Is position important? deep multi-task learning for aspect-based sentiment analysis
- PDF / 1,627,450 Bytes
- 12 Pages / 595.224 x 790.955 pts Page_size
- 3 Downloads / 177 Views
Is position important? deep multi-task learning for aspect-based sentiment analysis Jie Zhou1
· Jimmy Xiangji Huang2 · Qinmin Vivian Hu3 · Liang He4
© Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract The position information of aspect is essential and useful for aspect-based sentiment analysis, while how to model the position of the aspect effectively during aspect-based sentiment analysis has not been well studied. Inspired by the intuition that the position prediction can help boost the performance of aspect-based sentiment analysis, we propose a Deep MultiTask Learning (DMTL) model, which handles sentiment prediction (SP) and position prediction (PP) simultaneously. In particular, we first use a shared layer to learn the common features of the two tasks. Then, two task-specific layers are utilized to learn the features specific to the tasks and perform position prediction and sentiment prediction in parallel. Inspired by autoencoder structure, we design a position-aware attention and a deep bi-directional LSTM (DBi-LSTM) model for sentiment prediction and position prediction respectively to capture the position information better. Extensive experiments on four benchmark datasets show that our approach can effectively improve the performance of aspect-based sentiment analysis compared with the strong baselines. Keywords Aspect-based sentiment analysis · Multi-task · Deep learning · Position
1 Introduction Aspect-based sentiment analysis [12, 33, 35, 49] plays an important role in sentiment analysis [24, 31] and it has received great attention from both academic communities and industries. The goal of this task is to infer the sentiment polarity (e.g. positive, neutral or negative) of the review towards the given aspect. For example, as shown in Fig. 1, in the sentence “Moules were excellent, lobster ravioli was very salty.” there are two aspects, “moules” and “lobster ravioli”, and the user expresses positive and negative sentiments over them respectively.
Jie Zhou
[email protected] 1
School of Computer Science and Technology, East China Normal University, Shanghai, 200241, China
2
Information Retrieval and Knowledge Management Research Lab, York University, Toronto, Ontario, M3J 1P3, Canada
3
The School of Computer Science, Ryerson University, Toronto, Ontario, M5B 2K3, Canada
4
School of Computer Science and Technology, East China Normal University, Shanghai, 200241, China
It is obvious that the position information has significant effects on the sentiment polarities in the above example shown in Fig. 1. This is because the words with relatively shorter distances (e.g., “excellent”) play a more significant role in judging the sentiment polarity of the aspect (e.g., “moules”). This process is consistent with the way when people judge the sentiment polarity of the aspect. Users usually first observe the neighboring words of the aspect, judging whether these words can decide its sentiment polarity, after that users view other words with long distance. Furthermore, the e
Data Loading...
