Structural Bias for Aspect Sentiment Triplet Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 6736–6745 October 12–17, 2022. 6736 Structural Bias for Aspect Sentiment Triplet Extraction Chen Zhang , Lei Ren , Fang Ma , Jingang Wang , Wei Wu , Dawei Song Beijing Institute of Technology {czhang,mfang,dwsong}@bit.edu.cn Meituan NLP {wangjingang02,wuwei30}@meituan.com renlei_work@163.com Abstract Structural bias has recently been exploited for aspect sentiment triplet extraction (ASTE) and led to improved performance. On the other hand, it is recognized that explicitly incorporat- ing structural bias would have a negative im- pact on efficiency, whereas pretrained language models (PLMs) can already capture implicit structures. Thus, a natural question arises: Is structural bias still a necessity in the context of PLMs? To answer the question, we pro- pose to address the efficiency issues by using an adapter to integrate structural bias in the PLM and using a cheap-to-compute relative position structure in place of the syntactic de- pendency structure. Benchmarking evaluation is conducted on the SemEval datasets. The re- sults show that our proposed structural adapter is beneficial to PLMs and achieves state-of-the- art performance over a range of strong base- lines, yet with a light parameter demand and low latency. Meanwhile, we give rise to the concern that the current evaluation default with data of small scale is under-confident. Con- sequently, we release a large-scale dataset for ASTE. The results on the new dataset hint that the structural adapter is confidently effective and efficient to a large scale. Overall, we draw the conclusion that structural bias shall still be a necessity even with PLMs.1 1 Introduction Aspect sentiment triplet extraction (ASTE) is a task central to fine-grained opinion mining. Com- pared to aspect sentiment classification that only aims to predict sentiment polarities for various as- pects, ASTE instead extracts descriptive opinion units in the form of triplets (i.e., aspect-opinion- sentiment tuples). For example, (food, great, POS) and (service, dreadful, NEG) are aspect sentiment authors.Jingang Wang and Dawei Song are the corresponding 1Code and data are available at https://github.com/ GeneZC/StructBias. Triplet: Great food but the service was dreadful ! POS NEG Dependency: Great food but the service was dreadful ! AMOD NSUBJ CONJ Figure 1: Triplet (top) and dependency structures (bot- tom) of an illustrative sentence. Spans shaded in yellow are aspects, spans shaded in red are opinions, and arcs indicate either sentiment or structural relations. Irrele- vant structural relations are neglected for brevity. triplets for the sentence in Figure 1 (top), where {POS, NEG, NEU} respectively represent {positive, negative, neutral}. While ASTE can be generally tackled with neural models in either a pipeline manner (Peng et al., 2020) or a multi-task manner (Xu et al., 2020; Zhang et al., 2020; Wu et al., 2020a; Chen et al., 2021a; Xu et al., 2021), the aspect senti- ment triplets can be rather derivable from depen- dency structures (e.g., syntactic dependency trees) with hand-crafted rules (Wu et al., 2009; Sun et al., 2017). For the example in Figure 1 (bottom), the triplets can be recognized via certain structural de- pendency …

Content truncated to 500 words