CN107122494B - Topic Model Construction Method Based on Community Discovery - Google Patents

Abstract

The invention discloses a technical scheme of a topic model construction method based on community discovery, which comprises the following steps in sequence: extracting an implied relation network based on the short text data; dividing the relationship network into a plurality of communities by adopting a community discovery algorithm; expanding the short texts extracted from each community to obtain long documents with word co-occurrence relationship, and constructing a long document set by the obtained long documents; and (4) carrying out theme mining on the long document set to obtain a TMCD theme model discovered based on the community. According to the method, from the perspective of the community relation inherent in the data, the self-expansion of the short text is carried out on the basis of a community discovery algorithm, and the problem of data sparsity is solved.

Description

Topic Model Construction Method Based on Community Discovery

technical field

The invention relates to a topic model construction method based on community discovery, in particular to a topic mining technology for social short text data containing social networks.

Background technique

In the current network environment, with the enrichment of various online platforms, a large amount of social data is generated, and the social network has become a data source for information mining. Most of the data generated in this scenario is presented in the form of short text. Compared with long texts, short texts express concise semantics and transmit information faster, which is an obvious development trend of information dissemination. Short text is becoming one of the most important information carriers in today's society.

At present, in the analysis method of these data, it is a very effective way to mine the semantic information of text connotation through topic model. Classic topic model algorithms, such as PLSA, LDA, etc., mainly perform semantic analysis on texts based on dual patterns and word co-occurrence relationships. This kind of algorithm is effective when dealing with long documents, but when dealing with short texts, due to insufficient word co-occurrence relationship, the algorithm faces the problem of data sparsity, which will seriously affect the quality of the model.

At this stage, the topic model for this kind of short texts in the academic world mainly has the following five processing schemes: 1) use simple splicing to directly connect the short texts; 2) use the method of introducing external databases to aggregate short texts into long texts 3) Implemented from a heuristic method, such as extending the text based on the tag information of the Twitter content, the time stream information sent by the content, or the author who sent the content; 4) Using loose assumptions about the subject of the text, Suppose a short text contains only one topic; 5) Make changes to the modeled objects. A more representative one is the BTM model proposed by Yan et al. in 2013.

The above solutions have many shortcomings, such as forcibly erasing the boundary of the document or being interfered by external data.

SUMMARY OF THE INVENTION

The present invention proposes a method for constructing a topic model based on community discovery (ie, TMCD model, Topic Model based on Community Detection). Dig for solutions. From the perspective of the inherent community relationship in the data, the TMCD model performs self-expansion of short texts based on the community discovery algorithm, which solves the problem of data sparsity.

In order to solve the above problems, the technical solution of the method for constructing a topic model based on community discovery disclosed by the present invention includes the following steps:

Step 1. Extract the relational network contained in the short text data;

Step 2. Use the community discovery algorithm to divide the relational network into multiple communities;

Step 3. Expand the short texts extracted from each community to obtain long documents with word co-occurrence relationships,

And multiple long documents are obtained to form a long document collection;

Step 4. Perform topic mining on long document collections to obtain a TMCD topic model based on community discovery.

Further, the extraction process of the relational network in step 1 is as follows: using the subjects in the short text data as nodes, correlating and abstracting edges through the interaction between subjects, and forming a relational network together with the obtained nodes and edges.

Further, the closeness of the interaction between the subjects is used as the weight of the edge, and the active-passive relationship of the association is used as the direction of the edge.

Further, the community discovery algorithm described in step 2 includes one or more of agglomeration, splitting, label propagation and global exploration.

Further, in step 3, a self-expansion method is used to expand the short text.

Further, the short text data is social data containing a social network inside, and the relationship network is a social network.

The topic model construction method based on community discovery disclosed by the present invention provides a new solution for topic mining of social short text data, and has the following beneficial effects:

(1) This method uses the community network associations contained in the data as the basis for text classification, and on this basis, completes the expansion of short texts, and then solves the problem of data sparsity in the topic mining of short texts. This dataset topic model construction provides a solution.

