CN116882414A

CN116882414A - Automatic comment generation method and related device based on large-scale language model

Info

Publication number: CN116882414A
Application number: CN202311137360.1A
Authority: CN
Inventors: 孔奇松; 罗科林; 罗丹; 陈洁; 周庆; 刘琼
Original assignee: Shenzhen Amaqi Technology Co ltd
Current assignee: Shenzhen Amaqi Technology Co ltd
Priority date: 2023-09-05
Filing date: 2023-09-05
Publication date: 2023-10-13
Anticipated expiration: 2043-09-05
Also published as: CN116882414B

Abstract

The invention relates to the technical field of data processing, and discloses a method and a related device for automatically generating comments based on a large-scale language model, which are used for improving the accuracy of automatic generation of comments. Comprising the following steps: performing data division on the historical comment data to obtain a plurality of sub-data sets, performing semantic extraction on the plurality of sub-data sets to obtain a semantic information set, performing word embedding mapping to obtain a plurality of low-dimensional dense vectors, and performing clustering processing to obtain a semantic association relation; performing data conversion on the semantic association relationship to obtain an initial bag-of-words model, and performing topic number analysis on the initial bag-of-words model to determine the number of target topics; carrying out semantic distribution adjustment on the initial bag-of-words model to obtain a target bag-of-words model; and extracting the tag characteristics through the target word bag model to obtain target tag characteristics of each user, carrying out multi-mode fusion to obtain corresponding multi-mode fusion parameters, inputting the multi-mode fusion parameters into a preset comment generation model to generate comments, and obtaining target comments.

Description

Automatic comment generation method and related device based on large-scale language model

Technical Field

The invention relates to the technical field of data processing, in particular to a method and a related device for automatically generating comments based on a large-scale language model.

Background

In the contemporary information age, a large amount of user comments and semantic information are continuously generated, and the data have important values for user emotion, product evaluation, market trend and the like. Therefore, the development of intensive research on historical comment data is a challenging and practical task. At present, methods such as Natural Language Processing (NLP), word embedding technology, clustering algorithm and the like are widely applied to analysis and mining of text data.

Existing semantic extraction algorithms, although making great progress in semantic information extraction of text, still have limitations. For complex or implicit semantic relationships, traditional semantic extraction algorithms may not capture enough, resulting in insufficient comprehensiveness and accuracy of the extracted semantic information. Interpretive problem of word embedding mapping: word embedding techniques can map text data into a low-dimensional dense vector space, but the mapped vector lacks interpretability in some cases, and it is difficult to understand the specific meaning in the vector space, which affects the interpretation and understanding of the clustering results. The existing comment generation model has room for improvement in individuation and diversity. The generated comments are often too generalized, cannot well meet the personalized requirements of different users, and lack rich expression.

Disclosure of Invention

The invention provides a method and a related device for automatically generating comments based on a large-scale language model, which are used for improving the accuracy of automatically generating comments based on the large-scale language model.

The first aspect of the invention provides a method for automatically generating comments based on a large-scale language model, which comprises the following steps:

acquiring historical comment data, carrying out data division on the historical comment data to obtain a plurality of sub-data sets, and respectively carrying out semantic extraction on the plurality of sub-data sets through a semantic extraction algorithm to obtain a semantic information set, wherein the semantic information set comprises semantic information corresponding to each sub-data set;

based on the semantic information set, performing word embedding mapping on a plurality of sub-data sets to obtain a plurality of low-dimensional dense vectors, and performing clustering processing on the plurality of low-dimensional dense vectors to obtain semantic association relations;

performing data conversion on the semantic association relationship to obtain an initial bag-of-words model corresponding to the semantic association relationship, and performing topic number analysis on the initial bag-of-words model to determine the number of target topics;

Carrying out semantic distribution adjustment on the initial bag-of-words model based on the number of the target subjects to obtain a target bag-of-words model;

acquiring tag information sets of a plurality of users, inputting the tag information sets into the target word bag model for tag feature extraction, and obtaining target tag features of each user;

and carrying out multi-modal fusion on the target tag characteristics of each user and the semantic information set to obtain corresponding multi-modal fusion parameters, inputting the multi-modal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments.

With reference to the first aspect, in a first implementation manner of the first aspect of the present invention, the obtaining historical comment data, performing data division on the historical comment data to obtain a plurality of sub-data sets, and performing semantic extraction on the plurality of sub-data sets through a semantic extraction algorithm to obtain a semantic information set, where the semantic information set includes semantic information corresponding to each sub-data set, and includes:

acquiring the historical comment data, extracting invalid data from the historical comment data, and determining invalid comment data;

Performing data cleaning on the historical comment data based on the invalid comment data to obtain comment data to be processed;

performing time sequence division on the comment data to be processed to obtain a plurality of sub-data sets;

performing data type matching on a plurality of sub-data sets, determining a data type set, performing algorithm matching based on the data type set, and determining the semantic extraction algorithm;

extracting semantic vectors from a plurality of sub-data sets through the semantic extraction algorithm to obtain semantic fusion vectors corresponding to each sub-data set;

carrying out semantic representation analysis through semantic fusion vectors corresponding to each sub-data set to obtain target semantic representation information, and constructing the semantic information set through the semantic representation information.

With reference to the first aspect, in a second implementation manner of the first aspect of the present invention, performing word embedding mapping on the plurality of sub-data sets based on the semantic information set to obtain a plurality of low-dimensional dense vectors, and performing clustering processing on the plurality of low-dimensional dense vectors to obtain a semantic association relationship, where the method includes:

carrying out corpus matching through the semantic information set to determine a target corpus;

Respectively carrying out data preprocessing on each sub-data set to obtain a plurality of target sub-data sets;

generating word vectors for each target sub-data set respectively to obtain a word vector set corresponding to each target sub-data set;

word embedding mapping is carried out on a plurality of sub-data sets through a word vector set corresponding to each target sub-data set, so that a plurality of low-dimensional dense vectors are obtained;

carrying out data point mapping on a plurality of low-dimensional dense vectors through a preset hierarchical clustering algorithm to obtain a data point cluster;

performing distance analysis on each data point in the data point cluster to determine a distance data set;

and clustering based on the distance data set to obtain a semantic association relationship.

With reference to the first aspect, in a third implementation manner of the first aspect of the present invention, the converting the data of the semantic association relationship to obtain an initial bag-of-words model corresponding to the semantic association relationship, and performing topic number analysis on the initial bag-of-words model to determine a target topic number, where the method includes:

creating a vocabulary table for the semantic association relationship to obtain a target vocabulary table;

performing model construction based on the target vocabulary to obtain an initial bag-of-words model corresponding to the semantic association relationship;

Performing sparse matrix analysis on the initial bag-of-words model to obtain a target sparse matrix corresponding to the initial bag-of-words model;

inputting the target sparse matrix into the initial bag-of-words model for subject traversal analysis to obtain traversal analysis results;

and screening the number of the topics from the traversal analysis result based on a preset number threshold to obtain the number of the target topics.

With reference to the third implementation manner of the first aspect, in a fourth implementation manner of the first aspect of the present invention, the performing semantic distribution adjustment on the initial bag-of-words model based on the target topic number to obtain a target bag-of-words model includes:

performing topic name matching based on the target topic quantity to determine a plurality of topic names;

carrying out consistency score calculation based on a plurality of topic names to obtain a consistency score corresponding to each topic name;

and adjusting semantic distribution of the initial word bag model line through the consistency score corresponding to each theme name to obtain a target word bag model.

