Extracting opinion expression with neural attention

Abstract

Extracting opinion expressions from raw text is a fundamental task in sentiment analysis and it is usually formulated as a sequence labeling problem tackled by conditional random fields (CRFs). However CRF-based models usually need abundant hand-crafted features and require a lot of engineering effort. Recently deep neural networks are proposed to alleviate this problem. In order to extend neural-network-based models with ability to emphasize related parts in text, we propose a novel model which introduces the attention mechanism to Recurrent Neural Networks (RNNs) for opinion expression sequence labeling. We evaluate our model on MPQA 1.2 dataset, and experimental results show that the proposed model outperforms state-of-the-art CRF-based model on this task. Visualization of some examples show that our model can make use of correlation of words in the sentences and emphasize the crucial parts for this task to improve the performance compared with the vanilla RNNs.