Abstract
Argumentation Mining (AM) enables us to reason with natural language; effective AM
depends on our ability to accurately detect claims in text. In structured texts, the claim
detection performance of the state of the art model, ’BERT’, is well understood. But much
less is understood when BERT is applied to less structured text, such as social media, which is
more indicative of “real world natural language”.
We compare BERT’s performance in classical structured texts with that of semi-structured
texts. Then study the performance improvements obtained by pre-training BERT in the same
domain before training BERT for claim detection.
Overall, we have found that BERT performs well on semi-structured text, but pre-training in
the domain is not necessary to obtain good performance. This work can be continued through
the use of classical argumentation mechanisms to relate claims to one another for effective
argumentation mining from social media.