CONTEXT-AWARE SEMANTIC TEXT MINING AND REPRESENTATION LEARNING FOR TEXT DISAMBIGUATION AND ONLINE HARASSMENT CLASSIFICATION
Date:
Our contribution can be divided into two main parts; the first part focuses on the text ambiguity problem, and the second part focuses on the text classification problem, which are two related tasks in NLP. While analyzing and designing algorithms for text understanding and representation learning, we introduce algorithms to better understand the text and its exact meaning when there are different possible meanings for words present in the text. This problem has been known in NLP as a Word Sense Disambiguation (WSD) problem. In the first part of this thesis, we analyze the effect of different current available methods in text embedding on the WSD task, and based on the observations and experiments, we introduce a new method for text representation learning. In addition to general English text, to evaluate our method, we analyze the effect of our representation on Biomedical text as an application. This analysis shows how effective these embeddings are in capturing the context when we are looking to find the correct meaning behind the words in biomedical texts. We also investigate the problem of text classification in this study. Text classification is one other relevant problem in NLP to the problem of WSD. We consider a collection of tweet posts and try to classify them into two classes, one if a tweet includes harassment, and the other is the class of tweets without harassment. We apply classical machine learning approaches and show the effects and differences between them. In addition to this binary classification investigation, we focus on the first class, the tweets including harassment. We analyze the tweets and classify which type of harassment is a tweet using classical machine learning approaches, including logistic regression, Gaussian naive Bayes, decision trees, support vector machines, random forest, multi-layer perceptron, and AdaBoost in chapter five. The last chapter uses a deep learning approach, the graph convolution approach, to solve this problem. Our experiments show how effective using this deep learning method is compared to the previous classical machine learning approache