Exploiting Position Bias for Robust Aspect Sentiment Classification

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1352–1358 August 1–6, 2021. ©2021 Association for Computational Linguistics 1352 Exploiting Position Bias for Robust Aspect Sentiment Classification Fang Ma∗, Chen Zhang∗, Dawei Song† School of Computer Science, Beijing Institute of Technology Beijing, China {mfang,czhang,dwsong}@bit.edu.cn Abstract Aspect sentiment classification (ASC) aims at determining sentiments expressed towards dif- ferent aspects in a sentence. While state-of- the-art ASC models have achieved remarkable performance, they are recently shown to suffer from the issue of robustness. Particularly in two common scenarios: when domains of test and training data are different (out-of-domain scenario) or test data is adversarially perturbed (adversarial scenario), ASC models may at- tend to irrelevant words and neglect opinion expressions that truly describe diverse aspects. To tackle the challenge, in this paper, we hy- pothesize that position bias (i.e., the words closer to a concerning aspect would carry a higher degree of importance) is crucial for building more robust ASC models by reduc- ing the probability of mis-attending. Accord- ingly, we propose two mechanisms for cap- turing position bias, namely position-biased weight and position-biased dropout, which can be flexibly injected into existing models to en- hance representations for classification. Exper- iments conducted on out-of-domain and adver- sarial datasets demonstrate that our proposed approaches largely improve the robustness and effectiveness of current models.1 1 Introduction Aspect sentiment classification (ASC) is an impor- tant sub-task of sentiment classification. It aims to identify the sentiment polarity (i.e., negative, neutral, or positive) of a specified aspect in a sen- tence. Take “Great food but the service was bad.” as an example. For aspects food and service, their corresponding sentiment polarities are positive and negative, respectively. ∗ Fang Ma and Chen Zhang contribute equally to this work. The order is determined alphabetically. †Dawei Song is the corresponding author. 1The code and preprocessed data are available at https: //github.com/BD-MF/POS4ASC. Scenario Example Pred./Lb. I.D. Great food but the service was bad ! neg./neg. O.O.D. The battery has never worked well . pos./neg. Adv. Awful food but the service was great ! neg./pos. Figure 1: An illustration of how an ASC model IAN (Ma et al., 2017) might fail. Gradient saliency maps (Simonyan et al., 2014) with respect to the em- bedding of each word in I.D., O.O.D., and Adv. sce- narios, along with the model predictions (Pred.) and corresponding ground truth labels (Lb.), are provided. Underlined words are aspects. A challenge in ASC is how to model semantic relations between aspect terms and their contexts, which requires an ASC model to be only sensi- tive to the sentiment words actually depicting the target aspect terms. Although previous ASC mod- els (Tang et al., 2016b; Li et al., 2018; Zhang et al., 2019a; Xu et al., 2019; Wang et al., 2020; Tang et al., 2020) have achieved promising results by modeling complex interactions between aspects and contexts, these models have recently been shown to suffer from the lack of robustness (Xing et al., 2020). The issue is particularly severe in …

Content truncated to 500 words