CN114580436A - Social user theme analysis method and system based on semantics and word expansion - Google Patents

Abstract

The invention discloses a social user topic analysis method and a system based on semantics and word expansion, which relate to the technical field of network user information evaluation and solve the problem that any context correlation probably does not exist between user text information in the topic analysis method in the prior art, and the technical scheme is as follows: the short text topic model based on semantics and word expansion is provided, so that not only is the semantic information of a short text fused, but also word pair expansion is carried out through external corpus knowledge, and the topic tendency of the text is enhanced while the problem of short text sparsity is solved; on the basis of a short text topic model based on semantics and word expansion, a text-based topic weighting method based on the amount of interaction among users is adopted to distinguish the importance degree among texts, so that the topic distribution of the users is calculated. Compared with the traditional topic feature analysis method, the social user topic analysis method is more reasonable.

Description

A method and system for topic analysis of social users based on semantics and word expansion

technical field

The invention relates to the technical field of network user information assessment, and more particularly, to a method and system for analyzing social user topics based on semantics and word expansion.

Background technique

With the continuous development of the Internet, the short text data in the network has exploded, and people can publish various information on various social platforms. Therefore, how to analyze user characteristics from social platforms is of great research value and practical significance. Among them, the thematic characteristics of social users is one of the research focuses. At present, most of the information dissemination on major social platforms such as Weibo, Twitter, WeChat, etc. is in the form of short text. Short text data is different from chapter-level long text data, and short text data has different language rules. The user is the main body in the social platform, and each user can publish thousands of short text messages. For enterprises, analyzing the topic information of users can allow enterprises to propose targeted solutions to improve user experience, which has certain commercial value. For scholars, analyzing the topic distribution of users can be used as a social science research tool. results.

At present, the common method of topic feature extraction for users of social platforms is to integrate all the posted documents of the same user into a whole, making it a long document, and then inputting the long document into the topic model to obtain the topic distribution of the document. The distribution is regarded as the topic distribution of users. Although this can solve the problem of short text sparseness, it can also obtain "user-level" topic features. However, it is very likely that there is no contextual connection between the user's posts, or even about things that are not on the same topic. Therefore, it is unreasonable to simply merge all the posts of the same user into a pseudo-document.

Therefore, how to make the analyzed topic distribution characteristics of users more reasonable is an urgent problem to be solved at present.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a social user topic analysis method and system based on semantics and word expansion; it is very likely that there is no contextual correlation between user post information that solves the topic distribution characteristics obtained by the topic analysis method of the prior art, or even The problem of things that are not the same topic is described; the present invention improves the traditional short text topic model, and proposes a short text topic model based on semantics and word expansion. The model performs topic analysis on each social posting information to obtain a topic distribution matrix , and then calculate the weight matrix of the posted information according to the interactive information of each posted information, that is, the importance of the posted information, and finally use the topic distribution matrix and the weight matrix to calculate the user's topic distribution. Compared with the prior art The user topic analysis method for directly constructing pseudo-documents is more advantageous, and the social user topic analysis method based on semantics and word expansion of the present invention is more advantageous.

The above-mentioned technical purpose of the present invention is achieved through the following technical solutions:

In a first aspect, a method for analyzing social user topics based on semantics and word expansion is provided, including the following steps:

Obtaining the post information of the user and the interactive information generated by the post information in the social network platform;

Perform a preprocessing operation on the sent information to obtain text data of the sent information;

constructing a short text topic model based on semantics and word expansion according to the text data;

Inputting the posting information into the short text topic model for calculation, and obtaining a plurality of first topic distribution matrices of the posting information;

Calculate the weight matrix of each piece of the posted information of the user according to the interactive information;

Perform weighting processing on a plurality of the first topic distribution matrices according to the weight matrix to obtain a plurality of second topic distribution matrices;

Perform vector merging and normalization processing on the plurality of second topic distribution matrices to obtain the topic distribution of the user's post information.

