1. Introduction

Educational data can often reflect the development trend of education, such as the quality of teachers and teaching ability. Analyzing the website data of a college is a very effective way to know how it develops. The website is like a gateway, and the important decisions and major historical events are continuously published as news announcements. In this way, these data can be mined to obtain important information.

Chongqing Three Gorges Medical College is a medical-orientated college in China, which is responsible for cultivating medical graduates for society. Therefore, it is of great social value to analyze the historical trends of the college. This paper attempts to use the topic model to analyze the information on the official website of this college and then draw the development trends at the decision-making level from topic transfer over time to provide intellectual support for the comprehensive development of the college.

2. Related Work

Analyzing website data has always been a major research topic with many applications. Cha et al. improved the topic model to analyze the relationship graph of social-network data. They then categorized the edges and nodes in the graph based on the topic similarity ^[6]. Their model was quite effective in providing relevant recommendations within a social network. Rohani et al. ran the Latent Dirichlet Allocation (LDA) model on 90,527 social media records in Aviation and Airport Management domain in 2015. They detected the topic facets and extracted their dynamics over time ^[7]. Fuchs et al. proposed a social media analysis model based on trending topics extracted from Twitter and Google. This paper used the textual context to enrich the trends, which helped capture the identification of semantically related topics. Their work outperformed several baselines, including knowledge-graph modeling using DBPedia and directly comparing articles or terms ^[8]. Wang et al. proposed a hot topic detection approach based on bursty term identification to help people assimilate all the news immediately, which considered both frequency and topicality properties to detect the bursty terms and hot topics ^[9]. Later in 2018, Latent Dirichlet Allocation (LDA) was used to model the correlation of news items with stock price time series data ^[10]. The news items in the past were trained, and the similarity between the past and current news items was calculated to build the model. The paper switched the time points to predict the future. Li et al. extracted web news topics and used a model to detect the topic content evolution based on the topic clusters. A quantification method of the topic content was proposed in the model ^[11]. They showed the ``increase or decrease together'' law of topic intensity evolution. Sulova constructed models by combining structured and unstructured data from databases, web pages, and server log files to organize the data from web applications and then provided a summary of the data ^[12]. Wu et al. proposed a multi-view learning framework to incorporate news titles and bodies as different views of news to learn unified news representations. It could achieve good performance in news topic prediction ^[13]. Zhang et al. proposed a dynamic topical community detection method to detect communities and topics, which integrated link, text, and time ^[14]. The method could find communities and their topics with temporal variations. Since the Covid-19 pandemic presented a challenge to the global research community in 2020, Zhang et al. explored the research trace of the pandemic, which changed continuously, based on resilience theory ^[15]. By extracting the characteristic words from the early Covid-19 research articles, they found that the pandemic has significantly disrupted existing research. Bide et al. proposed a cross-event evolutionary detection framework to detect cross-events from similar time features ^[16]. Segmentation clustering was conducted based on similarity computation using the Bidirectional Encoder Representations from Transformers (BERT) model to encode tweets into vectors. In 2022, Zhang et al. designed a Bi-directional Long Short Term Memory-Conditional Random Field (Bi-LSTM-CRF) model for patent entity extraction. They also proposed a topic evolution path identifying method based on knowledge outflow and inflow to calculate the semantic relationship between topics of shared entities ^[17]. ZareRavasan et al. used the topic model to compute the abstracts of 2824 articles published between 1990 and 2020. They reported that topics, such as information system (IS) social practice, IS emerging services, and IS sustainability had gained momentum ^[18]. Feng et al. used Feature Maximization (FM) measurements to select features, combined with the contrast ratio, to perform the diachronic analysis ^[19]. They developed an integrated method based on the Keywords-based Text Representation Matrix (KTRM) and Lamirel’s EC index, and their method performed well in analyzing the diachronic topic evolution. Ding et al. proposed an enhanced latent semantic model, which was based on user comments and regularization factors, to capture the time evolution features of potential topics ^[20]. Their model could capture the changes in the users' interests and provide the evolutionary relationship between users' potential topics and product ratings. To make the research work more comprehensive, Churchill et al. presented a survey on the topic models, tracing back to the origin since the 1990s, comparing these models and their evaluation metrics, and laying the foundation for the next generation of topic models ^[21].

