Multi-task Learning Neural Networks for Comparative Elements Extraction

Dianqing Liu, Lihui Wang, Yanqiu Shao

Abstract: Comparative sentences are common in human languages. In online comments, a comparative sentence usually contains the subjective attitude or emotional tendency of a reviewer. Hence, comparative elements extraction (CEE) is valuable for opinion mining and sentiment analysis. Most of the existing CEE systems use rule-based or machine learning approaches that need to construct a rule base or spend a huge amount of effort on feature engineering. These approaches usually involve multiple steps, and the performance of each step relies on the accuracy of the previous step, risking error cascading oversteps. In this paper, we adopt a neural network approach to CEE, which supports end-to-end training and automatic learning of sentence representation. Furthermore, considering the high relevance of CEE and comparative sentences recognition (CSR), we propose a multi-task learning model to combine the two tasks, which can further improve the performance of CEE. Experiment results show that both our neural network approach and multi-task learning are effective for CEE.