BiTeM group report for TREC Chemical IR Track 2011
Julien Gobeill, Arnaud Gaudinat, Patrick Ruch, Emilie Pasche, Douglas Teodoro, Dina Vishnyakova · reroDoc Digital Library · 2011
For the third year, the BiTeM group participated in the TREC Chemical IR Track. For this campaign, we applied strategies that already showed their effectiveness, as the Citations Feedback, which takes benefit from the citations of the retrieved documents in order to re-arrange the ranking. But we also investigated a new inter-lingua model built with chemical annotations with concepts that we automatically mapped into documents. We used the MeSH controlled vocabulary for this purpose. For the Technology Survey task, the fusion of the MeSH and Text models led to a remarkable improvement (+71% for MAP) compared to the Text model alone. The most interesting aspect is that both models are highly complementary as the MeSH model brings 70% of new relevant documents that were not retrieved by the Text model. For the Prior Art task, we showed that there exist patterns of chemical patents that are interconnected (i.e. linked together with direct citations) and that are more likely to be present together in a prior art. Such patterns are efficiently retrieved with our Citations Feedback strategy. On the other hand, we pointed out that the less the prior art of a given topic is interconnected, the less efficient is the Information Retrieval. We hypothesize that such patents have a larger technical focus, maybe represented by a larger set of IPC codes, and then have a lower textual similarity with their prior art documents. These topics should gain to be recognized in order to be treated with complementary techniques.