3. Crawler Designing and Data Cleaning

A crawler was implemented to obtain the data on the website. First, all links were obtained. They were then judged, and the non-standard content for each link was filtered out. Finally, those pages that meet the conditions will stay. In particular, the criteria were set to be met, as shown in Table 1. The crawled data were cleaned based on this goal.

The crawled data were transformed into a dictionary according to the Table 1 format, indicating that it is the complete page data. Simultaneously, the simplified text was output with only the title and content into a ``.txt'' file, which is the data fed into the LDA model.

4. Topic Modeling

LDA ^[1] is a classic model that can generate topics automatically. It evolved from Unigram Model (UM), Latent Semantic Analysis (LSA) ^[2], probabilistic Latent Semantic Analysis (PLSA) ^[3], and then to LDA. Its concept is easy to understand, i.e., the whole corpus results from the term generation process. There are two distributions here, and for each document, there is a corresponding document-topic distribution; for each topic, there is a corresponding topic-term distribution. This generation process is that a topic was chosen from the document-topic distribution for each term position in a document. A term was then chosen from the topic-term distribution. Figs. 2 and 3 present the plate notation of LDA and the generation process, respectively. There are two ways to calculate the LDA algorithm. The first way is Variational Inference, which involves the EM (Expectation Maximization) algorithm ^[4]. In the E-step, the coupling relationship between the latent variables was canceled out through the variational assumption, and the variational distribution was obtained. In the M-step, the variational parameters were fixed, and the above expectations were maximized through a series of Newton steps. Through iterations, the parameters were obtained finally. In recent years, there has been an easier way of computing, Gibbs

sampling ^[5]. The idea behind it is that Markov stationary state can be achieved by continuously sampling from the distribution. Gibbs sampling is a special case of MCMC (Markov Chain Monte Carlo), used mainly in the multi-dimensional random variable situation. The system continuously samples the marginal probability distribution to obtain the joint probability distribution. In Gibbs sampling, the joint distribution function should be obtained first, and then we sample according to the Gibbs sampling formula. The joint distribution function can be obtained when the iteration number of samplings is large enough.

LDA has been to be an effective model in topic analysis and word clustering. The selection of the topic number K should be determined by evaluation, which we will describe in the next chapter.

Fig. 1. Flow chart of the proposed topic-analyzing framework.

Fig. 2. LDA plate notation.

Fig. 3. LDA generating process.

5. Experiment

5.1 Data

The BeautifulSoup library of Python was used to help with the implementation of the crawler, and 13682 total links were obtained. After data cleaning, 8264 qualified links were obtained. The data were transformed into two file types. The first was 8264 ``.txt'' files ready to be fed into the LDA model; each ``.txt'' was composed of only a title and content. The second was a large dictionary-form file comprised of 8264 items. Fig. 4 presents the format of each item.

For the 8264 ``.txt'' files, that is, 8264 documents, the title length in each document was from 10 to 20 Chinese characters, and the content was approximately 500 to 1000 Chinese characters.

After obtaining the pure 8264 documents, it is important to perform Chinese word segmentation before they can be fed into LDA. Here, the famous Jieba library of Python was used. The term-index map, index-term map, and the corresponding segmented documents were obtained after the text preprocessing.

Fig. 4. Dictionary format of items.

5.3 Results of Topics

The coherence score fluctuated with the highest value, around 15.9, when the topic number was 10 (Fig. 5). Topic numbers of 15, 25, 45, and 40 were obtained, which were the larger ones in the figure. Four points existed at the lower position, of which the topic numbers were 5, 20, 30, and 50, respectively. In Fig. 6, the PMI score increased when the topic number ranged from 5 to 40 and decreased drastically after that. The PMI score reached a peak when the topic number was 40. Therefore, 40 was chosen as the topic number because, at this point, the topic coherence was larger with the second-largest standard deviation, and the PMI score was the largest one with the second-largest standard deviation. Note that the status in each topic should be distinguished largely from others, so it is a good division.

