelligent analysis based on the classification results.

[0054] Specifically, during the topic clustering process, this application can employ classic clustering algorithms such as DBSCAN (Density-Based Spatial Clustering with Noise) and HDBSCAN (Hierarchical Density-Based Spatial Clustering), or utilize deep learning models such as BERTopic to achieve topic clustering of review information. These algorithms or models can automatically identify topics within reviews and group similar reviews into the same cluster. Subsequently, for each clustering result, this application can use algorithms such as TF-IDF and TextRank to extract keywords that represent the theme or core concept of the cluster.

[0055] Furthermore, when classifying keywords, this application can be based on a preset rule base or classification system. For example, this application can classify keywords into multiple categories such as positive reviews, negative reviews, product inquiries, and event suggestions. By matching keywords with entries in the rule base, the classification results for each keyword can be determined. This step facilitates a more detailed and targeted analysis of the review information later.

[0056] It is worth noting that this application fully considers the characteristics and diversity of live broadcast comments when extracting keywords and classifying them. For example, live broadcast comments may contain a large amount of colloquial expressions and internet buzzwords, which may be regarded as noise or invalid information in traditional text analysis. However, this application, by using advanced natural language processing technology and deep learning models, can accurately identify and extract valuable information from these expressions, thereby more comprehensively and accurately reflecting users' real needs and feedback.

[0057] Furthermore, this application offers flexible scalability. For example, the application can adjust and expand the pre-set rule base or classification system based on actual needs to adapt to different live broadcast scenarios and changing user needs. Furthermore, this application supports the optimization and improvement of clustering algorithms and keyword extraction algorithms to improve the accuracy and efficiency of analysis.

[0058] S130: Inputting the classification results of each keyword and the comment information in the cluster corresponding to each classification result into a preset large language model respectively, and obtaining the analysis results corresponding to each classification result output by the large language model.

[0059] In this step, after the comment information is clustered by topic through S120, multiple clustering results are obtained. The keywords in each clustering result are extracted respectively, and each keyword is classified to obtain the classification results. In this application, the classification results of each keyword and the comment information in the cluster cluster corresponding to each classification result can be input into the preset large language model respectively, so as to output the analysis results corresponding to each classification result through the large language model.

[0060] It is understandable that the above-mentioned large language model of this application refers to a model built on deep learning technology with powerful natural language processing capabilities. The model has been trained with a large amount of text data, and can deeply understand the content of the text and accurately capture the key information and contextual relationships in the text. In this application, the large language model is used to conduct in-depth analysis of the classified keywords and their corresponding clustered comment information, thereby outputting detailed analysis results associated with each classification result. These analysis results may include multiple dimensions such as emotional tendencies, product problems, and improvement suggestions, thereby providing comprehensive and in-depth data insights for live broadcast operators, helping them make more scientific and refined operational decisions.

[0061] For example, this application can use the Qwen-Plus model as a preset large language model, and use the Qwen-Plus module to perform multi-dimensional analysis on the classification results of each keyword and the comment information within the cluster corresponding to each classification result. For example, when this application inputs the keywords "delay", "frame drop", and "stuttering" to the Qwen-Plus model, and the classification label is "technical problem", and the comment information within the corresponding cluster includes information such as "live broadcast stuttering, affecting viewing", the model can associate technical terms (such as bandwidth, encoding) to generate corresponding reasons and improvement suggestions. For example, the cause analysis is "users concentratedly report the "stuttering" problem, which may be caused by insufficient server bandwidth or network fluctuations on the host side", and the improvement suggestion is "check the network stability of the live streaming equipment and increase the coverage of CDN nodes during high concurrency periods."

[0062] For example, if the keyword input into this application is "insufficient discount" and it is classified as a "negative review," the large language model can not only identify the dissatisfaction in the review, but also further analyze the user's potential discount range and the impact of this expectation on the purchase decision. This detailed analysis capability enables this application to provide more precise operational suggestions, such as adjusting discount strategies and enhancing user interaction.

