Aspect-Based Sentiment Analysis Applied in the News Domain Using Rule-Based Aspect Extraction and BiLSTM

Abstract

This research presents an aspect-based sentiment analysis (ABSA) model utilizing a rule-based aspect extraction method combined with the capabilities of deep learning techniques, specifically Bidirectional Long Short-Term Memory (BiLSTM). For precise sentiment prediction in news articles, BiLSTM's ability to capture long-term dependencies in text is essential. The study proposes a methodology that can be summarized into four main phases: data preprocessing, aspect term extraction, word embedding, and sentiment polarity prediction. Initially, the data was pre-processed, afterwards the model utilized a rule-based approach comprised of noun chunk extraction, candidate aspects selection, and candidate similarity filtering to extract the aspect terms in a news article. The text will then be transformed into feature vectors by using a word embedding model before the vectors are passed to the next phase of the model. The sentiment polarity prediction was performed by training the BiLSTM network where it learned to predict sentiment with a three-point granularity: positive, negative, and neutral. Upon testing the model on two test datasets, it yielded a macro-f1 score of 42% on the standard dataset, and a 39% on the frequent dataset wherein its macro-f1 are better than most, except for the BERT Classifier.