Discourse-Driven Argument Mining in Scientific Abstracts

Abstract

Argument mining consists in the automatic identification of argumentative structures in texts. In this work we address the open question of whether discourse-level annotations can contribute to facilitate the identification of argumentative components and relations in scientific literature. We conduct a pilot study by enriching a corpus of computational linguistics abstracts that contains discourse annotations with a new argumentative annotation level. The results obtained from preliminary experiments confirm the potential value of the proposed approach.

Argumentation Mining in Scientific Literature for Sustainable Development

“…Regarding argument relations, there have been also attempts to distinguish them into supporting and attacking. However, Accuosto and Saggion (2019) did not found any attack relation in 60 abstracts of the SciDTB corpus (Yang and Li, 2018). Also, Lauscher et al (2018) and Mayer et al (2020) found that the number of attacking relations in full papers and abstracts is relatively low.…”

“…The most common argumentative relationships employed are the "support" and "attack" ones (Peldszus and Stede, 2013;Mayer et al, 2020). Researchers also adapt additional relation types from Rhetorical Structure Theory (Mann and Thompson, 1988) such as "semantically same" (Lauscher et al, 2018), "detail", "sequence" (Kirschner et al, 2015), and "additional" (Accuosto and Saggion, 2019). Although those fine-grained relation types are of great value, their presence in the context of a scientific abstract is limited.…”

“…The most recent methods for the AU classification task leverage neural networks architectures. A common architecture is the one comprising a BiL-STM layer followed by a CRF layer for sequence tagging (Achakulvisut et al, 2019;Accuosto and Saggion, 2019). Mayer et al (2020) evaluate many different neural network models and shows that those based on Transformers have better performance on the AU extraction and the relation classification tasks.…”

“…Most of the datasets based on scientific literature are domain specific with strong emphasis on the biomedical domain (Blake, 2010;Alamri and Stevenson, 2016;Achakulvisut et al, 2019;Mayer et al, 2020). Other domains are educa-tion (Kirschner et al, 2015), computer graphics (Lauscher et al, 2018), and computational linguistics (Accuosto and Saggion, 2019). SciARK, as a multidisciplinary dataset, includes abstract from biomedical, social, environmental and other domains.…”

Comparative Analysis of Rhetorical and Argumentative Structures in the Study of Popular Science Discourse

“…In the presence of such a corpus, the task is to establish the relationship between rhetorical and argumentative structure of the text, the correlation of the components and relations between them in order to use existing resources and tools of discourse analysis to extract arguments. So, [9] describe the development and use in experiments of a two-level corpus of 112 short texts written in the genre of argumentative essays, and in [1], the material for annotation are scientific articles from the field of computer linguistics.…”