Learning Sentence Representations over Tree Structures for Target-Dependent Classification

Proceedings of NAACL-HLT 2018, pages 551–560 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Learning Sentence Representations over Tree Structures for Target-dependent Classification Junwen Duan, Xiao Ding, Ting Liu∗ Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, Harbin, China {jwduan,xding,tliu}@ir.hit.edu.cn Abstract Target-dependent classification tasks, such as aspect-level sentiment analysis, perform fine- grained classifications towards specific targets. Semantic compositions over tree structures are promising for such tasks, as they can poten- tially capture long-distance interactions be- tween targets and their contexts. However, previous work that operates on tree structures resorts to syntactic parsers or Treebank an- notations, which are either subject to noise in informal texts or highly expensive to ob- tain. To address above issues, we propose a reinforcement learning based approach, which automatically induces target-specific sentence representations over tree structures. The un- derlying model is a RNN encoder-decoder that explores possible binary tree structures and a reward mechanism that encourages structures that improve performances on downstream tasks. We evaluate our approach on two bench- mark tasks: firm-specific cumulative abnormal return prediction (based on formal news texts) and aspect-level sentiment analysis (based on informal social media texts). Experimental re- sults show that our model gives superior per- formances compared to previous work that op- erates on parsed trees. Moreover, our approach gives some intuitions on how target-specific sentence representations can be achieved from its word constituents. 1 Introduction We investigate target-dependent classification problem in this paper, with a special focus on the sentence level. Target-dependent classification aims to identify the fine-grained polarities of sen- tences towards specific targets, which is challeng- ing but also important for deep text understand- ing. The definitions of polarity vary across dif- ferent tasks, which can be positive or negative in ∗* Corresponding author Task Example Aspect-level Sen- timent Analysis The food is good but the service is dreadful. Stance Detection I don’t care about global climate change. Firm-specific Financial News Analysis Nike sues Wal-Mart for Patent In- fringement. Table 1: Samples of target-dependent classification tasks. The targets of interest are in bold. aspect-level sentiment analysis, favor or against in stance detection, and rise or drop in financial news analysis towards the stock price movement of a particular firm. Table 1 gives examples of three target- dependent classification tasks. We can find that there can be multiple target mentions in the same text scope, which makes it challenging for generic sentence representation approaches. For the first example, a restaurant manager or a potential cus- tomer may be interested in both food and service; however, the sentiment polarities towards the two targets are different. Hence, it would be beneficial for such tasks to tailor the sentence representations with respect to particular targets. Tree structures are promising for such tasks, as they can potentially capture long-distance depen- dencies between target words and their contexts (Li et al., 2015). Therefore, it is not surprising to find work that exploits the syntactically parsed trees for learning target-specific sentence repre- sentations. …

Content truncated to 500 words