Keywords

1 Introduction

The vigorous development of the Internet has made more and more netizens gather in the online community such as Weibo and post-bar and so on. In particular, college students express their views, discussed hot topics about their school in the university online community. As the active network community in Colleges and universities, the amount of information contained in the network community has increased geometrically. Correspondingly, the information related to all aspects of College Students’ life has become a hot research contents. This paper focuses on the contents of university online community, proposes an improved LDA method for text topic recognition, and gives weight to the results. By calculating the weight, we can judge whether the output results can represent the corresponding topic model. The text of university network community has the problem of large text noise, fast text updating and short single text. To solve these problems, firstly, this paper extends the LDA model’s stop word library and filters a large number of general useless replies to decrease the text noise. Then, uses the LDA topic model to generate a word vector group which can represent this topic, and each vector is given different weights. By calculating the weights, we can judge whether the topic model can truly represent the typical topic. Lastly, this paper establishes some topic models which can represent the typical events in universities. The experimental results show that the topic recognition model based on this paper can better reflect some of the hot events in universities than the traditional LDA topic model, so as to provide reasonable support for university management.

2 Related Research

2.1 Related Work

At present, some researches have used LDA topic model or improved LDA topic model to mine topic models in different fields. Zhu Maoran et al. use LDA model to identify the topic and its key words in the whole text set, and calculated the document-topic probability distribution in each time windows. Then the topic-lexical probability distribution is calculated by LDA model for each text set under each time window, and the similarity of different topics under different time windows is calculated, thus the evolution trend of topic strength is obtained. Finally, we get the change of thematic content through the probability distribution of words under similar themes. [1] Peng et al. embed a priori knowledge of word association, global feature words and subject emotion subordinate semantics in LDA model to improve LDA’s ability to recognize feature words, affective words and their relationships. And it uses LDA model to extract the feature words and emotional words of product reviews, and take advantage of affective analysis technology [2]. To classify emotional polarity of product reviews Yi et al. introduce the strategic coordinates into the subject analysis of LDA patents, and fuses the time factors, using the internal and external correlation index to measure the relationship between the various parts of the subject. [3] Tan et al. apply the LDA model to the mining of hot topics, and constructs the identification process of Weibo hot topics. [4] Zheng et al. propose an improved LDA model ST-LDA to analyze Weibo topics. This model assigns a topic to each Weibo. In particular, when assigning a topic to each Weibo, not only the influence of posting time on the topic is considered, but also the influence of posting location on the topic is added. Thus, the model can assign a topic to each Weibo. The semantic similarity of Weibo published in adjacent time and space is classified into the same topic, which improves the comprehensibility of the topic. [5] Xu et al. use LDA topic model to analyze the topic relevance of blog content, redefine the link relationship between bloggers, and then classify the topics. [6] Shi et al. mainly explore the evolution of BBS as a topic in the network media in time, so as to find hot topics and no-hot topics, and better guide the netizens to understand what is happening. [7] He et al. proposes a simple and easy symmetric learning data augmentation model (SLDAM) for underwater target radiate-noise data expansion and generation. The SLDAM, taking the optimal classifier of an initial dataset as the discriminator, makes use of the structure of the classifier to construct a symmetric generator based on antagonistic generation. It generates data similar to the initial dataset that can be used to supplement training data sets. [8] Wang et al. proposes a natural language semantic construction method based on cloud database to improve natural language comprehension ability and analytical skills of the machine [9].

The research above has done a lot of work on the application level of LDA, and has also been extended based on this theory to optimize the topic recognition in specific areas. However, we can also see the shortcomings: on the one hand, LDA is an unsupervised learning model, the input content will greatly affect the accuracy of the output results, and these methods do not effectively de-noise the input data, which still has a lot of room for optimization; on the other hand, the output of the LDA model is considered the final result of the corresponding topic without considering the intrinsic correlation between the output of the LDA model. Although LDA model treats the segmentation results equally, the results of LDA model can be re-weighted to determine whether the results can represent the corresponding topic model.

