CouRGe: Counterfactual reviews generator for sentiment analysis

Thumbnail Image
978-3-031-26438-2_24.pdf(335.98 KB)
Published version
Carraro, Diego
Brown, Kenneth N.
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Journal Issue
Past literature in Natural Language Processing (NLP) has demonstrated that counterfactual data points are useful, for example, for increasing model generalisation, enhancing model interpretability, and as a data augmentation approach. However, obtaining counterfactual examples often requires human annotation effort, which is an expensive and highly skilled process. For these reasons, solutions that resort to transformer-based language models have been recently proposed to generate counterfactuals automatically, but such solutions show limitations. In this paper, we present CouRGe, a language model that, given a movie review (i.e. a seed review) and its sentiment label, generates a counterfactual review that is close (similar) to the seed review but of the opposite sentiment. CouRGe is trained by supervised fine-tuning of GPT-2 on a task-specific dataset of paired movie reviews, and its generation is prompt-based. The model does not require any modification to the network’s architecture or the design of a specific new task for fine-tuning. Experiments show that CouRGe’s generation is effective at flipping the seed sentiment and produces counterfactuals reasonably close to the seed review. This proves once again the great flexibility of language models towards downstream tasks as hard as counterfactual reasoning and opens up the use of CouRGe’s generated counterfactuals for the applications mentioned above.
Natural language processing , Sentiment analysis , Language models , Counterfactual reasoning , Data augmentation
Carraro, D. and Brown, K. N. (2023) ‘Courge: counterfactual reviews generator for sentiment analysis’, AICS2022, in L. Longo and R. O’Reilly (eds) Artificial Intelligence and Cognitive Science. Cham: Springer Nature Switzerland, pp. 305–317. doi: 10.1007/978-3-031-26438-2_24.
Link to publisher’s version