[0063] In addition, the live comment analysis system of this application is designed with a multimodal analysis interface reserved to support future expansion to more data types such as live images and voice. This design will enable the system to better adapt to different application scenarios and business needs in the future, and enhance the comprehensive application value and competitiveness of the system. In addition, the live comment analysis system of this application also includes a log module. The main function of the log module is to comprehensively and carefully record various key information (such as request records, exception information, processing time, etc.) during the operation of the system. The recorded key information can be used for system debugging, performance monitoring, troubleshooting and operation auditing, and can also ensure system stability and traceability.

[0064] In the above embodiment, after obtaining the original comments generated by the live broadcast platform during the live broadcast process, the present application can perform a series of preprocessing operations on the original comments. The purpose of this step is to effectively eliminate various noise interference factors that may exist in the original comments, thereby significantly improving the quality and accuracy of subsequent analysis work; after successfully obtaining the preprocessed comment information, the present application can classify the comment information according to different topics, and accurately extract representative keywords in each clustering result. Subsequently, the present application can further classify these keywords, and input the classification results corresponding to each keyword and the comment information in the cluster cluster corresponding to each classification result into the preset large language model for further analysis and processing. Through the powerful processing capabilities of the large language model, detailed analysis results corresponding to each classification result are output. The advantage of doing this is that, on the one hand, the method of topic clustering can be used to capture in real time and automatically discover and identify emerging topic categories based on the dynamic changes in live broadcast comments. On the other hand, with the help of the comprehensive analysis capabilities of the large language model, this application can realize the correlation analysis of multiple dimensions such as sentiment tendencies, product problems, and improvement suggestions. In addition, this application uses a dual-path analysis architecture that combines the LLM model and the clustering model to ensure the efficiency of analysis while taking into account the depth and breadth of analysis. Through this intelligent analysis method, the cost of manual operation can be effectively reduced, the overall efficiency of live broadcast operations can be significantly improved, and solid data support can be provided for the comprehensive evaluation of live broadcast effects, ultimately helping to achieve more refined and scientific operational decisions.

[0065] In one embodiment, after pre-processing the original comments in S110, comment information is obtained, which may include:

[0066] S111: performing text cleaning on the original comments, and performing word segmentation processing on the cleaned original comments to obtain word segmentation results.

[0067] S112: After removing invalid characters from the word segmentation results, the word segmentation results after removing the invalid characters are normalized to obtain final comment information.

[0068] In this embodiment, when preprocessing the original comments, the original comments can be first cleaned and then the cleaned original comments can be segmented. Then, this application can remove invalid characters in the segmentation results and perform standardization, so that the final comment information can be obtained.

[0069] In a specific implementation method, the present application can first perform text cleaning on the original comments to remove invalid content such as garbled characters, advertising links, duplicate comments, etc., so as to improve the efficiency and accuracy of subsequent analysis. Subsequently, the present application can use a word segmentation tool (such as jieba) to perform word segmentation on the cleaned original comments to obtain word segmentation results. Word segmentation is a basic task in natural language processing, and its purpose is to divide continuous text into independent vocabulary units. Next, the present application can remove invalid characters in the word segmentation results, such as punctuation marks, special symbols, etc. These characters are of no substantial help for subsequent semantic analysis and topic clustering, but will increase the complexity of processing. Therefore, removing invalid characters can not only improve analysis efficiency, but also improve analysis quality. Finally, the present application can standardize the word segmentation results after removing invalid characters. Standardization processing includes unifying text formats (such as case conversion, traditional to simplified), normalizing numbers and date formats, etc.

[0070] After the above preprocessing steps, the final review information can be obtained, which will be used for subsequent topic clustering, keyword extraction and classification analysis.

[0071] In one embodiment, Figure 2 As shown, Figure 2 A schematic diagram of the process of subject-clustering comment information provided in an embodiment of the present application; after subject-clustering the comment information in S120, multiple clustering results are obtained, which may include:

[0072] S121: Convert the comment information into a semantic vector.

[0073] S122: Clustering the semantic vectors, and filtering noise data of the semantic vectors during the clustering process to obtain multiple clusters.

[0074] S123: After generating a topic label corresponding to each cluster, multiple clustering results are obtained.