2.2 LDA Model and Application

Latent Dirichlet Allocation (LDA), is a topic model commonly used in the field of machine learning and is often used for text categorization.

LDA was put forward in 2003 by Blei, David M., Ng, Andrew Y. and Jordan to predict the topic distribution of documents. It can give the topic of each document in the document set in the form of probability distribution. After analyzing some documents and extracting their topic distribution, it can cluster or classify the text according to the topic distribution.

Suppose there are M documents, corresponding to the d document, there are \( {\text{N}}_{\text{d}} \) words, that is, the input is as follows (Fig. 1):

Fig. 1.
figure 1

The LDA Input schematic diagram.

Full size image

In order to find the topic distribution of each document and the word distribution in each topic, this paper first assumes a topic number K, so that all distributions are based on K topics. Therefore, a specific LDA model can be obtained, as shown in the following Fig. 2:

Fig. 2.
figure 2

The LDA model schematic diagram.

Full size image

LDA assumes that the prior distribution of a document topic is a Dirichlet distribution, that is, for any document d, the topic distribution \( \uptheta_{\text{d}} \) is:

$$ \theta_{d} = Diriclet(\vec{\alpha }) $$
(1)

Among them, \( \alpha \) is a distributed super parameter, which is a K dimensional vector.

LDA assumes that the prior distribution of words in a topic is Dirichlet distribution, that is, for any topic k, the word distribution \( \beta_{k} \) is:

$$ \beta_{k} = Dirichlet\left( {\vec{\eta }} \right) $$
(2)

Among them, the parameter \( \eta \) is a V dimensional vector. V stands for the number of words in the word list.

For the nth word in any document d in the data, the distribution of the subject number \( z_{dn} \) is obtained from the topic distribution \( \uptheta_{\text{d}} \) as follows:

$$ z_{dn} = multi\left( {\theta_{d} } \right) $$
(3)

For this topic number, the probability distribution of word \( w_{dn} \) can be obtained as follows:

$$ w_{dn} = multi\left( {\beta_{{z_{dn} }} } \right) $$
(4)

Thus, for M document topics, the corresponding data has multiple distributions of M topic numbers, so that \( \alpha \to \theta_{d} \to \vec{z}_{d} \) constitutes a Dirichlet-multi conjugate. A posteriori distribution of document topics based on Dirichlet distribution can be obtained by Bayesian inference.

If the number of words for topic K in the dth document is \( n_{d}^{(k)} \), the count of polynomial distributions can be expressed as:

$$ \vec{n}_{d} = \left( {n_{d}^{\left( 1 \right)} ,n_{d}^{\left( 2 \right)} , \ldots n_{d}^{\left( K \right)} } \right) $$
(5)

The posterior distribution of \( \uptheta_{\text{d}} \) is obtained by using Dirichlet-multi conjugate is:

$$ Dirichlet(\theta_{d} |\vec{\alpha } + \vec{n}_{d} ) $$
(6)

For the distribution of subjects and words, if there is a Dirichlet distribution of K subjects and words, and the corresponding data has a polynomial distribution of K subject numbers, then \( \eta \to \beta_{k} \to \vec{w}_{\left( k \right)} \) constitutes a Dirichlet-multiconjugate. Bayesian inference method can be used to obtain a posterior distribution of subject words based on the Dirichlet distribution.

If the number of words V in topic K is \( n_{k}^{(v)} \), the number of polynomial distributions can be expressed as:

$$ \vec{n}_{k} = \left( {n_{k}^{\left( 1 \right)} ,n_{k}^{\left( 2 \right)} , \ldots n_{k}^{\left( V \right)} } \right) $$
(7)

By using Dirichlet-multi conjugation, the posterior distribution of \( \beta_{k} \) is obtained as:

$$ Dirichlet\left( {\beta_{k} |\vec{\eta } + \vec{n}_{k} } \right) $$
(8)

Because the subject-generated words do not depend on a specific document, the topic distribution and the subject-word distribution of the document are independent.

