Article Preview
Top1 Introduction
In the research area of text mining and natural language processing, one of the challenges is extracting information from large documents or collections of unstructured documents. Problems like clustering the documents, topic extraction, topic labeling, and content extraction are very common. The topic models or topic modeling techniques are unsupervised learning methods for the extraction of prevalent topics from documents. These methods give a powerful tool to represent words and topics in text collections. The assumption describes documents as a mixture of different topics not one e.g., a medical journal document could be about ‘cancer’, ‘Immunity’ and ‘DNA’ together. Topic discovery from medical literature is useful in browsing and organizing a large collection of medical documents. Topic models can also be used to assign research papers to reviewers. The main use of topic modeling lies in data analysis, classification, annotations, and regression of data ranging from text, speech, music, video, etc. Each topic is associated with a probability distribution over words in document collections; this distribution is unique to each collection. For example, the topic of POLITICS can be found in collections from many countries, but Narendra Modi would generally become frequently in collections of India. The text collections are available in different formats and sizes. The scientific articles, newspaper articles, and other medical documents are lengthy documents and so topic modeling on such articles can be done using standard techniques like Latent Semantic Analysis (LSA), pLSA (probabilistic LSA), Latent Dirichlet Allocation (LDA), and other versions of LDA. These standards methods assume that documents consist of diverse topics and the words are correlated.
Short texts like customer reviews, tweets, online chats, abstracts of articles, web posts are sparse texts so topic modeling on such text collections is challenging and inefficient when performed through mentioned standard methods. The unsupervised topic models such as LDA are good for clustering large-scale collections of annotated data. The information extraction from such collections has conceptually focused on a single collection of text but poses inadequacy for collections from different cultural backgrounds. Topic modeling can help in identifying associated medical terms or co-occurring symptoms in medical documents and ongoing trends in the domain. To perform topic modeling especially on text data we focus our study on abstracts from papers on Depression and substance abuse. Depression among adolescents is a serious problem in society. Adolescent depression is directly linked to suicide attempts, self-medication, mental and physical wellbeing (Wang, 2016). Among adolescents, a big health issue is substance use and depression (Fergusson & Boden, 2008). Research in this area is considerable which involves understanding the risk associated with it. With the help of topic modeling techniques, a great deal of information can be extracted related to mental disorders, their co-occurrence (Thapar et al., 2012). A very good understanding of the start of substance use is required to prevent it. Moreover, understanding the connection between substance use and depression is needed to complete such efforts. Text mining is used to find useful patterns and currents trends in the text corpus. Text mining is a very good filtering process where you remove irrelevant documents and focus on only useful ones (Lan et al., 2018), (Holzinger, 2014).
The motivation behind the research is to do a systematic study of various topic modeling techniques and to see their efficiency in analyzing large text corpora. In addition, evaluating and comparing different topic modeling techniques to identify limitations and strengths of each model. To perform topic modeling on mainly scientific articles, English text, and abstracts. The online public libraries have articles related to various health issues and issues related to depression among adolescents. These libraries have a vast collection of articles on substance use and tobacco use among adolescents that can be used for various research purposes. Many review articles and research articles summarize the outcome but mainly focus on certain issues only. Text mining utilizes methods like machine learning to find useful patterns in text data. We used papers related to substance use to discover primary reasons as to why substance use among adolescents is increasing.