Sentiment analysis based on deep learning

• Main Author: Jiang, Qi
• Other Authors: Mao Kezhi
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/149022
• Institution: Nanyang Technological University

As millions of messages are posted and thousands of articles are published every day, a lot of information is stored in form of natural unstructured text. Natural Language Processing (NLP) aims to extract information from text and understand text using computational methods. One of the most important tasks of NLP is sentiment analysis, also known as opinion mining, which studies people’s opinions, sentiment polarities, emotions, and attitudes. Sentiment analysis is of great significance as it can be applied to and benefit a wide range of industries including business sales, governments, social media etc. It is a challenging task and has been studied for decades with various rule-based approaches and machine learning approaches. This project aims to investigate sentiment analysis with deep learning, a machine learning method using large multi-layer artificial neural networks. In particular, this project focuses on sentiment classification which detects the polarity within text and classifies it into positive and negative classes. Fine-grained sentiment classification is more precise and more difficult than 2-way sentiment classification as it expands the polarity categories to 5 classes: very positive, positive, neutral, negative and very negative. In this project, different neural language models including CNN-based models, RNN-based models and BERT are studied, and their performances on sentiment classification are tested, compared, and analyzed. Then, knowledge-aided dual-channel CNN (KDCNN) model is proposed to utilize external knowledge for sentiment analysis to improve CNN’s performance. The external knowledge includes positive and negative sentimental lexicons, negation words, and intensity words. KDCNN can not only improve the precision of sentiment classification but also has less trainable model parameters and less reliance on large amount of training data, which can reduce the cost of training.