Causal Intervention Improves Implicit Sentiment Analysis

Proceedings of the 29th International Conference on Computational Linguistics, pages 6966–6977 October 12–17, 2022. 6966 Causal Intervention Improves Implicit Sentiment Analysis Siyin Wang†, Jie Zhou†*, Changzhi Sun§, Junjie Ye†, Tao Gui†, Qi Zhang†, and Xuanjing Huang† †School of Computer Science, Fudan Univerisity §ByteDance AI Lab {siyinwang20, jie_zhou, jjye19}@fudan.edu.cn {tgui, qz, xjhuang}@fudan.edu.cn sunchangzhi@bytedance.com Abstract Despite having achieved great success for senti- ment analysis, existing neural models struggle with implicit sentiment analysis. This may be due to the fact that they may latch onto spu- rious correlations (“shortcuts”, e.g., focusing only on explicit sentiment words), resulting in undermining the effectiveness and robust- ness of the learned model. In this work, we propose a CausaL intervention model for im- plicit sEntiment ANalysis using instrumental variable (CLEAN). We first review sentiment analysis from a causal perspective and analyze the confounders existing in this task. Then, we introduce an instrumental variable to eliminate the confounding causal effects, thus extracting the pure causal effect between sentence and sentiment. We compare the proposed CLEAN model with several strong baselines on both the general implicit sentiment analysis and aspect- based implicit sentiment analysis tasks. The re- sults indicate the great advantages of our model and the efficacy of implicit sentiment reason- ing. 1 Introduction The remarkable success that the field of sentiment analysis has achieved in the past few years has been derived from the use of increasingly high-capacity neural models to extract correlations from data (Pe- ters et al., 2018; Devlin et al., 2018; Liu et al., 2019). Although having reached state-of-the-art results, correlational predictive models can be un- trustworthy (Guidotti et al., 2018): they may latch onto spurious correlations (“shortcuts”), leading to poor generalization. One shortcut might be the explicit sentiment word which is a powerful feature cue. Unfortu- nately, such a shortcut severely harms the general- ization and the robustness of the learned models in implicit sentiment analysis (ISA), where there are no explicit sentiment words about the topic *Corresponding authors. sentiment word confounder �implicit, neg� The food was definitely good, but when all was said and done, I just couldn't justify it for the price (including 2 drinks, $100/person)... �explicit, pos� g Figure 1: An examples of confounding factors in im- plicit sentiment analysis for ABSA. or aspect (Russo et al., 2015). Figure 1 gives a sample of aspect-based sentiment analysis (ABSA) (Zhou et al., 2020b,a), which aims to predict the sentiments of the aspects in the sentence. The aspect “food” has explicit sentiment words “def- initely good”, but aspect “price” does not. If the model thoughtlessly relies on shortcuts to senti- ment words, it may make an incorrect sentiment prediction about the aspect “price”. In fact, there are many other kinds of shortcuts that models may learn, for example, the rhetorical question mood expressed by the users (Ranganath et al., 2018) and the co-occurrence of neutral words and sentiment polarities (Wang and Culotta, 2020). The above shortcomings can potentially be ad- dressed by the causal perspective: knowledge of causal relationships between observations and …

Content truncated to 500 words