(2) Through the self-expanding method based on content similarity, this method solves the problem of simple splicing of all content-related and non-content-related topics in the existing short text topic modeling solutions without introducing external help data. The problem of forcibly erasing the boundary of the document caused by the sexual text or the problem of external noise interference caused by the introduction of an external auxiliary corpus, and fundamentally avoid the influence of insufficient word co-occurrence relationship on the topic model.

Description of drawings

Figure 1 is a schematic diagram of the relationship between documents-topic-words in the topic model.

Figure 2 is a schematic diagram of a social network.

FIG. 3 is a flowchart of a method for constructing a topic model for a social data set in an embodiment.

FIG. 4 is a flowchart of the short text extension part in FIG. 3 .

Detailed ways

In order to better understand the technical content of the present invention, specific embodiments are given and described below in conjunction with the accompanying drawings.

Figure 1 shows the relationship between documents, topics, and vocabulary in the topic model. After the concept of "topic" is introduced into the data, the topic can be used as a "bridge" between documents and words. By observing the probability distribution between documents and topics and the probability distribution between topics and words, it can be obtained through relevant mathematical models. distribution of subjects. When obtaining the relationship between topics and words, the number of word co-occurrence relationships will affect the accuracy of the observation results, and this accuracy will further affect the quality of the final topic model. For long texts, there are enough word co-occurrence relationships as support during observation, while short texts lack sufficient word co-occurrence relationships, that is, the problem of data sparsity occurs. The TMCD model construction method proposed by the present invention is developed to solve this problem.

As shown in FIG. 2 , the TMCD model in the embodiment is aimed at social data, through the correlation between the key subjects in the data set (that is, the objects that generate data in the data set, generally contacts) and the subjects (that is, the dissemination paths that generate the data) The topic model obtained after abstraction will present an obvious social network. The abstraction here refers to the abstraction of meaningful contacts and relationships between contacts in the dataset into nodes and edges in the social network. Among them, the subject objects in the abstract data set are nodes, such as contacts in social data; the association between abstract subjects is an edge, and the closeness of the association is used as the weight of the edge. For example, in the social data, the contacts send messages to each other as the edge , the number of messages sent is used as the weight, and the active-passive relationship of sending messages is used as the direction of the edge. An important feature of the obtained social network is that it contains a community structure, and the community structure means that the social network can be divided into several communities through some algorithms, and the data in the same community are similar. In the results of the division, the nodes within the community are relatively closely related and closely connected, while the nodes between the communities are relatively sparsely connected.

Figure 3 is a flowchart of a method for constructing a topic model for a social data set in an embodiment. The method constructs a model based on community discovery, including the following steps:

Step 1: Extract the contained social network according to the communication relationship between the subjects within the social data and the data between subjects. Among them, social data includes all data sets that contain social networks, such as data sets composed of real-time information generated by contacts in instant messaging such as QQ and WeChat, and online social platforms such as Weibo and Zhihu are composed of forwarding and commenting data. The resulting dataset, etc. The specific extraction process is as follows:

1) The subject in the abstract data (ie social data) is used as the node, wherein the subject in the abstract data includes objects that can be used as nodes in the constructed social network, such as people, things or events;

2) Associate through the interaction relationship between subjects, and form an abstract edge, where the interaction relationship includes all the relationships that can form an effective association between two subjects, such as: the transmission of messages in instant messaging constitutes the association of subject contacts, online social networking The subject association formed by forwarding, commenting, and sharing on the platform, etc.;

3) Based on the above steps, the abstracted nodes and edges form an obvious social network.

Step 2: Use the community discovery algorithm to divide the social network into multiple community structures. Community discovery algorithms include all algorithms that can effectively divide communities for social networks, including but not limited to algorithms based on agglomeration process, splitting process, label propagation, and global exploration (including spectral analysis) ideas, which are also most community discovery algorithms. The design idea covers almost all community discovery algorithms that can be effectively divided.

Step 3: Expand the short text contained in each community according to the community structure division result. The expansion method mainly includes the following sub-steps:

1) Extract the short text data corresponding to multiple nodes contained under each divided community;

2) The short text is expanded into a long document by the traditional expansion method based on self-expansion;

3) Based on the above steps, several long documents containing rich word co-occurrence relationships can be obtained (depending on the number of divided communities) obtained by self-expansion of data with text similarity in the social network, and each community is expanded. Long documents constitute a long document collection.