With reference to the first aspect, in a fifth implementation manner of the first aspect of the present invention, the obtaining a tag information set of a plurality of users, and inputting the tag information set into the target bag-of-word model to perform tag feature extraction, to obtain target tag features of each user, includes:

Acquiring tag information sets of a plurality of users, inputting the tag information sets into the target word bag model for numerical processing, and obtaining a plurality of tag numerical information;

constructing a numerical matrix of the plurality of tag numerical information to obtain a plurality of numerical matrices;

performing matrix classification on the numerical matrices to obtain a plurality of matrix types;

and extracting the tag characteristics based on a plurality of matrix types to obtain the target tag characteristics of each user.

With reference to the first aspect, in a sixth implementation manner of the first aspect of the present invention, the performing multi-modal fusion on the target tag feature of each user and the semantic information set to obtain corresponding multi-modal fusion parameters, inputting the multi-modal fusion parameters into a preset comment generation model to generate comments, and obtaining a target comment includes:

performing first bias parameter analysis on the target tag characteristics of each user, and determining a first bias parameter set;

performing second bias parameter analysis on the semantic information set to determine a second bias parameter set;

carrying out parameter weighted average calculation on the first bias parameter set and the second bias parameter set to obtain a target bias parameter;

Performing multi-mode fusion on the target tag characteristics of each user and the semantic information set through the target bias parameters to obtain corresponding multi-mode fusion parameters;

inputting the multimodal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments.

The second aspect of the present invention provides a comment automatic generation device based on a large-scale language model, the comment automatic generation device based on the large-scale language model comprising:

the acquisition module is used for acquiring historical comment data, carrying out data division on the historical comment data to obtain a plurality of sub-data sets, and respectively carrying out semantic extraction on the plurality of sub-data sets through a semantic extraction algorithm to obtain a semantic information set, wherein the semantic information set comprises semantic information corresponding to each sub-data set;

the mapping module is used for carrying out word embedding mapping on the plurality of sub-data sets based on the semantic information set to obtain a plurality of low-dimensional dense vectors, and carrying out clustering processing on the plurality of low-dimensional dense vectors to obtain semantic association relations;

the conversion module is used for carrying out data conversion on the semantic association relationship to obtain an initial bag-of-words model corresponding to the semantic association relationship, carrying out topic quantity analysis on the initial bag-of-words model, and determining the number of target topics;

The adjusting module is used for carrying out semantic distribution adjustment on the initial bag-of-words model based on the number of the target subjects to obtain a target bag-of-words model;

the extraction module is used for acquiring tag information sets of a plurality of users, inputting the tag information sets into the target bag-of-words model for tag feature extraction, and obtaining target tag features of each user;

and the fusion module is used for carrying out multi-modal fusion on the target tag characteristics of each user and the semantic information set to obtain corresponding multi-modal fusion parameters, and inputting the multi-modal fusion parameters into a preset comment generation model to generate comments to obtain target comments.

A third aspect of the present invention provides a comment automatic generation apparatus based on a large-scale language model, comprising: a memory and at least one processor, the memory having instructions stored therein; the at least one processor invokes the instructions in the memory to cause the large-scale language model-based comment automatic generation apparatus to perform the large-scale language model-based comment automatic generation method described above.

A fourth aspect of the present invention provides a computer-readable storage medium having instructions stored therein that, when executed on a computer, cause the computer to perform the above-described method for automatically generating comments based on a large-scale language model.

According to the technical scheme provided by the application, historical comment data are acquired, the historical comment data are subjected to data division to obtain a plurality of sub-data sets, semantic extraction is respectively carried out on the plurality of sub-data sets through a semantic extraction algorithm to obtain a semantic information set, word embedding mapping is carried out on the plurality of sub-data sets based on the semantic information set to obtain a plurality of low-dimensional dense vectors, and clustering processing is carried out on the plurality of low-dimensional dense vectors to obtain a semantic association relation; performing data conversion on the semantic association relationship to obtain an initial word bag model corresponding to the semantic association relationship, and performing topic number analysis on the initial word bag model to determine the number of target topics; carrying out semantic distribution adjustment on the initial bag-of-words model based on the number of target subjects to obtain a target bag-of-words model; acquiring tag information sets of a plurality of users, inputting the tag information sets into a target word bag model for tag feature extraction, and obtaining target tag features of each user; and carrying out multi-modal fusion on the target tag characteristics and the semantic information sets of each user to obtain corresponding multi-modal fusion parameters, inputting the multi-modal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments. In the scheme of the application, semantic extraction is carried out on the historical comment data to obtain a semantic information set, and word embedding mapping is carried out on a plurality of sub-data sets to obtain a low-dimensional dense vector. The vectors can better express semantic information of comment data, so that more accurate and rich tag features are extracted. Clustering is performed on a plurality of low-dimensional dense vectors, so that comment data similar to semantic information can be classified into one category. The method is favorable for forming semantic association relations, and the comments of similar subjects are gathered together, so that the subsequent target subject number analysis and semantic distribution adjustment are facilitated. The initial word bag model is obtained by carrying out data conversion on the semantic association relation obtained by clustering, and the number of the subjects of the initial model is analyzed, so that the number of the target subjects can be more accurately determined. This helps to generate a more representative target bag-of-words model. And carrying out semantic distribution adjustment on the initial bag-of-words model through the number of target subjects to obtain a target bag-of-words model with more semantic expression. The model can better capture the subject information and semantic information of the comment data, and improves the performance of the comment generation model. The target label characteristics and the semantic information sets of each user are subjected to multi-mode fusion to obtain multi-mode fusion parameters, and the multi-mode fusion parameters are input into a preset comment generation model, so that more personalized, accurate and semantically enriched target comments can be generated. The multi-mode fusion can comprehensively utilize different information sources, and improves the quality and diversity of comment generation.

Drawings

FIG. 1 is a schematic diagram of one embodiment of a method for automatically generating comments based on a large-scale language model according to an embodiment of the present invention;

FIG. 2 is a flow chart of word embedding mapping for multiple sub-data sets in accordance with an embodiment of the present invention;

FIG. 3 is a flowchart of data conversion of semantic association according to an embodiment of the present invention;

FIG. 4 is a flowchart of semantic distribution adjustment of an initial bag-of-words model based on a target topic number in an embodiment of the present invention;

FIG. 5 is a schematic diagram of an embodiment of a device for automatically generating comments based on a large-scale language model according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of an embodiment of a device for automatically generating comments based on a large-scale language model according to an embodiment of the present invention.

Detailed Description

The embodiment of the invention provides a method and a related device for automatically generating comments based on a large-scale language model, which are used for improving the accuracy of automatically generating comments based on the large-scale language model.

The terms "first," "second," "third," "fourth" and the like in the description and in the claims and in the above drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged where appropriate such that the embodiments described herein may be implemented in other sequences than those illustrated or otherwise described herein. Furthermore, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed or inherent to such process, method, article, or apparatus.

For ease of understanding, the following describes a specific flow of an embodiment of the present invention, referring to fig. 1, and one embodiment of a method for automatically generating comments based on a large-scale language model in the embodiment of the present invention includes:

s101, acquiring historical comment data, carrying out data division on the historical comment data to obtain a plurality of sub-data sets, and respectively carrying out semantic extraction on the plurality of sub-data sets through a semantic extraction algorithm to obtain a semantic information set, wherein the semantic information set comprises semantic information corresponding to each sub-data set;

it can be understood that the execution subject of the present invention may be a comment automatic generation device based on a large-scale language model, and may also be a terminal or a server, which is not limited herein. The embodiment of the invention is described by taking a server as an execution main body as an example.

