Two-Level Convolutional Neural Network for Aspect Extraction.

Extract product aspect is an important task in fine-grained sentiment analysis. Though many models have been proposed to solve the problem. They either use handcrafted features or complex neural network architectures. In this paper, we propose a simple Two-level CNN model to extract product aspects from customer reviews, namely Char- and Word-level CNN. The Char-level CNN learns char-level representation of each word (also named morphological information), while the Word-level CNN captures features from the concatenation of char-level representations and word embeddings. Compared to previous neural architectures, our model do not use any external resources like dependency parsing tree or lexicons. To the best of our knowledge, this is the first time to couple Char- and Word-level CNN for aspect extraction. We conduct comparison experiments on two product review datasets. Experimental results demonstrate the effectiveness of our proposed model.