Ontology-based Technical Text Annotation

Powerful tools could help users explore and maintain domain specific documentations, provided that documents have been semantically annotated. For that, the annotations must be sufficiently specialized and rich, relying on some explicit semantic model, usually an ontology, that repre- sents the semantics of the target domain. In this paper, we learn to annotate biomedical scientific publications with respect to a Gene Regulation Ontology. We devise a two-step approach to an- notate semantic events and relations. The first step is recast as a text segmentation and labeling problem and solved using machine translation tools and a CRF, the second as multi-class classi- fication. We evaluate the approach on the BioNLP-GRO benchmark, achieving an average 61% F-measure on the event detection by itself and 50% F-measure on biological relation annotation. This suggests that human annotators can be supported in domain specific semantic annotation tasks. Under different experimental settings, we also conclude some interesting observations: (1) For event detection and compared to classical time-consuming sequence labeling approach, the newly proposed machine translation based method performed equally well but with much less computation resource required. (2) A highly domain specific part of the task, namely proteins and transcription factors detection, is best performed by domain aware tools, which can be used separately as an initial step of the pipeline.