In particular, the server obtains historical comment data, which may come from different sources, such as a student's classroom assessment, a product's user assessment, and so forth. And (3) extracting invalid data from the obtained historical comment data, and removing comment data which have no practical meaning or are noisy by the server. After the invalid data is extracted, the historical comment data is subjected to data cleaning, some special symbols, non-Chinese characters and the like are removed, and comment data to be processed are obtained. The method comprises the steps of carrying out time sequence division on comment data to be processed, and dividing the data into a plurality of sub-data sets according to time sequence, wherein each sub-data set represents comment data in a specific time period. The server can process comment data of different time periods respectively and capture semantic changes in the time dimension. And carrying out data type matching on the plurality of sub-data sets, and classifying the data according to the characteristics or the labels of the comment data, wherein the classification is according to disciplines, product types and the like. Based on the data type sets, the server selects an appropriate semantic extraction algorithm to extract semantic information from each of the sub-data sets. After the semantic extraction algorithm is selected, the server applies the algorithms to the plurality of sub-data sets respectively, so that the semantic vector corresponding to each sub-data set is obtained. These semantic vectors reflect information such as topics, emotions, etc. in the comment, which provide the basis for subsequent semantic representation analysis. The server performs semantic representation analysis through semantic vectors corresponding to each sub-data set, and adopts technologies such as cluster analysis, dimension reduction and the like to find semantic association relations among the comments. The server obtains target semantic representation information that will reflect semantic relationships between different time periods and types of comment data. And constructing a complete semantic information set by combining the target semantic representation information corresponding to each sub-data set. This set will contain semantic information for different time periods and types of comments in the historical comment data, providing useful clues and information for the comment generation model to generate target comments related to the user's tag and containing the desired subject matter. For example, assume that the server is to implement a student class comment generation system. The server first collects student class rating data for different subjects over the last few years. Invalid data extraction is performed on the data, and comments which are too short or do not contain valid information are removed. And (3) cleaning the rest data, and removing special symbols and non-Chinese characters in the rest data to obtain the comment data to be processed. And carrying out time sequence division on the comment data to be processed according to the academic period and the discipline to obtain a plurality of sub-data sets, wherein each sub-data set corresponds to comment data of a specific discipline in one academic period. For each sub-dataset, the server performs data type matching according to disciplines, and then selects the appropriate semantic extraction algorithm. It is assumed that the server uses a topic modeling algorithm (e.g., LDA) for the data of the semantic evaluation and an emotion analysis algorithm for the data of the semantic evaluation. Through these algorithms, the server extracts semantic vectors from each sub-dataset representing topics or emotion information in the individual comments. The server performs semantic representation analysis on the semantic vector of each sub-data set, clusters together the comments of similar subjects by using a clustering algorithm, and visualizes semantic association relations among the comments by dimension reduction. The server gathers together the corresponding target semantic representation information of each sub-data set, and constructs a semantic information set containing different scholars and disciplines. This set provides valuable information for subsequent comment generation, ensuring that the generated comment content is relevant to the user tag and contains the intended topic.

S102, carrying out word embedding mapping on a plurality of sub-data sets based on a semantic information set to obtain a plurality of low-dimensional dense vectors, and carrying out clustering processing on the plurality of low-dimensional dense vectors to obtain a semantic association relation;

specifically, the server performs corpus matching according to the semantic information set, and determines a target corpus. This target corpus should contain enough text data to generate word vectors with rich semantic information. And respectively carrying out data preprocessing on each sub-data set to obtain a plurality of target sub-data sets. In the data preprocessing stage, the server performs operations such as word segmentation, stop word removal, part-of-speech tagging and the like to prepare data for word vector generation. Word vector generation is performed on each target sub-dataset, which can be accomplished by commonly used Word embedding algorithms such as Word2Vec, gloVe, or FastText. These algorithms can map words to a low-dimensional dense vector space, capturing semantic relationships between words. And using the word vector set corresponding to each target sub-data set for word embedding mapping on the plurality of sub-data sets to obtain a plurality of low-dimensional dense vectors. These vectors represent the semantic representation of the words in each sub-dataset in the target corpus. And carrying out data point mapping on the plurality of low-dimensional dense vectors through a preset hierarchical clustering algorithm to obtain a data point cluster. Hierarchical clustering is an efficient clustering algorithm that gradually merges data points into clusters of different hierarchies. Each data point in the cluster of data points is subjected to a distance analysis to determine a distance data set. This distance data set reflects similarities or correlations between data points, helping to discover potential semantic associations. And clustering based on the distance data set to obtain a semantic association relationship. This clustering process will identify closely related data points in the data point clusters, revealing semantic associations between sub-data sets. For example, assume that the server is to generate comments based on user ratings data for different movie types. The server collects user rating data from different time periods and movie types and performs data cleaning and preprocessing. The server determines a target corpus containing rich movie related text. For each sub-data set, the server performs Word vector generation, such as using Word2Vec algorithm, respectively. And obtaining the low-dimensional dense vector corresponding to each sub data set by the server through word embedding mapping. The server clusters the vectors through a hierarchical clustering algorithm to obtain a data point cluster. Further, the server performs a distance analysis on each data point in the data point cluster to obtain a distance data set. And clustering based on the distance data set, obtaining a semantic association relationship by the server, and finding the semantic relationship among different film types in the user evaluation data, so as to generate a more targeted comment.

S103, performing data conversion on the semantic association relationship to obtain an initial bag-of-words model corresponding to the semantic association relationship, and performing topic quantity analysis on the initial bag-of-words model to determine the number of target topics;

specifically, the vocabulary is created for the semantic association relationship, and a target vocabulary is obtained from the vocabulary. A vocabulary is a collection that contains all the words that appear in a semantic association. And carrying out model construction based on the target vocabulary to obtain an initial bag-of-words model corresponding to the semantic association relationship. The bag of words model is a common method of text representation that represents text as the number of occurrences of words in a document. And carrying out sparse matrix analysis on the initial bag-of-words model to obtain a target sparse matrix corresponding to the initial bag-of-words model. A sparse matrix is a matrix with a majority of elements being zero, which can more effectively represent large-scale text data. And inputting the target sparse matrix into the initial bag-of-words model to perform topic traversal analysis. Topic traversal is a technique for finding topic structures, i.e., grouping together associated words, in a bag of words model. And obtaining a candidate set of the topics by the server through the results of the topic traversal analysis. These candidate topics are formed by combining similar words into topics based on relevance information in the text data. And screening the number of the topics from the traversal analysis result based on a preset number threshold. The server determines the final number of target topics based on certain criteria or metrics. For example, assume that the server has a set of movie ratings data, and the server has obtained a semantic association and word vector representation. The server performs vocabulary creation on the text data to obtain a target vocabulary containing all words. The server builds an initial bag of words model using the target vocabulary, representing each text as the number of occurrences of the term in the document. And carrying out sparse matrix analysis on the initial bag-of-words model to obtain a target sparse matrix, wherein the matrix can more efficiently represent text data. Inputting the target sparse matrix into an initial bag-of-words model for subject traversal analysis, and obtaining a series of candidate subjects by the server. The server screens the number of topics based on a preset number threshold. Such as the server setting a minimum number of words for a topic or a relevance score threshold for a topic. By applying these thresholds, the server determines the final number of target topics. The server obtains an initial bag of words model based on semantic association and determines the number of target topics that are appropriate, which will provide valuable information and guidance for subsequent comment generation.