Compared with the prior art, the present invention improves the traditional short text topic model, and proposes a short text topic model based on semantics and word expansion. The model performs topic analysis on each piece of social posting information to obtain a topic distribution matrix, Then, according to the interactive information of each post, calculate the weight matrix of the post, that is, the importance of the post, and finally use the subject distribution matrix and the weight matrix to calculate the user's subject distribution, compared with the prior art. The user topic analysis method for directly constructing pseudo-documents, the social user topic analysis method based on semantics and word expansion of the present invention is more advantageous.

Further, the preprocessing operation includes performing a word segmentation operation on the sent information, removing stop words in the sent information, and removing interference symbols in the sent information.

Further, the steps of constructing a short text topic model based on semantics and word expansion according to the text data are as follows:

Extracting the first word pair with semantic association in the context of the text data by using the semantic dependency analysis method;

The keywords to be expanded of the text data are acquired, the keywords to be expanded are input into an external corpus, and the correlation between the keywords to be expanded and words in the external corpus is calculated by using point mutual information.

Further, a judgment threshold of the correlation is set, and if the obtained correlation is greater than the judgment threshold, a second word pair is formed between the keyword to be expanded and the word in the external corpus.

Further, merge all the vectors in the plurality of second topic distribution matrices to obtain the user's topic distribution feature vector, and use the softmax function to normalize the topic distribution feature vector to obtain the user's post information. topic distribution.

In a second aspect, a semantic and word expansion-based social user topic analysis system is provided, including:

an information acquisition unit, configured to acquire the post information of the user and the interactive information generated by the post information in the social network platform;

a preprocessing unit, configured to perform a preprocessing operation on the sent information to obtain text data of the sent information;

a model building unit for building a short text topic model based on semantics and word expansion according to the text data;

a first computing unit, configured to input the post information into the short text topic model for calculation, and obtain a plurality of first subject distribution matrices of the post information;

a second calculation unit, configured to calculate the weight matrix of each piece of the posted information of the user according to the interactive information;

a weighting processing unit, configured to perform weighting processing on a plurality of the first topic distribution matrices according to the weight matrix to obtain a plurality of second topic distribution matrices;

The merging processing unit is configured to perform vector merging and normalization processing on the plurality of second topic distribution matrices to obtain the topic distribution of the post information of the user.

Further, the preprocessing unit includes a word segmentation unit and a removal unit;

The word segmentation unit is configured to perform a word segmentation operation on the posted information, and the removing unit is configured to remove stop words in the posted information and remove interference symbols in the posted information.

Further, the model building unit includes a semantic extraction unit and a word expansion unit;

The semantic extraction unit is used for extracting the first word pair with semantic association in the context of the text data by using a semantic dependency analysis method;

The word expansion unit is used to obtain the keywords to be expanded of the text data, input the keywords to be expanded into an external corpus, and use point mutual information to calculate the relationship between the keywords to be expanded and the words in the external corpus. Correlation.

Further, the word expansion unit also includes a judgment unit;

The judging unit is configured to set a judging threshold for the correlation, and if the obtained correlation is greater than the judging threshold, a second word pair is formed between the keyword to be expanded and the word in the external corpus.

Further, the merging processing unit also includes a vector merging unit and a processing unit;

The vector merging unit is used for merging all the vectors in the plurality of second topic distribution matrices to obtain the topic distribution feature vector of the user;

The processing unit is configured to use a softmax function to normalize the topic distribution feature vector to obtain the topic distribution of the user's post information.

Compared with the prior art, the present invention has the following beneficial effects:

1. The analysis method of the present invention proposes a short text topic model based on semantics and word expansion. On the basis of the traditional short text topic model, this model not only integrates the semantic information of the short text, but also uses an external corpus knowledge base. The word pair expansion is carried out to enhance the topic tendency of the text while solving the problem of short text sparseness.

2. Based on the short text topic model based on semantics and word expansion, the present invention adopts a topic weighting method based on the interactive information between the messages sent by the users to distinguish the importance of the messages, so as to calculate the topic distribution of users. . Compared with the traditional topic feature calculation method, it is more reasonable.

