Aspect Extraction from Customer Reviews Using Convolutional Neural Networks

Abstract

Convolutional Neural Networks (CNN) have emerged as a viable solution for text classification, including aspect extraction. Yet the existing state-of-the-art CNN architectures used in aspect extraction do not incorporate improvements (e.g. non-static CNN, multi-kernel convolution layers, and dropout regularization) that have been identified as beneficial for general text classification tasks. In this research, we present an improved CNN architecture with these enhancements for aspect extraction. Our modifications to the basic CNN model improve feature extraction of the convolutional layer, while allowing the dense layers to represent complex non-linear relationships between selected features and the output layer. Furthermore, we show that using non-static CNN instead of static CNN further fine-tunes the word embedding features for the specific domain of customer reviews in the absence of domain-specific corpora to train word2Vec models. We show that using Skip-gram trained word2vec models opposed to Continuous bag of word (CBOW) architecture improves the quality of word2vec embedding features for aspect extraction and text classification tasks. This new CNN model and baseline CNN are evaluated using SemEval-2016 Task 5 Restaurant and Laptop domain datasets. Our CNN model shows significant improvements over the baseline CNN model for aspect extraction from customer reviews for both restaurant and laptop domains. It also outperforms the currently available state-of-the-art systems for restaurant domain.