S104, carrying out semantic distribution adjustment on the initial word bag model based on the number of target subjects to obtain a target word bag model;

specifically, according to the number of target topics, matching the topic names is performed, and a plurality of topic names are determined. These topic names represent different topics or topics, for example, in movie rating data, there may be topics such as "scenario", "performance", "visual effect", etc. And carrying out consistency score calculation based on the plurality of topic names to obtain a consistency score corresponding to each topic name. The consistency score is used to measure the consistency and stability of a topic throughout the data set, and can help the server determine which topics are more representative in the data. And carrying out semantic distribution adjustment on the initial word bag model through the consistency scores corresponding to the topic names to obtain a target word bag model. The adjustment process adjusts the weight distribution of different words in the word bag model according to the consistency scores of the topics, so that the target word bag model is more in line with the semantic structure of the data. For example, assume that the server has a set of movie ratings data, an initial bag of words model and a target number of topics have been obtained. The server performs semantic distribution adjustment. The server performs topic name matching based on the number of target topics, assuming that the server determines three topic names, namely "scenario", "performance" and "visual effect". And the server calculates the consistency score corresponding to each topic name by using a consistency score calculation method. This score reflects how consistent and stable the topic represented by these topic names is throughout the dataset. For example, if a "scenario" topic appears more frequently and is more consistently distributed in a dataset, its consistency score will be higher. And the server performs semantic distribution adjustment on the initial bag-of-words model through the consistency scores obtained through calculation. Assuming that the consistency score of a "scenario" topic is higher, words related to the "scenario" are given higher weight in the target bag of words model, thereby better reflecting the importance and representativeness of the topic in the data. Similarly, for other topics, the server will also adjust accordingly based on their consistency scores so that the target bag of words model will more accurately represent the semantic structure of the entire dataset. Through such semantic distribution adjustment, the server obtains a target bag-of-words model which will more conform to the semantic features of the data and provide more accurate and meaningful results for subsequent comment generation.

S105, acquiring tag information sets of a plurality of users, and inputting the tag information sets into a target word bag model to extract tag characteristics so as to obtain target tag characteristics of each user;

specifically, a set of tag information of a plurality of users is obtained, and the tags can be marks or descriptions of a certain theme, product or event by the users. And inputting the label information set into a target bag-of-word model for numerical processing to obtain a plurality of label numerical information. The server converts the text-form labels into numerical representations for subsequent numerical matrix construction and processing. And constructing a numerical matrix for the numerical information of the plurality of labels to obtain a plurality of numerical matrixes. Each numerical matrix represents label information for a user, wherein the rows represent different labels and the columns represent different numerical features or attributes. And performing matrix classification on the numerical matrices to obtain a plurality of matrix types. The matrix classification is used for identifying the structure and characteristics among different user tag information, so that the target tag characteristics of the user are better extracted. And extracting the tag characteristics based on a plurality of matrix types to obtain the target tag characteristics of each user. And extracting tag features with more representativeness and distinguishing property according to the distribution and correlation of the user tag information in different matrix types. For example, assume that the server has a set of tag information for different movies, such as "comedy", "action", "scenario", etc. The server wishes to extract the target tag characteristics for each user to better understand their preferences and interests. The server gathers a collection of tag information for a plurality of users, each labelling a different movie. The server inputs the tag information into the target bag-of-words model, and converts the text tag into a numerical representation, such as using word frequency or TF-IDF. And constructing a numerical matrix of the numerical information of the labels to obtain the numerical matrix of each user, wherein the rows represent different movie labels and the columns represent numerical characteristics. For each user's numerical matrix, the server classifies the matrix, finding that some users prefer to mark comedy movies, while others prefer action, resulting in different matrix types. Based on these matrix types, the server performs tag feature extraction, finding that some users tag more movies in comedy types and less in action types. The server can extract the target tag characteristics of each user and learn about their movie preferences and interests.

S106, carrying out multi-modal fusion on the target tag characteristics and the semantic information sets of each user to obtain corresponding multi-modal fusion parameters, and inputting the multi-modal fusion parameters into a preset comment generation model to generate comments to obtain target comments.

It should be noted that, the first bias parameter analysis is performed on the target tag feature of each user, and the first bias parameter set is determined. The aim is to extract information of representative and importance from the target tag features. For example, in movie rating data, the server analyzes the user's preference level for different movie types, and thus obtains a first bias parameter set, which represents the user's importance level for different types of tag features. And carrying out second bias parameter analysis on the semantic information set to determine a second bias parameter set. The purpose is to extract information related to the target task (comment generation) from the semantic information set. For example, in movie ratings data, the server analyzes which semantic information is more relevant to the generation of ratings content, resulting in a second set of bias parameters representing the importance of different semantic information. And carrying out parameter weighted average calculation on the first bias parameter set and the second bias parameter set to obtain a target bias parameter. And obtaining the comprehensive target bias parameters by carrying out weighted average on the first bias parameter set and the second bias parameter set. The weight of the weighted average can be set according to the task requirement and the actual situation. And carrying out multi-mode fusion on the target tag characteristics of each user and the semantic information set through the target bias parameters to obtain corresponding multi-mode fusion parameters. The importance degree of the user on the tag characteristics and the semantic information is considered, so that the fusion of the multi-mode data is realized. The specific fusion method can be weighted average, feature stitching and the like. Inputting the multimodal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments. The server passes the multimodal fusion parameters as input to a pre-set comment generation model that will use these parameters and other relevant information to generate the final target comment. For example, assume that the server has a set of user-to-movie tag information and corresponding semantic information sets. The server wishes to generate a comment for the movie for each user based on this information. The server analyzes the first bias parameters of the target label characteristics of each user, and discovers that the user pays more attention to the two labels of the scenario and the performance. And carrying out second bias parameter analysis on the semantic information set, and finding that the emotion information in the comment is more critical to comment generation. And obtaining a target bias parameter by carrying out weighted average calculation on the first bias parameter set and the second bias parameter set, wherein for example, the weight of the first bias parameter set is 0.6, and the weight of the second bias parameter set is 0.4. And carrying out multi-mode fusion on the target tag characteristics and the semantic information sets of each user according to the target bias parameters. This means that the server will pay more attention to the user's tendency to "scenario" and "performance" labels, as well as to the affective information in the comments, when generating the comments. Inputting the multimodal fusion parameters into a preset comment generation model, wherein the model comprehensively considers target tag characteristics, semantic information and other influencing factors of a user to generate a final target comment. Through multi-mode fusion and the introduction of bias parameters, the server better personalizes and customizes the comment generation process, so that the generated comments better meet the preference and the requirement of the user.

