Marker Words for Negation and Speculation in Health Records and Consumer Reviews

Conditional random fields were trained to detect marker words for negation and speculation in two corpora belonging to two very different domains: clinical text and consumer review text. For the corpus of clinical text, marker words for speculation and negation were detected with results in line with previously reported interannotator agreement scores. This was also the case for speculation markers in the consumer review corpus, while detection of negation markers was unsuccessful in this genre. Also a setup in which models were trained on markers in consumer reviews, and applied on the clinical text genre, yielded low results. This shows that neither the trained models, nor the choice of appropriate machine learning algorithms and features, were transferable across the two text genres.