Neural Architecture for Negative Opinion Expressions Extraction

Abstract

Opinion expressions extraction is one of the main frameworks in opinion mining. Extracting negative opinions is more difficult than positive opinions because of indirect expressions. Especially, in the domain of consumer reviews, consumers are easier to be influenced by negative reviews when making decision. In this paper, we focus on the extraction of negative opinion expressions of consumer reviews. State-of-art methods heavily depend on task specific knowledge in the form of handcrafted features and data pre-processing. In this paper, we use a neural architecture by combining word embeddings, Bi-LSTM and CRF. We add a conditional random fields (CRF) layer to bidirectional long-short term memory (Bi-LSTM) recurrent neural network language model, which provides sentence level tag information and improves the result of experiment. Our model requires no feature engineering and outperforms feature dependent methods when experimenting on real-world reviews from Amazon.com.