In the embodiment of the application, history comment data are acquired, data division is carried out on the history comment data to obtain a plurality of sub-data sets, semantic extraction is carried out on the plurality of sub-data sets respectively through a semantic extraction algorithm to obtain a semantic information set, word embedding mapping is carried out on the plurality of sub-data sets based on the semantic information set to obtain a plurality of low-dimensional dense vectors, and clustering processing is carried out on the plurality of low-dimensional dense vectors to obtain a semantic association relation; performing data conversion on the semantic association relationship to obtain an initial word bag model corresponding to the semantic association relationship, and performing topic number analysis on the initial word bag model to determine the number of target topics; carrying out semantic distribution adjustment on the initial bag-of-words model based on the number of target subjects to obtain a target bag-of-words model; acquiring tag information sets of a plurality of users, inputting the tag information sets into a target word bag model for tag feature extraction, and obtaining target tag features of each user; and carrying out multi-modal fusion on the target tag characteristics and the semantic information sets of each user to obtain corresponding multi-modal fusion parameters, inputting the multi-modal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments. In the scheme of the application, semantic extraction is carried out on the historical comment data to obtain a semantic information set, and word embedding mapping is carried out on a plurality of sub-data sets to obtain a low-dimensional dense vector. The vectors can better express semantic information of comment data, so that more accurate and rich tag features are extracted. Clustering is performed on a plurality of low-dimensional dense vectors, so that comment data similar to semantic information can be classified into one category. The method is favorable for forming semantic association relations, and the comments of similar subjects are gathered together, so that the subsequent target subject number analysis and semantic distribution adjustment are facilitated. The initial word bag model is obtained by carrying out data conversion on the semantic association relation obtained by clustering, and the number of the subjects of the initial model is analyzed, so that the number of the target subjects can be more accurately determined. This helps to generate a more representative target bag-of-words model. And carrying out semantic distribution adjustment on the initial bag-of-words model through the number of target subjects to obtain a target bag-of-words model with more semantic expression. The model can better capture the subject information and semantic information of the comment data, and improves the performance of the comment generation model. The target label characteristics and the semantic information sets of each user are subjected to multi-mode fusion to obtain multi-mode fusion parameters, and the multi-mode fusion parameters are input into a preset comment generation model, so that more personalized, accurate and semantically enriched target comments can be generated. The multi-mode fusion can comprehensively utilize different information sources, and improves the quality and diversity of comment generation.

In a specific embodiment, the process of executing step S101 may specifically include the following steps:

(1) Acquiring historical comment data, extracting invalid data from the historical comment data, and determining the invalid comment data;

(2) Performing data cleaning on the historical comment data based on the invalid comment data to obtain comment data to be processed;

(3) Performing time sequence division on comment data to be processed to obtain a plurality of sub-data sets;

(4) Performing data type matching on the plurality of sub-data sets, determining a data type set, performing algorithm matching based on the data type set, and determining a semantic extraction algorithm;

(5) Respectively extracting semantic vectors from a plurality of sub-data sets through a semantic extraction algorithm to obtain semantic fusion vectors corresponding to each sub-data set;

(6) Carrying out semantic representation analysis through semantic fusion vectors corresponding to each sub-data set to obtain target semantic representation information, and constructing a semantic information set through the semantic representation information.

In particular, historical comment data is obtained from a database or other data source. These data contain user ratings and comments on different types of merchandise, movies, restaurants, etc. Invalid data extraction is performed on the historical comment data, and the invalid comment data is identified and removed, such as data containing messy codes, incomplete data or data irrelevant to tasks. And cleaning the historical comment data based on the invalid comment data, removing the invalid comment, and reserving the valid comment data to obtain the comment data to be processed. The time series division is carried out on the comment data to be processed, the comment data is divided into a plurality of sub-data sets according to the time sequence, for example, the comment data are divided according to the day, the week or the month, and the comment data in different time periods can be analyzed and processed. And carrying out data type matching on the plurality of sub data sets, determining a data type set, carrying out algorithm matching based on the data type set, and selecting a proper semantic extraction algorithm. Different comment data contains different types of information, such as emotional tendency, subject matter, etc., and therefore different semantic extraction algorithms need to be selected according to the data type. And respectively extracting semantic vectors from the plurality of sub-data sets through the selected semantic extraction algorithm to obtain semantic fusion vectors corresponding to each sub-data set. These vectors may represent semantic information in each sub-dataset, such as emotional polarity, key topics, etc. And carrying out semantic representation analysis through the semantic fusion vector corresponding to each sub-data set to obtain target semantic representation information. The semantic fusion vector is comprehensively analyzed, higher-level semantic information is extracted from the semantic fusion vector, and for example, emotion tendencies, topic distribution and the like of the whole data set are analyzed. And constructing a semantic information set through the semantic representation information, and combining the target semantic representation information of each sub-data set to form a complete semantic information set. This set will contain the semantic features and dependencies of all sub-data sets, providing an important reference for subsequent comment generation. For example, assume that the server has a set of user ratings data for different restaurants, each rating containing ratings, comment text, and time information. The server wishes to perform the task of automatically generating comments on these evaluations. Historical comment data is obtained from the database, including the user's score and comment text. Invalid data extraction is performed on the historical comment data, for example, data with unreasonable scores or messy codes in comment texts are identified and eliminated. And cleaning the data based on the invalid comment data, and reserving the valid evaluation data to obtain the comment data to be processed. And carrying out time sequence division on the comment data to be processed according to time sequence, for example, division according to month, so as to obtain a plurality of sub-data sets. For these sub-data sets, the server performs data type matching, and finds that some data sets mainly contain emotional tendency of users, while other data sets pay more attention to specific evaluation contents of dishes and services. Based on the different data types, the server selects a suitable semantic extraction algorithm, for example, for emotion-prone data, the server uses an emotion analysis algorithm, and for specific evaluation content, the server uses a topic extraction algorithm. And extracting semantic vectors from each sub-data set through a selected semantic extraction algorithm to obtain semantic fusion vectors corresponding to each sub-data set, wherein the vectors can represent emotion and theme information of each sub-data set. And the server performs semantic representation analysis on the semantic fusion vectors, for example, comprehensively considering the relation between emotion and theme, and obtaining target semantic representation information. And constructing a semantic information set through the target semantic representation information, wherein the semantic information set comprises semantic features and relativity of each sub-data set, and providing references for subsequent comment generation.

In a specific embodiment, as shown in fig. 2, the process of executing step S102 may specifically include the following steps:

s201, corpus matching is carried out through a semantic information set, and a target corpus is determined;

s202, respectively carrying out data preprocessing on each sub-data set to obtain a plurality of target sub-data sets;

s203, generating word vectors for each target sub-data set respectively to obtain a word vector set corresponding to each target sub-data set;

s204, word embedding mapping is carried out on the plurality of sub-data sets through the word vector set corresponding to each target sub-data set, and a plurality of low-dimensional dense vectors are obtained;

s205, carrying out data point mapping on a plurality of low-dimensional dense vectors through a preset hierarchical clustering algorithm to obtain a data point cluster;

s206, performing distance analysis on each data point in the data point cluster to determine a distance data set;

s207, clustering is conducted based on the distance data set, and semantic association relations are obtained.