[0075] In this embodiment, in order to achieve topic clustering of comment information, the present application can first convert the comment information into semantic vectors. The purpose of this step is to convert the comment information in text form into a numerical form that can be understood and processed by the computer, so as to perform subsequent clustering operations. Specifically, the present application can use pre-trained language models such as BERT, GPT, Sentence-BERT, or use traditional text vectorization methods such as Word2Vec and TF-IDF to convert the comment information into points in a high-dimensional semantic vector space. These semantic vectors can capture the key information and contextual relationships in the comment information, thereby providing strong support for subsequent clustering operations.

[0076] After obtaining the semantic vectors of the comment information, the present application can cluster these semantic vectors. Clustering is an unsupervised learning method that can classify similar objects into one category and separate dissimilar objects. In the present application, a clustering algorithm is used to divide the semantic vectors into multiple clusters, each cluster representing a theme or topic. To achieve this goal, the present application can use clustering algorithms such as DBSCAN, HDBSCAN, or use deep learning models (such as DeepCluster, etc.) to cluster semantic vectors. During the clustering process, the present application can also filter out noise data in the semantic vector, such as invalid comments, duplicate comments, or comments that are not related to other topics, so as to improve the accuracy and efficiency of clustering.

[0077] Finally, this application can also generate corresponding topic labels for each cluster. Topic labels are intuitive descriptions of the clustering results, which can help users quickly understand the theme or topic represented by each cluster. To generate accurate and meaningful topic labels, this application can extract key words or phrases in the cluster and conduct comprehensive analysis and judgment based on contextual information. The resulting topic labels are not only concise and clear, but also accurately reflect the core content and characteristics of the cluster.

[0078] Through the above examples, the present application can obtain multiple clustering results, each of which contains a topic label and the corresponding clustered comment information. This information will be used for subsequent keyword extraction, classification analysis, and large language model processing.

[0079] In one embodiment, before clustering the semantic vectors, the following steps may be further included:

[0080] S1211: Obtain user basic information, live broadcast metadata, and comment interaction data corresponding to the original comment.

[0081] S1212: Convert the user basic information, the live broadcast metadata, and the comment interaction data into a structured vector.

[0082] S1213: Fusing the structured vector with the semantic vector, and clustering the fused semantic vector.

[0083] In this embodiment, before performing semantic vector clustering, in order to further improve the accuracy and effectiveness of clustering, this application may also consider introducing more relevant information. Specifically, this application can obtain user basic information, live broadcast metadata, and comment interaction data related to the original comments. Among them, user basic information (such as user ID, level, historical interaction records) is used to associate user portraits with comment content; live broadcast metadata (such as live broadcast session ID, start time, product list) is used for contextual association analysis (such as comment focus on specific products); comment interaction data (such as the number of likes and replies) is used to evaluate the influence of comments and user participation. This information can reflect user behavior and preferences, as well as the characteristics and effects of live broadcasts from different dimensions, thereby providing a richer and more comprehensive basis for clustering operations.

[0084] Once this information is obtained, the application can convert it into a structured vector. It is understood that structured vectors are a technique for converting unstructured data (such as text, images, etc.) into a structured form (such as a numerical vector). This technique preserves the key information and features of the original data while facilitating computer processing and analysis. For basic user information, live broadcast metadata, and comment interaction data, the application can employ existing feature extraction and vectorization methods to convert them into points in a high-dimensional vector space.

[0085] Next, this application can fuse these structured vectors with semantic vectors. The purpose of fusion is to integrate information from different sources and in different forms for more in-depth and comprehensive analysis. In order to achieve this goal, this application can use methods such as vector splicing, weighted averaging or deep learning models to organically fuse structured vectors with semantic vectors to obtain fused semantic vectors. These fused semantic vectors not only contain the key content and contextual relationships in the comment information, but also incorporate useful information from user basic information, live broadcast metadata and comment interaction data, thereby providing a more comprehensive and accurate basis for subsequent clustering operations.

