Aspect-Based Sentiment Analysis Using Convolutional Neural Network and Bidirectional Long Short-Term Memory

摘要

In order to improve performance of previous aspect-based sentiment analysis (ABSA) on restaurant reviews in Indonesian language, this paper adapts the research achieving the highest F1 at SemEval 2016. We use feedforward neural network with one-vs-all strategy for aspect category classification (Slot 1), Conditional Random Field (CRF) for opinion target expression extraction (Slot 2), and Convolutional Neural Network (CNN) for sentiment polarity classification (Slot 3). Aside from lexical features we also use additional features learned from neural networks. We train our model on 992 sentences and evaluate them on 382 sentences. Higher performances are achieved for Slot 1 (F1 0.870) and Slot 3 (F1 0.764) but lower on Slot 2 (F1 0.787).