It should be noted that the server determines the target corpus, which may be a database or a corpus containing a large amount of text data, including the corpus related to the comment data of the server. Data preprocessing is performed on each sub-data set, which includes text word segmentation, stop word removal, word drying, etc., for subsequent word vector generation and word embedding mapping. Word vector generation is performed on each target sub-dataset, and Word bag models, word2Vec, gloVe and other algorithms can be used for converting text into vector representations, so that a Word vector set corresponding to each target sub-dataset is obtained. Word embedding mapping is carried out on a plurality of sub-data sets through a word vector set corresponding to each target sub-data set, and texts in different data sets are converted into unified low-dimensional dense vector representation. The word embedding mapping technology can map words in different data sets to the same vector space, so that subsequent semantic association analysis is facilitated. And carrying out data point mapping on the plurality of low-dimensional dense vectors through a preset hierarchical clustering algorithm, and clustering similar text vectors together to form a data point cluster. Hierarchical clustering is a common unsupervised learning algorithm that can cluster according to the similarity between text vectors. And (3) carrying out distance analysis on each data point in the data point cluster to determine the distance between similar texts, wherein measuring methods such as Euclidean distance, cosine similarity and the like can be used. Clustering is carried out based on the distance data set, and data points with closer distances are classified into the same category, so that semantic association relations are obtained. Similar comment data is identified according to the distance data set, so that semantic association relations are formed. For example, assume that the server's comment data is a movie comment and the target corpus is a corpus containing a large amount of movie-related text. The server performs data preprocessing on the comment data, and performs word segmentation, stop word removal and other processing on the text. Aiming at the target corpus, the server generates word vectors to obtain word vector sets corresponding to each target sub-data set. And the server maps the text vector in the comment data to a word vector set of the target corpus to obtain a low-dimensional dense vector. The vectors are clustered through a hierarchical clustering algorithm to form a data point cluster. In the data point cluster, the server performs distance analysis to determine the distance between similar texts. Clustering is carried out based on the distance data set, and data points with closer distances are classified into the same category, so that semantic association relations among movie reviews are obtained. These associations may help the server discover similar topics or emotional trends in the comment data to better understand the user's ratings for movies.

In a specific embodiment, as shown in fig. 3, the process of executing step S103 may specifically include the following steps:

s301, creating a vocabulary table for the semantic association relationship to obtain a target vocabulary table;

s302, performing model construction based on a target vocabulary to obtain an initial word bag model corresponding to the semantic association relation;

s303, carrying out sparse matrix analysis on the initial bag-of-words model to obtain a target sparse matrix corresponding to the initial bag-of-words model;

s304, inputting the target sparse matrix into an initial bag-of-words model for subject traversal analysis to obtain traversal analysis results;

s305, screening the number of topics from the traversal analysis result based on a preset number threshold to obtain the number of target topics.

It should be noted that vocabulary creation is performed on semantic association relationships. Extracting all related vocabularies from the semantic association relationship, and constructing a target vocabulary. The target vocabulary consisting of high-frequency vocabularies can be obtained by counting the occurrence frequency of all vocabularies in the semantic association relationship and removing stop words and low-frequency vocabularies. Model construction is based on the target vocabulary, which may be a bag of words model. The bag of words model is a common method of text representation that converts text into a vector representation of words. And mapping the text in the semantic association relationship into a vector in the target vocabulary to obtain an initial word bag model corresponding to the semantic association relationship. And carrying out sparse matrix analysis on the initial bag-of-words model to obtain a target sparse matrix corresponding to the initial bag-of-words model. Sparse matrix is a special matrix representation method, and is commonly used for representing sparse features of text in a bag-of-word model. And obtaining sparse characteristic representation of the initial bag-of-words model on the target vocabulary through sparse matrix analysis. Inputting the target sparse matrix into the initial bag-of-words model for subject traversal analysis to obtain traversal analysis results. The topic traversal is a common text clustering and topic extraction method, and a hidden topic structure in the text can be obtained by performing traversal analysis on the text. In this step, the server discovers topic information in the semantic association by using a topic traversal method. And screening the number of the topics from the traversal analysis result based on a preset number threshold to obtain the number of the target topics. In the result of the topic traversal analysis, a plurality of topic candidates are obtained, and the number of topics which are relevant to the task and have proper number can be screened out by setting a preset number threshold. This results in a target number of topics. For example, assume that the server has a semantic association of movie ratings, including the content of ratings of users for different movies. The server wishes to extract the subject information from these ratings, knowing which aspects the user's ratings for movies are focused on. And the server creates a vocabulary table for the semantic association relation, counts the vocabulary in all evaluation texts, removes low-frequency vocabulary and constructs a target vocabulary table. A bag of words model is constructed based on the target vocabulary, and each evaluation text is mapped to a vector representation in the target vocabulary. And carrying out sparse matrix analysis on the word bag model to obtain sparse feature representation of each text on the target vocabulary. And inputting the sparse matrix into a subject traversal analysis to obtain a potential evaluation subject. And screening out evaluation topics which are related to the tasks and have proper quantity according to a preset topic quantity threshold value to obtain the target topic quantity.

In a specific embodiment, as shown in fig. 4, the process of executing step S104 may specifically include the following steps:

s401, performing topic name matching based on the number of target topics to determine a plurality of topic names;

s402, carrying out consistency score calculation based on a plurality of topic names to obtain a consistency score corresponding to each topic name;

s403, adjusting semantic distribution of the initial word bag model line through the consistency score corresponding to each topic name to obtain a target word bag model.

Specifically, topic name matching is performed based on the number of target topics, and a plurality of topic names are determined. The names of the topics are obtained from the target number of topics, and each topic can be given a representative name according to task requirements or domain knowledge. And carrying out consistency score calculation based on the plurality of topic names to obtain a consistency score corresponding to each topic name. The consistency score is an index that measures the degree of consistency between a topic name and the corresponding topic content. Natural language processing techniques, such as word vector similarity, topic models, and the like, may be used to calculate a consistency score between topic names and topic content. And carrying out semantic distribution adjustment on the initial bag-of-words model through the consistency score corresponding to each topic name to obtain a target bag-of-words model. And adjusting semantic distribution in the initial bag-of-words model according to the consistency score so that the topic names are more accurately matched with the corresponding topic contents. For example, assume that the server has an initial bag-of-words model of movie reviews that contains user ratings for movies. The server hopes to extract the topic names according to the number of the target topics, calculates the consistency scores between the topic names and the topic contents, and then carries out semantic distribution adjustment on the initial word bag model to obtain a more accurate target word bag model. The server sets the number of target topics to 3, i.e. it wants to extract 3 topics from movie reviews. Performing topic name matching based on the number of target topics, and respectively giving the following names to the 3 topics by the server: "scenario evaluation", "actor performance" and "visual effect". The server uses natural language processing techniques to calculate a consistency score between each topic name and the corresponding topic content. For example, the server uses word vector similarity to measure the degree of similarity between keywords in the topic names and keywords in the corresponding topic content, resulting in a consistency score. And according to the calculated consistency scores, the server performs semantic distribution adjustment on the initial bag-of-words model. For example, for the topic of "scenario evaluation," if the vocabulary distribution related to "scenario evaluation" in the initial bag-of-words model is found to be not accurate enough, the server increases the vocabulary weight related to "scenario evaluation" to adjust the semantic distribution of the topic in the target bag-of-words model.

In a specific embodiment, the process of executing step S105 may specifically include the following steps:

(1) Acquiring tag information sets of a plurality of users, inputting the tag information sets into a target word bag model for numerical processing, and obtaining a plurality of tag numerical information;

(2) Constructing a numerical matrix of the numerical information of the plurality of labels to obtain a plurality of numerical matrixes;

(3) Performing matrix classification on the numerical matrices to obtain a plurality of matrix types;

(4) And extracting the tag characteristics based on a plurality of matrix types to obtain the target tag characteristics of each user.