Description of drawings

The accompanying drawings described herein are used to provide further understanding of the embodiments of the present invention, and constitute a part of the present application, and do not constitute limitations to the embodiments of the present invention. In the attached image:

1 is a schematic flowchart of a method for analyzing social user topics provided by an embodiment of the present invention;

2 is a schematic diagram of the acquisition of short text topic model word pairs in the prior art;

3 is a schematic diagram of word pair acquisition based on a short text topic model based on semantics and word expansion according to an embodiment of the present invention;

FIG. 4 is a schematic flowchart of obtaining user topic distribution according to an embodiment of the present invention;

5 is a schematic flowchart of a method for constructing a short text topic model based on semantics and word expansion provided by an embodiment of the present invention;

FIG. 6 is a schematic structural diagram of a social user topic analysis system provided by an embodiment of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and the accompanying drawings. as a limitation of the present invention.

It should be noted that the terms "first" and "second" are only used for the purpose of description, and cannot be understood as indicating or implying relative importance or implying the quantity of the indicated technical features. Thus, a feature defined as "first" or "second" may expressly or implicitly include one or more of that feature. In the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

Example 1:

As shown in FIG. 1 , Embodiment 1 of the present application provides a method for analyzing social user topics based on semantics and word expansion, including the following steps:

S1, obtain the post information of the user and the interactive information generated by the post information in the social network platform;

The post information of the user includes the fixed personal information of the user, such as the user ID, and the original work information, such as the article ID, post content, etc. The post information mainly refers to short text information. The interactive information that generates interactive behavior with each post information, including the number of forwarding, number of likes, number of favorites, number of comments, number of followers and number of shares, etc. To analyze the topic distribution of posts posted by social network users, this application uses all post information among social network users as a basis, as well as interactive information that generates interactive behaviors with each post information.

S2, performing a preprocessing operation on the sent information to obtain text data of the sent information;

It is a conventional technical means in the art to perform a preprocessing operation on the sent message to obtain the text data of the sent message, so no redundant description will be given.

S3, construct a short text topic model based on semantics and word expansion according to the text data;

Specifically, as shown in FIG. 2, FIG. 2 is the method used by the short text topic model (BTM model for short) in the prior art to obtain the word pair information of the short text, but when the word pair information is obtained by using a sliding window The semantic association of the context of word pairs in short texts is ignored, so the short text topic model based on semantics and word expansion of this application adopts the method of semantic dependence to extract word pairs in the text, so that the BTM model can integrate the semantics of the context book. information. In the prior art, the topic features of users are mainly extracted by constructing pseudo-documents, which can solve the problem of sparse short texts. The word pair expansion method based on word association (namely, the short text topic model of word expansion) is proposed, so that the BTM model can solve the sparse problem of short texts.

S4, inputting the post information into the short text topic model for calculation, and obtaining a plurality of first subject distribution matrices of the post information;

Input all the posted information of the user into the short text topic model based on semantics and word expansion, and the first topic distribution matrix of each posted information can be obtained, then the topic distribution of all the posted information of each user is essentially a D*K dimension of the first topic distribution Topic distribution matrix, D represents all the posted information of the user, and K represents the set number of topics. The first topic distribution matrix of D*K dimension is denoted as Tuser.

S5, calculating the weight matrix of each piece of the posted information of the user according to the interactive information;

Each user can obtain a weight matrix of D*1 dimension, D represents all the post information of the user, and the weight matrix of D*1 dimension is denoted as H _user . The weight matrix calculated according to the interactive information is the "hotness" of the posted information. The value of the weight matrix can be used as the weight value of the corresponding posted information. The weight value indicates the importance of the posted information to the user. The more important the topic distribution, the stronger the influence.

S6, performing weighting processing on a plurality of the first topic distribution matrices according to the weighting matrix to obtain a plurality of second topic distribution matrices;

In S5, the weight value represents the importance of the post information to the user. Therefore, the weight matrix H _user calculated by the interactive information is used to weight the first topic distribution matrix of the post information Tuser, and the second topic of the user can be obtained. Distribution matrix X _user . Since the first topic distribution matrix is a multi-dimensional matrix calculated by the short text topic model based on semantics and word expansion on all the post information of the user, the corresponding post information is weighted by each piece of interactive information, so that Calculate the topic distribution of users.

S7: Perform vector merging and normalization processing on the plurality of second topic distribution matrices to obtain the topic distribution of the user's post information.