Overall, this paper considered K = 40 as a proper value for the topic number. All 40 topics with corresponding titles given by human interpretation were then listed, as shown in Table 3. Eight representative topics were chosen to categorize due to the limited space, with the top 10 Chinese phrases in each topic, as shown in Table 4. For example, topic-1, topic-12, and topic-13 were categorized as ``Welcome'', but the facet that each topic focused on was different from the others. Topic-1 concerns the aspect of orientation programs for the students, like some training and the dean’s message. Topic-12 tells the story about student enrollment, such as sign-up and specialty selection. Topic-30 focuses on the welcoming preparations done by the college.

Fig. 5. Topic coherence score with different topic numbers K.

Fig. 6. PMI score with different topic numbers K.

Table 2. Details of Coherence Score and PMI with the Respective Standard Deviation.

Topic number	Coherence score	Coherence score std	PMI	PMI std
5	-17.0769	1.700748	0.782046	0.033435
10	-15.9396	3.794480	0.779588	0.056387
15	-16.1289	2.706822	0.783537	0.050057
20	-16.6534	4.594621	0.788284	0.061898
25	-16.1298	3.916092	0.787955	0.069197
30	-16.7025	3.558920	0.794719	0.062617
35	-16.6258	3.692576	0.793037	0.062819
40	-16.2859	4.276845	0.804979	0.068083
45	-16.2009	3.541288	0.796885	0.060292
50	-17.1378	3.734030	0.786428	0.052768

Table 3. Title of all topics.

Topic ID	Topic title	Topic id	Topic title
0	Poverty alleviation	20	Traditional Chinese medicine
1	New semester starts	21	Party construction
2	Cooperation	22	Punishment
3	Innovation & Research	23	College’s mission
4	Dean’s message	24	Students’ competitions
5	Medical alliance	25	Meetings
6	Staff promotion	26	Jobs & Employment
7	Teacher training	27	Project declaration
8	Organization construction	28	Academic lectures
9	Security & Pandemics	29	Students’ internship
10	Vocational education	30	Welcome new students
11	Inspection of teaching	31	Nation development
12	Student sign up	32	College activities
13	Discipline construction	33	Construction of work style
14	Specialties construction	34	Enrollment affairs
15	Community & Volunteers	35	Examination
16	College development	36	Caring for teachers
17	Faculty & Department	37	Teaching ability
18	Academic affairs	38	Practical education
19	Strong rainfall	39	Recruitment

Table 4. Some Representative Topics with Their Top 10 Chinese Phrases.

Category	Topic ID	Top 10 Chinese phrase
Welcome	1	representative, classmate, health training, students, hope, graduation, ceremony, youth, scholarship, sports meeting, newborn;
	12	sign up, students, examinee, examination, Admission, time, top up exam, complete, recruit students, information, ordinary, selection;
	30	work, students, prepare, school opens, service, welcome new students, library, parents, guarantee, scene, canteen, security;
Activities	15	activity, healthy, volunteer, society, service, resident, science popularization, community, practice, theme;
	24	competition, contest, skill, big match, competitor, national, the first prize, the final round, won, test;
	32	activity, classmate, dorm, culture, apartment, campus, exhibition, civilization, cultural festival, the most beautiful star in campus;
Teaching	7	train, skill, theory, participate in, assistant, practicing, operation, ability, conduct, physician;
	37	teacher, teaching, curriculum, promote, basics, level, design, teaching quality, young teachers, develop, to open up;

6. Conclusion

Manually analyzing the huge amount of information on the network and obtaining knowledge from it is an exhausting task. An analyzing framework was constructed to analyze its topic evolution automatically and solve this problem in educational data. Taking Chongqing Three Gorges Medical College as a case, a crawler was designed to obtain website data, which was cleaned and re-constructed. The data were fed into the topic-processing module. The optimal topic number, the term distribution in each topic, and the topic transfer over the period were obtained. Owing to time limits, Gaussian LDA has not been considered to perform experiment comparisons, and frequent item mining has not been designed for this paper, which could be conducted later. This paper can provide an administration summary for the leadership level of an educational organization, enable them to have better control of the development direction, and give ideas and management experiences of other colleges.