UiO Language Technology Internationally Leading
The IFI Language Technology Group, collaborating with the University of Potsdam, developed the best overall system in the 2010 Shared Task (i.e. system bake-off) of the International Conference on Computational Natural Language Learning. In this competition, participating systems were evaluated quantitatively on how well they can automatically resolve so-called hedges in biomedical research literature.In scientific texts, expressions like 'can', 'suggest', 'likely', and many others often (but crucially not always) serve to signal uncertainty or speculation. The UiO system detects with high accuracy whether or not there is speculation, and if so which parts of a sentence are affected. The technology combines machine learning with in-depth grammatical analysis; two techniques with a strong tradition at UiO.Leading groups world-wide participated in the competition, including prestigious universities like Cambridge, Columbia, Shanghai Jiao Tong, Stanford, or Tokyo. The problem was subdivided into two subtasks: (a) flagging speculative sentences, and (b) finding specific hedges and affected sentence parts. The UiO team ranked fourth (of 24 submissions) on the first subtask, and third (of 15) on the second. As the only system performing well on both tasks, when combining results across the two tracks, the UiO system was best overall.The successful team was comprised of post-doctoral fellow Erik Velldal, researcher Lilja Øvrelid, and professor Stephan Oepen. More information is available through two research articles that have been submitted for publication, or directly from Erik Velldal upon request.