SenPoi at SemEval-2022 Task 10: Point me to your Opinion, SenPoi

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1313 - 1323 July 14-15, 2022 ©2022 Association for Computational Linguistics SenPoi at SemEval-2022 Task 10: Point me to your Opinion, SenPoi Jan Pfister University of Würzburg Sebastian Wankerl University of Würzburg DHBW Mosbach {pfister,wankerl,hotho}@informatik.uni-wuerzburg.de Andreas Hotho University of Würzburg Abstract Structured Sentiment Analysis is the task of extracting sentiment tuples in a graph struc- ture commonly from review texts. We adapt the Aspect-Based Sentiment Analysis pointer network BARTABSA to model this tuple ex- traction as a sequence prediction task and ex- tend their output grammar to account for the increased complexity of Structured Sentiment Analysis. To predict structured sentiment tu- ples in languages other than English we swap BART for a multilingual mT5 and introduce a novel Output Length Regularization to mit- igate overfitting to common target sequence lengths, thereby improving the performance of the model by up to 70%. We evaluate our approach on seven datasets in five languages including a zero shot crosslingual setting. 1 Introduction The goal of sentiment analysis is to understand a writer’s opinions expressed in a text. In recent years, this topic became of particular research inter- est since with the advent of social media platforms, users were encouraged to share their opinions about a wide range of subjects with the world. For exam- ple, websites like Yelp collect and share opinions about restaurant visits and various online retail- ers allow customers to publish their reviews about items in their assortment. Consequently, various text corpora emerged that made it feasible to apply machine learning based approaches to this task (Ya- dav and Vishwakarma, 2020). Sentiment analysis can be approached with var- ious degrees of granularity. It is possible to only classify the overall polarity of a sentence as pos- itive or negative, or a more precise intermediate value (Nguyen et al., 2020; Devlin et al., 2019). Furthermore, the sentiment can be predicted in a more fine-grained manner, like with regard to the various targets addressed in the sentence. This can be of particular interest if the writer is ambigu- ous in their review and no overall sentiment label can be assigned to a sentence. For example, a cus- tomer might describe the service or the interior of a restaurant as unpleasant but the food itself as good. This can be extended to more complex sentence structures, taking into account not only the aspects or the sentiment, but also its holder (i.e. who is expressing the sentiment) or the opinion term itself, e.g. (Yan et al., 2021; Mukherjee et al., 2021). In this paper, we work on the challenge described by Barnes et al. (2022). Here, the goal is to extract an arbitrary number of opinion tuples from a text. Each tuple can consist of a holder, a target, a sen- timent expression, and a polarity. However, it is also possible that some of these mentioned enti- ties are not present in a certain tuple. Moreover, a sentence may include no sentiment and …

Content truncated to 500 words