Aspect-based sentiment analysis for user profiles

• Main Author: Ng, Zhiyong
• Other Authors: Lihui CHEN
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Document and text processing; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/149012
• Institution: Nanyang Technological University

With the rise of online e-commerce shopping, spam and scam through online reviews have become a burgeoning problem. Aspect sentiments have been used to help create user profiles which in turn assist in detecting deceptive reviews. This experimental study aimed to implement novel architectures to evaluate and improve the aspect-based sentiment analysis (ABSA) subtask. A combination of text corpora from different product domains, namely SemEval14, Yelp.com, and Edmunds, were used for the aspect extraction and aspect sentiment classification stage. Previous literature had implemented both supervised deep learning and unsupervised pattern-based approaches to extract aspect terms. Hence, we sought to improve the processes mentioned above to generate higher quality aspect terms classified into different aspects of sentiment polarity. Various word embedding models such as Skip-Gram, CBOW, and GloVe were used to create word vectors for the supervised aspect extraction approach. Besides, dependency and POS tag features were introduced into the word vectors to test their efficacy during aspect term extraction. The word vectors were used to train a convolutional neural network (CNN) to predict aspect term locations in a text corpus. Furthermore, an unsupervised syntactic pattern-based system was implemented using dependency parsers such as POS tag and SpaCy. The aspect terms extracted using the experimental models were evaluated with VaderSentiment, Sentiwordnet, and E2E-BERT-ASC for the aspect sentiment classification task. The project concluded with the most appropriate aspect extraction, sentiment classification, and explanation of the findings. Keywords: word embeddings, aspect extraction, aspect tagging, supervised learning, glove, word2vec, spacy, vadersentiment, opinion lexicon, convolutional neural network