Argumentative Sequence Labelling Using Transfer Learning

This work presents an approach for Argumentative Sequence Labelling using Transfer Learning. Specifically, a famous pre-trained neural architecture, BERT, has been employed using the Transfer Learning technique known as “fine-tuning” and employing two different data formats for sequence labelling (BIO and BILUO). The neural architecture has been fine-tuned on two famous corpora to recognize not only the boundaries of argumentative units, but also the specific types of argumentative component. The resulting model not only outperforms the results of previous models, but it is also easier to implement, since it does not require highly-engineered features. An evaluation at token-level is performed, as well as a preliminary error analysis.