HYBRID MACHINE LEARNING AND DEEP LEARNING MODELS FOR SENTIMENT ANALYSIS OF STUDENT FEEDBACK AND SARCASM DETECTION

Abstract

Sentiment analysis or opinion mining is classification of feedback/reviews given by people about a product, event, service, or policy. It is used in various domains such as finance, business, healthcare, politics, education, and many more. Education plays an important role in the growth of a country. Feedback given by students is a valuable resource that can be analyzed to know which aspect (teacher, curriculum, facilities) needs to be improved for better education quality. But there is a scarcity of resources and tools, such as annotated corpora, lexicons, and benchmark datasets, making sentiment analysis challenging in the education domain. Hence it is worth creating datasets and developing efficient machine learning and deep learning models in the education domain for evaluating instructor performance. Sarcasm is a special form of sentiment in which people use positive words to express negative feelings. For example, "I love this phone," in this sentence, the positive word love is used to express the negative feeling. While communicating verbally, people express sarcasm using hand gestures, facial expressions, and eye movements. These clues are missing in text data, making sarcasm detection challenging. It is worth developing deep learning models for sarcasm detection from textual data. The major objectives of the present work are as follows: (i) Creation of datasets containing feedback about instructor given by students for evaluating instructor performance. (ii) Analysis and exploration of feature selection, feature extraction, machine learning (classification and clustering) models, and development of machine learning and deep learning models to evaluate instructor performance based on the student’s feedback. (iii) Development of deep learning models for sarcasm detection on social media textual data. The thesis is organized into six chapters. Chapter 1: Introduction The introduction contains an overview of sentiment analysis, and levels of sentiment analysis. It includes detailed information about different sentiment analysis techniques (machine learning, lexicon-based, and hybrid methods). It includes applications of sentiment analysis in various domains such as education, healthcare, recommendation system, and government intelligence. It presents several research challenges, objectives, and contributions of the thesis. Chapter 2: This chapter covers the concept of machine learning, feature selection, feature extraction, and ensemble models for mining student feedback about the instructor in higher education. This chapter has two parts, in first part presents an exploratory analysis of feature selection, feature extraction, and classification algorithms to find the best combination for prediction of instructor performance. The second part contains the proposal and development of hybrid machine learning model which is a fusion of stacking and voting ensemble to evaluate instructor performance. It presents the comparative analysis proposed approach with state-of-the-art methods and baseline models. A statistical method (ANOVA) used to validate the results statistically is also presented in this chapter. Chapter 3: This chapter covers the concept of different types of text representation methods and ensemble model. It presents a developed two-level hybrid model which applies the concept of feature fusion and hyperparameter optimization of ensemble model to evaluate instructor performance based on textual student feedback. This chapter also presents the concept of web scrapping to create the dataset in resource constrained education domain. It presents the comparative analysis of proposed approach with state-of-the-art methods and baseline models. Chapter 4: This chapter covers the concept of domain specific transformer model for generating contextual word embeddings. It presents a development process of domain specific (social media) transformer model called LMTweets. It presents a developed hybrid deep learning model for sarcasm detection from three benchmark textual datasets. It presents the comparative analysis of developed approach with state-of-the-art methods and baseline models. A statistical method (ANOVA) used to validate the results statistically is also presented in this chapter. Chapter 5: This chapter presents a developed hybrid deep learning model for sarcasm detection on three benchmark datasets. It also presents the comparative study of recent studies on multimodal datasets and textual datasets. It presents the concepts of contextual word embedding generated from transformer model and its importance in text classification task. It presents the comparative analysis of developed approach with state-of-the-art methods and baseline models. A statistical method (ANOVA) used to validate the results statistically is also presented in this chapter. Chapter 6: This chapter concludes the thesis and presents the future scope of the work and research directions.