Using LSTM Encoder-Decoder for Rhetorical Structure Prediction

Abstract

The importance of identifying rhetorical categories in texts has been widely acknowledged in the literature, since information regarding text organization or structure can be applied in a variety of scenarios, including genre-specific writing support and evaluation, both manually and automatically. In this paper we present a Long Short-Term Memory (LSTM) encoder-decoder classifier for scientific abstracts. As a large corpus of annotated abstracts was required to train our classifier, we built a corpus using abstracts extracted from PUBMED/MEDLINE. Using the proposed classifier we achieved approximately 3% improvement in per-abstract accuracy over the baselines and 1% improvement for both per-sentence accuracy and f1-score.