[0086] In one embodiment, after generating the topic labels corresponding to each cluster in S123, multiple clustering results are obtained, which may include:

[0087] S1231: For each cluster: perform word segmentation on the semantic vectors in the cluster, and extract representative words from the cluster based on the word segmentation results.

[0088] S1232: After generating topic tags corresponding to the representative words, a clustering result of the cluster is obtained.

[0089] In this embodiment, when generating topic tags corresponding to each cluster, in order to ensure the accuracy and representativeness of the tags, this application can generate corresponding topic tags (such as "product quality" and "logistics speed") for each cluster based on BERTopic technology, and this technology also supports dynamic topic updates and visual displays (such as topic distribution histograms), thereby adapting to the dynamic changes in live comments.

[0090] Specifically, for each cluster, the present application can first perform detailed word segmentation on the semantic vector or the fused semantic vector. Word segmentation is a basic link in natural language processing. It can divide continuous text into independent vocabulary units, which are the key basis for subsequent analysis. After successfully completing the word segmentation process, the present application can further filter out representative words from these word segmentation results. These representative words are usually words that appear frequently in the cluster and can accurately reflect the core content and characteristics of the cluster. In order to achieve this goal, the present application can use algorithms such as TF-IDF (term frequency-inverse document frequency), TextRank, or use deep learning models for keyword extraction.

[0091] After obtaining representative words, this application can further generate topic tags corresponding to these words. Topic tags are intuitive and concise descriptions of clustering results, which can help users quickly understand the theme or topic represented by each cluster. In order to generate accurate and meaningful topic tags, this application can comprehensively consider the word frequency, contextual relationships, and overall content of the representative words, thereby ensuring that the generated tags are both concise and clear, and can accurately reflect the core content and characteristics of the cluster.

[0092] Through the above embodiments, this application can achieve refined topic clustering of live broadcast comment information and generate accurate and meaningful topic labels for each cluster. This information not only helps live broadcast operators quickly understand the comment content, but also provides them with comprehensive and in-depth data insights, thereby helping them make more scientific and refined operational decisions.

[0093] In one embodiment, extracting keywords from each clustering result in S120 may include:

[0094] S124: For each clustering result: use TF-IDF to calculate the word frequency weight of each word in the clustering result to obtain the weight corresponding to each word.

[0095] S125: Perform stop word filtering on each word in the clustering result, and select the word with the highest weight from the filtered words as the keyword.

[0096] In this embodiment, when extracting keywords from each clustering result, to ensure the accuracy and effectiveness of the keywords, this application can use the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm to calculate the word frequency weights of each word. It can be understood that TF-IDF is a commonly used text mining and information retrieval technology, which can evaluate the importance of a word for a document set or a single document in a corpus. Among them, the term frequency (TF) reflects the frequency of a word appearing in a document, while the inverse document frequency (IDF) reflects how much information a word provides. That is, the more common a word is, the lower the IDF value; conversely, the less common a word is, the higher the IDF value. Therefore, the higher the TF-IDF value, the more important this word is in the document and the more likely it is to be a keyword.

[0097] In a specific implementation, for each clustering result, this application can first calculate the TF-IDF values of each word therein to obtain the weights corresponding to each word. The purpose of this step is to provide a basis for subsequent keyword selection by quantifying the importance of each word in the clustering result. Subsequently, this application can perform stop word filtering on each word in each clustering result. Stop words refer to words that appear frequently in the text but contribute little to the text theme, such as "de" (的), "le" (了), etc. The purpose of filtering stop words is to reduce noise and improve the accuracy and efficiency of keyword extraction. After filtering out the stop words, this application can select the word with the highest weight from the remaining words as the keyword. These keywords can accurately reflect the core content and features of each clustering result and provide strong support for subsequent classification analysis and large language model processing and other links.

[0098] Through the above embodiments, this application can achieve refined keyword extraction for live comment information and generate an accurate and meaningful keyword list for each clustering result.

[0099] In one embodiment, as Figure 3 shown, Figure 3 is a schematic diagram of the process of classifying each keyword provided by the embodiment of this application; after classifying each keyword in S120, the classification result can include:

[0100] S126: Obtain a preset rule library.

