Clause sentiment identification based on convolutional neural network with context embedding

Identifying sentiment of opinion target is an essential component of many tasks for sentiment analysis. We firstly identify the sentiment of the clause in which the specific opinion target lie and then infer the sentiment of opinion target from the sentiment of clause. In order to utilize context more adequately, We propose a novel model using Long Short-Term Memory(LSTM) and Convolutional Neural Network(CNN) together to identify the sentiment of clause. LSTM is used for generating context embedding and CNN is treated as a trainable feature detector. In the experiment using product reviews data, our model outperforms traditional methods in the aspect of accuracy. What's more, the time of model training is acceptable and our model is more scalable because we don't need to discovery rules manually and prepare lots of external language resources which is laborious and time-consuming.

I. Introduction

The topic about product review takes a great proportion in the microblogs and doing sentiment analysis of product opinions is beneficial for product recommendation and product improvement. The fine-grained sentiment analysis includes opinion target extraction and sentiment identification of opinion target. In this paper, we mainly focus on sentiment classification of opinion target if the opinion targets have been extracted.