Specifically, a tag information set of a plurality of users is acquired. The tag information may be a keyword or phrase that a user describes, classifies, or evaluates something or content. For example, in movie rating data, tag information may include a type of movie, actor performance, scenario rating, etc. by a user. And inputting the label information set into a target bag-of-word model for numerical processing to obtain a plurality of label numerical information. The process of digitizing is a process of converting tag information into numerical features, and a common method is to use a bag-of-words model or a word vector representation. The bag of words model represents each tag as a vector, and each dimension of the vector represents a word whose value is the number of times the word appears in the tag. Word vector representation then represents each tag as a dense vector, mapping the vocabulary into a high-dimensional space. By such a numerical process, the server converts the tag information into a numerical form that can be processed by the computer. And constructing a numerical matrix for the numerical information of the plurality of labels to obtain a plurality of numerical matrixes. Each numerical matrix represents label information of a user, rows of the matrix correspond to different labels, and columns correspond to dimensions of words or word vectors of the word bag model. The elements in the matrix represent the number of occurrences of each word in each tag or the value of the word vector. And performing matrix classification on the numerical matrices to obtain a plurality of matrix types. Matrix classification is the process of classifying a plurality of numerical matrices according to their characteristics. The matrix of numbers may be classified using a clustering algorithm or classifier to classify matrices with similar features into the same type. Different matrix types represent different types of users or tag information, e.g. one type represents users like action movies and another type represents users like love movies. And extracting the tag characteristics based on a plurality of matrix types to obtain the target tag characteristics of each user. For each matrix type, the target tag features of the type of users can be obtained through methods such as statistics, clustering, feature extraction and the like. These target tag characteristics may be the user's preference for different types of tags, comment content for a particular tag, etc. For example, assume that the server has a data set of movie ratings, which contains tag information of a plurality of users for different movies, such as "action", "love", "actor's performance", etc. The server wishes to extract the target tag characteristics from the tag information of the user to understand the user's preferences and ratings of different types of movies. The server acquires tag information sets of a plurality of users, and each user corresponds to a list containing tags. And inputting the label information sets into a target bag-of-words model for numerical processing to obtain a plurality of label numerical information, and obtaining the bag-of-words model or the word vector representation corresponding to each user. And constructing numerical matrix of the plurality of tag numerical information, wherein each matrix represents the tag information of one user. For example, if there are 3 users, and the corresponding tag list is [ "action", "comedy" ], [ "love", "scenario" ], [ "action", "love" ], three numerical matrices can be constructed as follows: numerical matrix for user 1: action, comedy, love, scenario; tag 1: [1, 0]; tag 2: [0, 1]. Numerical matrix for user 2: action, comedy, love, scenario; tag 1: [0, 1]; tag 2: [1, 0]. Numerical matrix for user 3: action, comedy, love, scenario; tag 1: [1, 0, 1, 0]; tag 2: [0, 1, 0, 1]. The matrix classification is performed on the numerical matrices, and the server is assumed to classify the matrices by using a clustering algorithm to obtain two matrix types, which respectively represent users who like action movies and users who like love movies. And extracting the tag characteristics based on the two matrix types. For users who like action movies, the server will find that they have a higher preference for "action" tags and a lower preference for "comedy" and "love" tags, and will mention more of the content related to the action in the comment. And for users who like love movies, content related to love will be mentioned more in comments. The server successfully extracts the target tag characteristics of each user and provides valuable information for subsequent comment generation and personalized recommendation.

In a specific embodiment, the process of executing step S106 may specifically include the following steps:

(1) Performing first bias parameter analysis on target tag characteristics of each user, and determining a first bias parameter set;

(2) Performing second bias parameter analysis on the semantic information set to determine a second bias parameter set;

(3) Carrying out parameter weighted average calculation on the first bias parameter set and the second bias parameter set to obtain a target bias parameter;

(4) Performing multi-mode fusion on the target tag characteristics and the semantic information sets of each user through the target bias parameters to obtain corresponding multi-mode fusion parameters;

(5) Inputting the multimodal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments.

Specifically, a first bias parameter analysis is performed on the target tag characteristics of each user, and a first bias parameter set is determined. Information of representative and importance is extracted from the target tag features. For example, in movie rating data, the server analyzes the user's preference level for different movie types, and thus obtains a first bias parameter set, which represents the user's importance level for different types of tag features. And carrying out second bias parameter analysis on the semantic information set to determine a second bias parameter set. Information related to the target task (comment generation) is extracted from the semantic information set. For example, in movie ratings data, the server analyzes which semantic information is more relevant to the generation of ratings content, resulting in a second set of bias parameters representing the importance of different semantic information. And carrying out parameter weighted average calculation on the first bias parameter set and the second bias parameter set to obtain the target bias parameter. And obtaining the comprehensive target bias parameters by carrying out weighted average on the first bias parameter set and the second bias parameter set. The weight of the weighted average can be set according to the task requirement and the actual situation. And carrying out multi-mode fusion on the target tag characteristics and the semantic information sets of each user through the target bias parameters to obtain corresponding multi-mode fusion parameters. The importance degree of the user on the tag characteristics and the semantic information is considered, so that the fusion of the multi-mode data is realized. The specific fusion method can be weighted average, feature stitching and the like. Inputting the multimodal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments. And generating a personalized target comment by using the fused multi-modal parameters as input and combining a comment generation model. The comment generation model can be a natural language processing model based on a neural network, and generates comments conforming to the characteristics and semantic meanings of the user according to the input multimodal parameters and other context information. For example, assume that the server has a data set of movie ratings, which includes tag information of a plurality of users for different movies and comment sentences for movies. The server wishes to generate personalized comments based on the user's tag characteristics and comment content. The server performs a first bias parameter analysis on the tag characteristics of each user, e.g., find user a more interested in "action" type tags, and user B more prefers "love" type tags. A second bias parameter analysis is performed on the set of semantic information, for example, to determine which semantic information is more important in the generation of the ratings, such as actor performance and scenario development of the movie. And obtaining a target bias parameter through weighted average calculation, wherein the weight considers the preference of the individual user and the importance of semantic information. And applying the target bias parameters to the tag characteristics and semantic information sets of each user, and carrying out multi-mode fusion. For example, user a's "action" type tab feature and the actor performance of a movie are taken as important inputs, while user B is more focused on the "love" type tab feature and the drama development of the movie. Inputting the multimodal fusion parameters into a preset comment generation model to generate personalized target comments. Through multi-mode fusion and personalized comment generation, the server better understands the preference and evaluation of the user and provides comments more in line with the interests of the user.

The description of the automatic comment generating method based on the large-scale language model in the embodiment of the present invention is given above, and the description of the automatic comment generating device based on the large-scale language model in the embodiment of the present invention is given below, referring to fig. 5, one embodiment of the automatic comment generating device based on the large-scale language model in the embodiment of the present invention includes:

the acquiring module 501 is configured to acquire historical comment data, perform data division on the historical comment data to obtain a plurality of sub-data sets, and perform semantic extraction on the plurality of sub-data sets through a semantic extraction algorithm to obtain a semantic information set, where the semantic information set includes semantic information corresponding to each sub-data set;

the mapping module 502 is configured to perform word embedding mapping on the multiple sub-data sets based on the semantic information set to obtain multiple low-dimensional dense vectors, and perform clustering processing on the multiple low-dimensional dense vectors to obtain a semantic association relationship;

the conversion module 503 is configured to perform data conversion on the semantic association relationship to obtain an initial bag-of-words model corresponding to the semantic association relationship, and perform topic number analysis on the initial bag-of-words model to determine a target topic number;

The adjusting module 504 is configured to perform semantic distribution adjustment on the initial bag-of-words model based on the number of target topics, to obtain a target bag-of-words model;

the extracting module 505 is configured to obtain tag information sets of a plurality of users, and input the tag information sets into the target bag-of-word model to perform tag feature extraction, so as to obtain target tag features of each user;

and the fusion module 506 is configured to perform multi-modal fusion on the target tag feature of each user and the semantic information set to obtain corresponding multi-modal fusion parameters, and input the multi-modal fusion parameters into a preset comment generation model to generate comments, so as to obtain a target comment.

