On the Contribution of Specific Entity Detection in Comparative Constructions to Automatic Spin Detection in Biomedical Scientific Publications

In this article, we address the problem of providing automated aid for the detection of misrepresentation (“spin”) of research results in scientific publications from the biomedical domain. Our goal is to identify automatically inadequate claims in medical articles, i.e. claims that present the beneficial effect of the experimental treatment to be greater than it is actually proven by the research results. To this end, we propose a Natural Language Processing (NLP) approach. We first make a review of related work and an NLP analysis of the problem; then we present our first results obtained on the articles that report results of Randomized Controlled Trials (RCTs), i.e. clinical trials comparing two or more interventions by randomly assigning them to patients. Our first experiments concern the identification of entities specific to RCTs (outcomes and patient groups), obtained with basic methods (local grammars) on a corpus extracted from the PubMed open archive. We explore the possibility to extract outcomes from comparative constructions that are commonly used to report results of clinical trials. Our second set of experiments consists in extracting outcomes from a manually annotated corpus using deep learning methods.