For multiple second topic distribution matrices X _user , all vectors in multiple matrices need to be merged to represent the topic distribution of users, but due to the different number of interactions of each user, the value of topic distribution among users is very large Therefore, normalization is required to reduce the variability of topic distributions.

In yet another embodiment of Embodiment 1 of the present application, the preprocessing operation includes performing a word segmentation operation on the posted information, removing stop words in the posted information, and removing interference symbols in the posted information.

It should be noted that the preprocessing operation may also include part-of-speech tagging processing for the posted information, and keyword extraction, topic extraction, abstract, event extraction and semantic extraction for the posted information, so that the transformed text data is more diverse.

In yet another embodiment of Embodiment 1 of the present application, as shown in FIG. 5 , the steps of constructing a short text topic model based on semantics and word expansion according to the text data are as follows:

Extracting the first word pair with semantic association in the context of the text data by using the semantic dependency analysis method;

The keywords to be expanded of the text data are acquired, the keywords to be expanded are input into an external corpus, and the correlation between the keywords to be expanded and words in the external corpus is calculated by using point mutual information.

Specifically, the short text topic model based on semantics and word expansion improves the traditional BTM model. As shown in Figure 2, the traditional BTM model uses a sliding window to obtain word pair information, ignoring the contextual semantic association of word pairs in the text. As shown in Figure 3, the improved BTM model uses the semantic dependency method to extract word pairs in the text, so that the BTM model integrates text semantic information. “President of xx country” as an example, after preprocessing the post information, in the process of constructing the word pair data set, the sliding window will construct “election” and “become” into a word pair, but the human opinion of this sentence Geographically understand the analysis, there is no grammatical and logical connection between "election" and "become", it is unreasonable to form two words into word pairs and enhance the probability that the two words belong to the same topic, "xx" and "president" are two words. There are obvious semantic and logical associations between words, and it is more reasonable to enhance the probability that they belong to the same topic, but due to the size of the sliding window, "xx" and "president" will not constitute a word pair. In the first word pair extracted by the semantic dependency method, still take the sentence "xx won the election of xx country to become the next president of xx country" as an example, "xx" and "become" belong to the subject-predicate relationship, "become ” and “President” belong to the relationship of verb and guest, and “xx” and “President” have the relationship of host and guest. Therefore, in the word pair construction, the words with strong correlation are selected to form the first word pair, so that the formed first word pair has the semantic association of the context in the post information.

At the same time, in order to ensure that the model can further solve the sparse problem when the text information is short text, a word pair expansion method based on word association is proposed. to calculate the correlation between the keyword x to be expanded and the word y in the external corpus. The point mutual information of x and y is denoted as PMI(x,y), as shown in the following formula:

Among them, P(x) is the probability that the keyword x appears in the external corpus; P(y) is the probability that the word y appears in the external corpus. For example, taking words such as "Xiaomi", "Apple", "Huawei", and "computer" as examples (if these words are extracted from the information posted by users in the text), extract the words related to these words in the external news data set. Target words with the highest PMI values:

Table 1 Example of the result of word expansion calculation

In the end, the word pairs that can be expanded from the word "Xiaomi" are "Xiaomi Coolpad", "Xiaomi MIUI", "Xiaomi Ecological Chain" and so on. Word pairs for which the word "apple" can be expanded: "apple apple", "apple jobs", "apple cook", etc. Other words are also similar word expansion methods. Thus, the sparse problem when the sent information is short text information is solved.

In yet another embodiment of the first embodiment of the present application, a judgment threshold of the correlation is set, and if the obtained correlation is greater than the judgment threshold, the keywords to be expanded and the words in the external corpus are formed into a second word right.

Specifically, if the correlation value obtained by the correlation calculation satisfies the set judgment threshold, and the judgment threshold is generally 200, it is considered that the keyword to be expanded has a strong correlation with the words in the external corpus, and a second word pair can be formed .

In yet another embodiment of Embodiment 1 of the present application, all vectors in the plurality of second topic distribution matrices are combined to obtain the topic distribution feature vector of the user, and the softmax function is used to normalize the topic distribution feature vector process to obtain the topic distribution of the post information described by the user.