3 Establishing the Theme Model of University Online Community

3.1 Text Topic Recognition Method in University Online Community

The concept of topic area is mostly adopted in the network community of colleges and universities. The topics discussed in the same topic area can be summed up as a kind of campus-related topics. Every post in the topic area revolves around a particular topic. If you ignore the useless responses in a single post, you can assume that all the responses are thematic responses. Considering these characteristics of college network community, this paper preprocesses the input text before generating topic distribution using LDA model, as shown in the following Fig. 3:

Fig. 3.
figure 3

Schematic diagram of text preprocessing

Full size image

After pretreatment, the topic posts can be used as the input of LDA model to get the distribution of the topic words, because all the noises that affect the results are removed. And because each topic Posts contain only one topic, we can easily get the distribution of the topic words of a hot topic. As the LDA model is a bag of words model, the result can only show the frequency distribution of the subject words, and can not be divided in detail according to the weight of different words. According to the characteristics of the subject words in the topic model, this paper proposes a method of weight assignment according to the category of the subject words. By calculating the final weight results, we can get the result of weight assignment. Determine whether these topic phrases can correctly represent the corresponding topic models. Therefore, the whole process of text topic recognition in this paper is shown in the following Fig. 4:

Fig. 4.
figure 4

Flow chart of text topic recognition

Full size image

3.2 The Establishment of Popular Topic Model in University Online Community

The heat of the topic area in the network community reflects the students’ attention to the corresponding events in universities. This paper investigates a large number of hot topics in the network community of colleges and universities, and obtains the topics closely related to colleges and universities shown in the following Fig. 5:

Fig. 5.
figure 5

Flow chart of text topic recognition

Full size image

From the above chart shows, the hot topics related to the campus is divided into three categories, six small classes, which basically covers all aspects of campus life. This paper establishes the corresponding topic model bases on these six categories. The right-most side of the graph shows the weight of each level of the hot topic. For the model level, the letter u is used to represent the level, and \( {\text{u}}_{1} {\text{u}}_{2} {\text{u}}_{3} \ldots {\text{u}}_{\text{n}} \) is used to represent the corresponding level. For each level, this article uses the letter w to represent the corresponding weight, \( {\text{w}}_{1} {\text{w}}_{2} {\text{w}}_{3} \ldots {\text{w}}_{\text{n}} \) denotes the weight of the corresponding hierarchy, the letter \( \lambda = \delta /\sum\nolimits_{0}^{n} u \) denotes the compensation value, and the value of delta is related to the word frequency input by the LDA model. If the maximum word frequency is less than 0.1, then δ = 0.1, and so on. The formula for calculating weight can be obtained as follows:

$$ {\text{f}}\left( {{\text{u}},w} \right) = \frac{{{\text{u}}_{1} *{\text{w}}_{1} + {\text{u}}_{2} *{\text{w}}_{2} + {\text{u}}_{3} *{\text{w}}_{3} + \ldots + {\text{u}}_{\text{n}} *{\text{w}}_{\text{n}} }}{{{\text{u}}_{1} + {\text{u}}_{2} + {\text{u}}_{3} + \ldots {\text{u}}_{\text{n}} }}*\lambda $$
(9)

The weighted result of the topic model can be obtained by using the above formula. If the result is greater than 0.5, the input text can be used to represent the corresponding topic model.