[0101] S127: Match each keyword with the preset rule library respectively, and determine the classification result corresponding to each keyword according to the matching result.

[0102] In this embodiment, in order to classify each keyword, the present application can pre-build a corresponding rule base. Then, when performing keyword classification, the preset rule base is obtained and keyword matching is performed using the preset rule base. The preset rule base can be established based on the knowledge and experience of domain experts, or obtained through analysis and mining of a large amount of historical data. The preset rule base contains multiple classification standards and corresponding keyword features. These rules can accurately reflect the characteristics and differences of keywords in different categories, thereby providing strong support for classification operations.

[0103] After obtaining the preset rule base, the present application can match each keyword with the preset rule base respectively. The purpose of this step is to determine the category to which the keyword belongs by comparing the keyword with the features in the preset rule base. To achieve this goal, the present application can use string matching, regular expression matching or machine learning-based classification algorithms to compare the keywords with the entries in the preset rule base one by one. During the matching process, the present application can consider the keyword's word frequency, contextual relationship and various features in the preset rule base to ensure the accuracy and comprehensiveness of the match.

[0104] After the matching is completed, this application can determine the classification results corresponding to each keyword based on the matching results. These classification results can accurately reflect the theme and category of the keywords, providing a strong basis for subsequent analysis and processing. For example, for keywords related to product quality, they can be classified into the "product evaluation" category; for keywords related to logistics speed, they can be classified into the "logistics service" category. Through such classification operations, this application can achieve refined management and analysis of live broadcast comment information, providing live broadcast operators with more comprehensive and in-depth data insights.

[0105] The live broadcast comment analysis device provided in an embodiment of the present application is described below. The live broadcast comment analysis device described below and the live broadcast comment analysis method described above can be referenced to each other.

[0106] In one embodiment, Figure 4 As shown, Figure 4 This is a schematic diagram of the structure of a live comment analysis device provided in an embodiment of the present application. The present application also provides a live comment analysis device, which may include an information acquisition module 210, a topic clustering module 220, and an intelligent analysis module 230, specifically including the following:

[0107] The information acquisition module 210 is used to obtain original comments generated by the live broadcast platform during the live broadcast process, and obtain comment information after pre-processing the original comments.

[0108] The topic clustering module 220 is used to perform topic clustering on the comment information to obtain multiple clustering results, extract keywords from each clustering result, and classify each keyword to obtain a classification result.

[0109] The intelligent analysis module 230 is used to input the classification results of each keyword and the comment information in the cluster corresponding to each classification result into a preset large language model, and obtain the analysis results corresponding to each classification result output by the large language model.

[0110] In the above embodiment, after obtaining the original comments generated by the live broadcast platform during the live broadcast process, the present application can perform a series of preprocessing operations on the original comments. The purpose of this step is to effectively eliminate various noise interference factors that may exist in the original comments, thereby significantly improving the quality and accuracy of subsequent analysis work; after successfully obtaining the preprocessed comment information, the present application can classify the comment information according to different topics, and accurately extract representative keywords in each clustering result. Subsequently, the present application can further classify these keywords, and input the classification results corresponding to each keyword and the comment information in the cluster cluster corresponding to each classification result into the preset large language model for further analysis and processing. Through the powerful processing capabilities of the large language model, detailed analysis results corresponding to each classification result are output. The advantage of doing this is that, on the one hand, the method of topic clustering can be used to capture in real time and automatically discover and identify emerging topic categories based on the dynamic changes in live broadcast comments. On the other hand, with the help of the comprehensive analysis capabilities of the large language model, this application can realize the correlation analysis of multiple dimensions such as sentiment tendencies, product problems, and improvement suggestions. In addition, this application uses a dual-path analysis architecture that combines the LLM model and the clustering model to ensure the efficiency of analysis while taking into account the depth and breadth of analysis. Through this intelligent analysis method, the cost of manual operation can be effectively reduced, the overall efficiency of live broadcast operations can be significantly improved, and solid data support can be provided for the comprehensive evaluation of live broadcast effects, ultimately helping to achieve more refined and scientific operational decisions.

