Using the Argumentative Structure of Scientific Literature to Improve Information Access

MEDLINE/PubMed contains structured abstracts that can provide argumentative labels. Selection of abstract sentences based on the argumentative label has shown to improve the performance of information retrieval tasks. These abstracts make up less than one quarter of all the abstracts in MEDLINE/PubMed, so it is worthwhile to learn how to automatically label the non-structured ones. We have compared several machine learning algorithms trained on structured abstracts to identify argumentative labels. We have performed an intrinsic evaluation on predicting argumentative labels for non-structured abstracts and an extrinsic evaluation to predict argumentative labels on abstracts relevant to Gene Reference Into Function (GeneRIF) indexing. Intrinsic evaluation shows that argumentative labels can be assigned effectively to structured abstracts. Algorithms that model the argumentative structure seem to perform better than other algorithms. Extrinsic results show that assigning argumentative labels to non-structured abstracts improves the performance on GeneRIF indexing. On the other hand, the algorithms that model the argumentative structure of the abstracts obtain lower performance in the extrinsic evaluation.