It is worth noting that the expansion based on its own data set does not introduce external help data. Specifically, the direct splicing method can be used as an example, that is, the extracted short text is directly connected. This expansion method itself does not consider whether the text is not. With similarity, the specific operation in this scenario is to expand all the texts corresponding to multiple nodes in the same community as a long document.

Step 4: Perform topic modeling for long document collections and obtain TMCD models. Using traditional topic model construction methods (such as LDA, probabilistic latent semantic analysis, PLSA, etc.), the word-topic observation results are obtained based on the rich word co-occurrence relationship in the document, and then combined with the observed document-topic results, through a certain Mathematical methods (such as Gibbs sampling, etc.) complete the process of topic analysis and mining, and obtain a TMCD model for social data sets. The TMCD model can intuitively output the information of the topic and corresponding keywords contained in the document. Compared with the traditional topic model method directly acting on the short text, the TMCD model additionally performs the text expansion process based on community discovery, so that the text There are enough word co-occurrence relationships in the t to greatly improve the quality of the results of topic mining.

As shown in Figure 4, it is the flowchart of the short text expansion part described in the 2nd) sub-step of step 3 in the embodiment, and specifically includes the following steps:

S3-1 is a short text extraction operation, extracts a plurality of nodes included in an unexpanded community according to the result of community division in step 2 of Figure 3, and then extracts the corresponding short text data from the information of each node;

S3-2 is the short text expansion operation, and the short text extracted in step 3-1 is expanded based on the self-expansion method. Concatenate and expand all texts in this community into one long document;

S3-3 is a judging condition, judging whether all the short texts have been expanded according to the community division result. If there is an unexpanded community, go to step 3-1, otherwise go to step 3-4;

S3-4 is to return the expanded long document set, and the short text expansion step according to the community division result ends.

To sum up, in the embodiment, the method for constructing a topic model based on social discovery provides a new idea for the topic mining of social data sets. , and based on this, self-expansion of short texts forms a long document set, which solves the data sparsity problem faced by topic mining directly on short texts, greatly improves the quality of topic models, and is a social dataset. The topic model provides a solution.

Although the present invention has been disclosed above in preferred embodiments, it is not intended to limit the present invention. Those skilled in the art to which the present invention pertains can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be determined according to the claims.

Claims (4)

1. A topic model construction method based on community discovery is characterized by comprising the following steps:

step 1, extracting an implied relation network based on short text data;

step 2, dividing the relationship network into a plurality of communities by adopting a community discovery algorithm;

step 3, expanding the short texts extracted from each community to obtain long documents with word co-occurrence relationship, and forming a long document set by the obtained long documents;

step 4, performing theme mining on the long document set to obtain a theme model discovered based on the community;

in step 3, the short text is expanded by adopting a self-expansion method, and the method specifically comprises the following steps:

s3-1, extracting a plurality of nodes contained in an unexpanded community from the community division result in the step 2, and then extracting corresponding short text data from the information of each node;

s3-2, expanding the short text extracted in the step 3-1 based on a self-expansion mode to obtain a long document with a word co-occurrence relation;

s3-3, judging whether all short texts are expanded according to community division results, if a community which is not expanded exists, entering the step 3-1, otherwise, entering the step 3-4;

s3-4 is a long document set after returning and expanding;

the short text data is social data with a social network inside, and the relationship network is a social network.

2. The topic model construction method of claim 1 wherein the extraction process of the relationship network in step 1 is: the main bodies in the short text data are used as nodes, correlation is carried out through interactive relations among the main bodies, and edges are extracted to form edges, and the obtained nodes and the edges form a relation network together.

3. The topic model construction method of claim 2 wherein the closeness of the interaction relationships between the bodies is used as the weight of the edge and the associated active and passive relationships are used as the direction of the edge.

4. The topic model construction method of claim 1, wherein the community discovery algorithm of step 2 comprises one or more of aggregation, fragmentation, tag propagation, and global exploration.

Images