Extracting Structure from Scientific Abstracts Using Neural Networks

摘要

Objective: Knowing the structure of a scientific paper abstract is useful in a variety of NLP tasks such as information extraction or paper writing assistance. Existing methods classify abstract sentences into predefined structural roles, achieving good results but only on specific fields. In this poster we investigate a method that works well across domains. Method: We propose a classifier based on neural networks, and compare it with conventional classifiers, using labeled abstracts from the bio-medical domain as a training corpus, and manually annotated computer science abstracts as test data. Result: Early experiments demonstrate that our neural-network-based method significantly outperforms conventional methods in the cross-domain classification task.