U_topic表示用户发文的主题分布，X_d表示主题分布特征向量，D表示用户所有的发文信息(可以为用户发了多少篇文章)。Specifically, as shown in FIG. 4 , the existing technical solution is to combine all the texts sent by the user into a long document, and then directly input it into the traditional BTM model for topic analysis. However, in the present invention, after subject analysis is performed on each post of the user separately, the subject distributions of all posts by the same user are merged and normalized to be used as the subject feature of the user. The normalization process is as follows:

U _topic represents the topic distribution of the user's post, X _d represents the topic distribution feature vector, and D represents all the post information of the user (how many posts can be posted for the user).

Combining the above technical solutions, the social user topic analysis method in Embodiment 1 of the present application proposes a short text topic model based on semantics and word expansion based on the traditional short text topic model. Analysis, get the topic distribution matrix, and then calculate the weight matrix of the posted information according to the interactive information of each posted information, that is, the importance of the posted information, and finally use the topic distribution matrix and the weight matrix to calculate the user's topic distribution, Compared with the user topic analysis method of directly constructing pseudo-documents in the prior art, the social user topic analysis method based on semantics and word expansion of the present invention is more advantageous.

Embodiment 2:

Based on the same concept, as shown in FIG. 6 , the second embodiment provides a social user topic analysis system based on semantics and word expansion on the basis of the first embodiment, including:

an information acquisition unit 110, configured to acquire the post information of the user and the interactive information generated by the post information in the social network platform;

a preprocessing unit 120, configured to perform a preprocessing operation on the sent information to obtain text data of the sent information;

A model building unit 130, configured to build a short text topic model based on semantics and word expansion according to the text data;

A first calculation unit 140, configured to input the posted information into the short text topic model for calculation, and obtain a plurality of first topic distribution matrices of the posted information;

The second calculation unit 150 is configured to calculate, according to the interaction information, a weight matrix of each piece of the post information of the user;

a weighting processing unit 160, configured to perform weighting processing on a plurality of the first topic distribution matrices according to the weight matrix to obtain a plurality of second topic distribution matrices;

The merging processing unit 170 is configured to perform vector merging and normalization processing on the plurality of second topic distribution matrices to obtain the topic distribution of the post information of the user.

The social user topic analysis system in the second embodiment of the present application proposes a short text topic model based on semantics and word expansion on the basis of the traditional short text topic model. The model performs topic analysis on each social posting information to obtain topic distribution. Matrix, and then calculate the weight matrix of the posted information according to the interactive information of each posted information, that is, the importance of the posted information, and finally use the topic distribution matrix and weight matrix to calculate the user's topic distribution. Compared with the existing The technical user topic analysis system for directly constructing pseudo-documents, and the social user topic analysis system based on semantics and word expansion of the present invention is more advantageous.

In yet another embodiment of the second embodiment of the present application, the preprocessing unit 120 includes a word segmentation unit and a removal unit;

The word segmentation unit is configured to perform a word segmentation operation on the posted information, and the removing unit is configured to remove stop words in the posted information and remove interference symbols in the posted information.

In yet another embodiment of the second embodiment of the present application, the model construction unit 130 includes a semantic extraction unit and a word expansion unit;

The semantic extraction unit is used for extracting the first word pair with semantic association in the context of the text data by using a semantic dependency analysis method;

The word expansion unit is used to obtain the keywords to be expanded of the text data, input the keywords to be expanded into an external corpus, and use point mutual information to calculate the relationship between the keywords to be expanded and the words in the external corpus. Correlation.

In yet another embodiment of the second embodiment of the present application, the word expansion unit further includes a judgment unit;

The judging unit is configured to set a judging threshold for the correlation, and if the obtained correlation is greater than the judging threshold, a second word pair is formed between the keyword to be expanded and the word in the external corpus.

In yet another embodiment of the second embodiment of the present application, the merging processing unit 170 further includes a vector merging unit and a processing unit;

The vector merging unit is used for merging all the vectors in the plurality of second topic distribution matrices to obtain the topic distribution feature vector of the user;

The processing unit is configured to use a softmax function to normalize the topic distribution feature vector to obtain the topic distribution of the user's post information.