4 Experimental Results and Analysis

4.1 Text Topic Recognition Analysis of University Network Community

The University Forum of Beijing University of Posts and Telecommunications is one of the typical network communities in Colleges and universities. According to the different topics, the forum is divided into eight discussion areas, each of which includes the website, the campus of Beijing University of Posts and Telecommunications, academic science and technology, information society, humanities and arts, life style, leisure and entertainment, physical fitness, games and love of hometown. There are other discussion areas behind those areas. By analyzing the content of the discussion area, this paper draws a conclusion that posts posted on specific plates have obvious characteristics. For example, in the life fashion section, the discussion is mostly about life-related topics, such as canteen, bathroom, supermarket, and so on, while in the health care section, the discussion is mostly about hospitals, nursing and other topics. These topic posts also have a strong attribute: that is, the response below a topic post, except for some irrelevant replies, the rest of the replies are related to the topic of the topic post discussion.

This paper selects the posts related to campus life in the life style area of the forum of Beijing University of Posts and Telecommunications, and obtains the contents of each posts in the zone by means of web crawler.

Web crawler is also called web spider, or web robot. It is a kind of programs or scripts according to certain Rules that automatically grab information from the World Wide Web. In other words, it can automatically get the content of the web page according to the link address of the web page. If we compare the Internet to a big spider web, it has a lot of web pages, web spiders can get all of webpage’s content. [10] In this paper, we can find the origin contents on the website like this (Fig. 6):

Fig. 6.
figure 6

The origin text on the website

Full size image

The origin contents above can’t use by LDA model before text preprocessing completed. We can see the whole text below (Fig. 7):

Fig. 7.
figure 7

The noise text without text preprocessing

Full size image

So, if we use LDA directly, the results obtained by directly using LDA model are as follows (Fig. 8):

Fig. 8.
figure 8

LDA result without text preprocessing

Full size image

As you can see, the phrases with the highest frequency are in turn “bd”, “的”, “了” and other characters that have nothing to do with the subject.

After the text preprocessing, the output of the LDA thematic model is shown in the following Fig. 9:

Fig. 9.
figure 9

LDA result using text preprocessing

Full size image

Furthermore, a detailed comparison of the above results is shown in the following Table 1:

Table 1. A detailed comparison of the two results
Full size table

We can see that after the text preprocessing process proposed in this paper, the correlation between the topic distribution obtained by LDA model and the actual topic distribution involved has been effectively improved.

4.2 Verification of Popular Topic Model in University Online Community

After obtaining the frequency of the above-mentioned topic distribution words, different weights are assigned to them, and a topic model similar to the one shown in the Fig. 10 is obtained:

Fig. 10.
figure 10

An example model using the theory of this paper

Full size image

So the weight of the vector frequency of the above words can be calculated like this:

$$ {\text{f}}\left( {{\text{u}},{\text{w}}} \right) = \frac{{\left( {\left( {0.017 + 0.011} \right)*0.45 + 0.008*0.35 + \left( {0.011 + 0.010} \right)*0.2} \right)*0.1}}{{\left( {0.017 + 0.011 + 0.08 + 0.011 + 0.010} \right)^{2} }} = 0.6033 $$
(10)

Because of 0.6033 > 0.5, this paper can draw the conclusion that the above topic models can well represent the corresponding themes. That is: the content of the theme is related to the basketball court and playground closely related to the life of college students.

This conclusion also applies to other similar campus themes if we strictly fellow the steps above.

5 Summary and Prospect

The university online community is closely related to schools and students. How to dig up the information from it and serve the schools and students is the current concern of all colleges and universities. This paper focuses on the characteristics of university online community, proposes an improved LDA method for text topic recognition. Firstly, this paper extends the LDA model’s stop word library and filters a large number of general useless replies to decrease the text noise. Then, uses the LDA topic model to generate a word vector group which can represent this topic, and each vector is given different weights. By calculating the weights, we can judge whether the topic model can truly represent the typical topic. Lastly, this paper establishes some topic models which can represent the typical events in universities. In the future, on the basis of this study, we can also establish other new thematic models according to the characteristics of universities, improve the calculation method of thematic model weight, and explore the development trend of thematic models of typical events in universities.