ENHANCED DEEP NEURAL NETWORK FOR ASPECT-OPINION TERMS CO-EXTRACTION

Abstract

Aspect-Opinion Co-extractions are two important sub-tasks of  aspect-based sentiments analysis (ABSA)  that involve simultaneous identification of product’s aspects and the associated opinion words from user textual reviews. Traditional approaches to the aspect-opinion co-extraction typically depend on the handcrafted and rule-based methods which are known to be labor-intensive and often less accurate. Recently, Convolutional Neural Networks (CNNs) have been widely applied for the aspect-opinion co-extraction task. However, the existing approaches rely solely on the word embedding models such as  Glove or word2vec. As such they cannot guarantee more fine-grained semantic information due to the proble of the  . Thus, in this paper, wet propose a lexicalized CNN (LCNN) technique that can help to better capture the fine-grained semantic information for better coextraction process. The proposed method consists of lexicon embeddings in addition to the word embeddings as inputs to the network. For the word embedding input, we use general embedding (GE) which is pre-trained based on a large corpus of Google news and domain-specific embedding (DSE) which is trained based on the Amazon and Yelp reviews. For the lexicon embeddings, we use lexicon resources based on the SenticNet model. The word embedding and lexicon embedding are concatenated and fed into the convolutional network to generate local features which are then max-pooled to generate input to the softmax function for the final aspect-opinion co-extraction task. The proposed model was evaluated using various benchmark datasets and the experimental results have shown that our proposed model performed better than to the baseline approaches.