Development of a research support tool for literature review facilitating data mining techniques

Aleksandr Kazantsev

Master's thesis

Abstract:

Numerous research papers and scholarly data have been released online over the past few years. The simplicity of accessing the data and tremendous growth of knowledge brings significant advantages to the research community, but it also creates the information overload problem, especially in academia. Therefore, there is a growing demand for the ability to segment and automatically analyze research papers in the research field. The unsupervised machine learning method of topic modelling is an automated method to extract information from scholarly data that has become increasingly popular. The application of topic modelling for generating topics to segment, explore and describe the Geophysical Research Abstracts (GRA) has been explored in this thesis. The topic modelling takes advantage of enormous amounts of text data to discover topics that run through a collection of documents by using statistical relationships between the terms in these documents. The Latent Dirichlet Allocation (LDA), Non-Negative Matrix Factorization (NMF) algorithms and Okapi BM25 information retrieval model are implemented and evaluated. The effectiveness of the topic modelling is assessed by the ability to serve as a uniform categorization framework for research papers and by showing that algorithms can generate meaningful topics and keywords. Extensive data preparation and preprocessing of Geophysical Research Abstracts were required to apply topic models successfully. The results demonstrate that LDA and NMF algorithms could create topic models with meaningful topics and that topic modelling can be used for content analysis and potentially as an unsupervised categorization framework. However, it was found that NMF and LDA algorithms have different characteristics and should be applied in different usage cases. One of the significant weaknesses is that topics created by topic models, compared to manual methods, are more unreliable and could produce misleading results, which is an effect of its uncontrollable nature. On the other hand, its strength is the ability to analyze large amounts of text in a short time and at a low cost, deriving insights from many research papers. Topic modelling can complement other methods for content analysis or categorization, and it is a powerful method for aggregating and presenting the results to generate insights for efficiently analyzing and segmenting research papers.

Keywords

topic modelling, NMF, LDA, topics, document keywords

Language used: English

Date on which the thesis was submitted / produced: 31. 3. 2022

Thesis defence

• Supervisor: doc. Mgr. Ing. Ioannis Markonis, Ph.D.

Czech University of Life Sciences Prague

Faculty of Environmental Sciences

Master programme / field:
Landscape Engineering / Environmental Modelling