It should be noted that, for the methods executed by the above program units in the second embodiment, reference may be made to the various embodiments in the social user topic analysis method of the present invention, and thus will not be described again.

The specific embodiments described above further describe the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above descriptions are only specific embodiments of the present invention, and are not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. A social user topic analysis method based on semantics and word expansion is characterized by comprising the following steps:

acquiring user's text information and interactive information generated by the text information in a social network platform;

preprocessing the text sending information to obtain text data of the text sending information;

constructing a short text topic model based on semantics and word expansion according to the text data;

inputting the text information into the short text topic model for calculation to obtain a plurality of first topic distribution matrixes of the text information;

calculating a weight matrix of each piece of text information of the user according to the interaction information;

weighting the first theme distribution matrixes according to the weight matrix to obtain second theme distribution matrixes;

and carrying out vector merging and normalization processing on the plurality of second theme distribution matrixes to obtain the theme distribution of the text information of the user.

2. The method according to claim 1, wherein the preprocessing operation comprises performing a word segmentation operation on the text information, removing stop words of the text information, and removing interference symbols in the text information.

3. The method for social user topic analysis based on semantic and word expansion as claimed in claim 1, wherein the step of constructing the short text topic model based on semantic and word expansion according to the text data comprises:

extracting a first word pair with semantic association in the text data context by adopting a semantic dependency analysis method;

and acquiring keywords to be expanded of the text data, inputting the keywords to be expanded into an external corpus, and calculating the correlation between the keywords to be expanded and words in the external corpus by using the point-to-point information.

4. The method as claimed in claim 3, wherein a threshold for determining the correlation is set, and if the correlation is greater than the threshold, the keywords to be expanded are combined with words in an external corpus to form a second word pair.

5. The social user topic analysis method based on semantics and word expansion according to claim 1, wherein all vectors in the second topic distribution matrices are combined to obtain topic distribution feature vectors of users, and a softmax function is used to normalize the topic distribution feature vectors to obtain topic distribution of the user text information.

6. A social user topic analysis system based on semantics and word expansion, comprising:

the information acquisition unit is used for acquiring the text sending information of a user and the interactive information generated by the text sending information in the social network platform;

the preprocessing unit is used for preprocessing the text information to obtain text data of the text information;

the model building unit is used for building a short text topic model based on semantics and word expansion according to the text data;

the first calculation unit is used for inputting the text information into the short text topic model for calculation to obtain a plurality of first topic distribution matrixes of the text information;

the second calculation unit is used for calculating a weight matrix of each piece of text information of the user according to the interaction information;

the weighting processing unit is used for weighting the plurality of first theme distribution matrixes according to the weight matrix to obtain a plurality of second theme distribution matrixes;

and the merging processing unit is used for carrying out vector merging and normalization processing on the plurality of second theme distribution matrixes to obtain the theme distribution of the text information of the user.

7. The system of claim 6, wherein the preprocessing unit comprises a word segmentation unit and a removal unit;

the word segmentation unit is used for performing word segmentation operation on the text information, and the removal unit is used for removing stop words of the text information and removing interference symbols in the text information.

8. The system for social user topic analysis based on semantics and word expansion according to claim 6, wherein the model building unit comprises a semantics extracting unit and a word expanding unit;

the semantic extraction unit is used for extracting a first word pair with semantic association in the text data context by adopting a semantic dependency analysis method;

the word expansion unit is used for acquiring the keywords to be expanded of the text data, inputting the keywords to be expanded into an external corpus, and calculating the correlation between the keywords to be expanded and words in the external corpus by using the point mutual information.

9. The system according to claim 8, wherein the word expansion unit further comprises a judgment unit;

and the judging unit is used for setting a judging threshold value of the correlation, and forming a second word pair by the keyword to be expanded and the words in the external corpus if the obtained correlation is greater than the judging threshold value.

10. The system of claim 6, wherein the merging processing unit further comprises a vector merging unit and a processing unit;

the vector merging unit is configured to merge all vectors in the plurality of second topic distribution matrices to obtain topic distribution feature vectors of the user;

the processing unit is used for carrying out normalization processing on the theme distribution characteristic vector by utilizing a softmax function to obtain the theme distribution of the text information of the user.

Images