PUB - Publikationen an der Universität Bielefeld

We explore the possibility of leveraging model explainability methods for weakly supervised claim localization in scientific abstracts. The resulting approaches require only abstract-level supervision, i.e., information about the general presence of a claim in a given abstract, to extract spans of text that indicate this specific claim. We evaluate our methods on the SciFact claim verification dataset, as well as on a newly created dataset that contains expert-annotated evidence for scientific hypotheses in paper abstracts from the field of invasion biology. Our results suggest that significant performance in the claim localization task can be achieved without any explicit supervision, which increases the transferability to new domains with limited data availability. In the course of our experiments, we additionally find that injecting information from human evidence annotations into the training of a neural network classifier can lead to a significant increase in classification performance.