[0111] In one embodiment, the present application also provides a computer-readable storage medium, which stores computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the live commentary analysis method as described in any of the above embodiments.

[0112] In one embodiment, the present application also provides a computer device, including: one or more processors, and a memory.

[0113] The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the live comment analysis method as described in any one of the above embodiments are performed.

[0114] Schematically, as Figure 5 As shown, Figure 5 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 300 can be provided as a server. Figure 5 Computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by memory 301 for storing instructions executable by processing component 302, such as an application. The application stored in memory 301 may include one or more modules, each corresponding to a set of instructions. In addition, processing component 302 is configured to execute the instructions to perform the live commentary analysis method of any of the above-mentioned embodiments.

[0115] The computer device 300 may further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate based on an operating system stored in the memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or the like.

[0116] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0117] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0118] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referenced to each other.

[0119] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A live comment analysis method, characterized in that: The method comprises: Obtaining original comments generated by the live broadcast platform during the live broadcast process, and obtaining comment information after pre-processing the original comments; After subject-clustering the review information, a plurality of clustering results are obtained, keywords in each clustering result are extracted, and each keyword is classified to obtain a classification result; The classification results of each keyword and the comment information in the cluster corresponding to each classification result are respectively input into a preset large language model to obtain analysis results corresponding to each classification result output by the large language model.

2. The live comment analysis method according to claim 1, characterized in that: After the original comments are preprocessed, comment information is obtained, including: Performing text cleaning on the original comments, and performing word segmentation processing on the cleaned original comments to obtain word segmentation results; After removing invalid characters from the word segmentation results, the word segmentation results after removing the invalid characters are normalized to obtain final comment information.

3. The live comment analysis method according to claim 1, characterized in that: After the review information is clustered by topic, multiple clustering results are obtained, including: Convert the review information into a semantic vector; Clustering the semantic vectors, and filtering noise data of the semantic vectors during the clustering process to obtain a plurality of clusters; After generating topic labels corresponding to each cluster, multiple clustering results are obtained.

4. The live comment analysis method according to claim 3, characterized in that: Before clustering the semantic vectors, the method further includes: Obtaining user basic information, live broadcast metadata, and comment interaction data corresponding to the original comment; Converting the user basic information, the live broadcast metadata, and the comment interaction data into a structured vector; The structured vector is fused with the semantic vector, and the fused semantic vector is clustered.

5. The live comment analysis method according to claim 3 or 4, characterized in that: After generating the topic labels corresponding to each cluster, multiple clustering results are obtained, including: For each cluster: Segment the semantic vectors in the cluster, and extract representative words from the cluster based on the segmentation results; After generating topic labels corresponding to the representative words, a clustering result of the cluster is obtained.

6. The live comment analysis method according to claim 1, characterized in that: The step of extracting keywords from each clustering result includes: For each clustering result: Use TF-IDF to calculate the frequency weight of each word in the clustering result and obtain the weight corresponding to each word; Stop word filtering is performed on each word in the clustering result, and the word with the highest weight is selected from the filtered words as the keyword.

7. The live comment analysis method according to claim 1, characterized in that: After classifying each keyword, the classification results are obtained, including: Get the preset rule library; Each keyword is matched with the preset rule library respectively, and a classification result corresponding to each keyword is determined according to the matching result.

8. A live comment analysis device, characterized in that: include: An information acquisition module is used to obtain original comments generated by the live broadcast platform during the live broadcast process, and obtain comment information after pre-processing the original comments; A topic clustering module is used to perform topic clustering on the review information to obtain multiple clustering results, extract keywords from each clustering result, and classify each keyword to obtain a classification result; The intelligent analysis module is used to input the classification results of each keyword and the comment information in the cluster corresponding to each classification result into a preset large language model, and obtain the analysis results corresponding to each classification result output by the large language model.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the live commentary analysis method as described in any one of claims 1 to 7.

10. A computer device, characterized in that: include: one or more processors, and memory; The memory stores computer-readable instructions, which, when executed by the one or more processors, execute the steps of the live commentary analysis method according to any one of claims 1 to 7.

Images