Through the cooperative cooperation of the components, historical comment data are obtained, the historical comment data are subjected to data division to obtain a plurality of sub-data sets, semantic extraction is respectively carried out on the plurality of sub-data sets through a semantic extraction algorithm to obtain a semantic information set, word embedding mapping is carried out on the plurality of sub-data sets based on the semantic information set to obtain a plurality of low-dimensional dense vectors, and clustering processing is carried out on the plurality of low-dimensional dense vectors to obtain a semantic association relation; performing data conversion on the semantic association relationship to obtain an initial word bag model corresponding to the semantic association relationship, and performing topic number analysis on the initial word bag model to determine the number of target topics; carrying out semantic distribution adjustment on the initial bag-of-words model based on the number of target subjects to obtain a target bag-of-words model; acquiring tag information sets of a plurality of users, inputting the tag information sets into a target word bag model for tag feature extraction, and obtaining target tag features of each user; and carrying out multi-modal fusion on the target tag characteristics and the semantic information sets of each user to obtain corresponding multi-modal fusion parameters, inputting the multi-modal fusion parameters into a preset comment generation model to generate comments, and obtaining target comments. In the scheme of the application, semantic extraction is carried out on the historical comment data to obtain a semantic information set, and word embedding mapping is carried out on a plurality of sub-data sets to obtain a low-dimensional dense vector. The vectors can better express semantic information of comment data, so that more accurate and rich tag features are extracted. Clustering is performed on a plurality of low-dimensional dense vectors, so that comment data similar to semantic information can be classified into one category. The method is favorable for forming semantic association relations, and the comments of similar subjects are gathered together, so that the subsequent target subject number analysis and semantic distribution adjustment are facilitated. The initial word bag model is obtained by carrying out data conversion on the semantic association relation obtained by clustering, and the number of the subjects of the initial model is analyzed, so that the number of the target subjects can be more accurately determined. This helps to generate a more representative target bag-of-words model. And carrying out semantic distribution adjustment on the initial bag-of-words model through the number of target subjects to obtain a target bag-of-words model with more semantic expression. The model can better capture the subject information and semantic information of the comment data, and improves the performance of the comment generation model. The target label characteristics and the semantic information sets of each user are subjected to multi-mode fusion to obtain multi-mode fusion parameters, and the multi-mode fusion parameters are input into a preset comment generation model, so that more personalized, accurate and semantically enriched target comments can be generated. The multi-mode fusion can comprehensively utilize different information sources, and improves the quality and diversity of comment generation.

The automatic comment generating apparatus based on a large-scale language model in the embodiment of the present invention is described in detail above in fig. 5 from the point of view of a modularized functional entity, and the automatic comment generating apparatus based on a large-scale language model in the embodiment of the present invention is described in detail below from the point of view of hardware processing.

Fig. 6 is a schematic structural diagram of a large-scale language model-based comment automatic generating apparatus according to an embodiment of the present invention, where the large-scale language model-based comment automatic generating apparatus 600 may have relatively large differences according to different configurations or performances, and may include one or more processors (central processing units, CPU) 610 (e.g., one or more processors) and a memory 620, and one or more storage media 630 (e.g., one or more mass storage devices) storing applications 633 or data 632. Wherein the memory 620 and the storage medium 630 may be transitory or persistent storage. The program stored on the storage medium 630 may include one or more modules (not shown), each of which may include a series of instruction operations in the large-scale language model-based comment automatic generation apparatus 600. Still further, the processor 610 may be configured to communicate with the storage medium 630 to execute a series of instruction operations in the storage medium 630 on the large-scale language model based comment automatic generation device 600.

The large-scale language model-based comment automatic generation device 600 may also include one or more power supplies 640, one or more wired or wireless network interfaces 650, one or more input/output interfaces 660, and/or one or more operating systems 631, such as Windows Server, mac OS X, unix, linux, freeBSD, and the like. It will be appreciated by those skilled in the art that the large-scale language model-based comment automatic generation apparatus structure shown in fig. 6 does not constitute a limitation of the large-scale language model-based comment automatic generation apparatus, and may include more or less components than those illustrated, or may combine some components, or may be a different arrangement of components.

The invention also provides a large-scale language model-based comment automatic generation device, which comprises a memory and a processor, wherein the memory stores computer readable instructions, and the computer readable instructions, when executed by the processor, cause the processor to execute the steps of the large-scale language model-based comment automatic generation method in the above embodiments.

The present invention also provides a computer readable storage medium, which may be a non-volatile computer readable storage medium, or may be a volatile computer readable storage medium, where instructions are stored in the computer readable storage medium, where the instructions, when executed on a computer, cause the computer to perform the steps of the method for automatically generating comments based on a large-scale language model.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.

The integrated units, if implemented in the form of software functional units and sold or passed as separate products, may be stored in a computer readable storage medium. Based on the understanding that the technical solution of the present invention may be embodied in essence or in a part contributing to the prior art or in whole or in part in the form of a software product stored in a storage medium, comprising instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a read-only memory (ROM), a random access memory (random access memory, RAM), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A comment automatic generation method based on a large-scale language model is characterized by comprising the following steps:

2. The automatic comment generation method based on a large-scale language model according to claim 1, wherein the obtaining of the history comment data, the data division of the history comment data, obtaining a plurality of sub-data sets, and the semantic extraction of the plurality of sub-data sets by a semantic extraction algorithm, respectively, obtaining a semantic information set, wherein the semantic information set includes semantic information corresponding to each sub-data set, includes:

3. The method for automatically generating comments based on a large-scale language model according to claim 1, wherein the word embedding mapping is performed on a plurality of sub-data sets based on the semantic information set to obtain a plurality of low-dimensional dense vectors, and clustering is performed on a plurality of low-dimensional dense vectors to obtain semantic association relations, and the method comprises the following steps:

4. The automatic comment generation method based on a large-scale language model of claim 1, wherein the performing data conversion on the semantic association relationship to obtain an initial bag-of-words model corresponding to the semantic association relationship, and performing topic quantity analysis on the initial bag-of-words model to determine a target topic quantity includes:

5. The method for automatically generating comments based on a large-scale language model according to claim 4, wherein the performing semantic distribution adjustment on the initial bag-of-words model based on the number of target topics to obtain a target bag-of-words model comprises:

6. The method for automatically generating comments based on a large-scale language model according to claim 1, wherein the steps of obtaining tag information sets of a plurality of users, inputting the tag information sets into the target word bag model to perform tag feature extraction, and obtaining target tag features of each user comprise:

7. The automatic comment generation method based on a large-scale language model according to claim 1, wherein the performing multi-modal fusion on the target tag feature of each user and the semantic information set to obtain corresponding multi-modal fusion parameters, inputting the multi-modal fusion parameters into a preset comment generation model to perform comment generation to obtain a target comment, includes:

8. The automatic comment generation device based on the large-scale language model is characterized by comprising the following components:

9. A large-scale language model-based comment automatic generation apparatus, characterized by comprising: a memory and at least one processor, the memory having instructions stored therein;

the at least one processor invoking the instructions in the memory to cause the large-scale language model based comment automatic generation apparatus to perform the large-scale language model based comment automatic generation method of any of claims 1-7.

10. A computer readable storage medium having instructions stored thereon, which when executed by a processor, implement the method for automatically generating a large-scale language model-based comment according to any one of claims 1 to 7.