Filtering Entity Centric Documents using Numerics and Temporals features within RF Classifier.
Vincent Bouvier, Patrice Bellot · Text REtrieval Conference · 2013
This article introduces the system designed to work for the Knowledge Base Acceleration (KBA) track at Text REtrieval Conference (TREC). This is the second time this track is run with yet this year there is even more challenge. KBA is focused on keeping up to date knowledge bases (KB) such as Wikipedia1 (WP) where each KB node is considered as an entity (WP page for wikipedia example). It has been shown in (Frank et al., 2012) that the time lag between the publication date of cited news articles and the date of an edit to wikipedia article creating the citation can be really big (median 356 days) for non-popular entities. KBA is to give a chance to non-popular entities information to be updated as soon as a useful information is published on the internet. The KBA organizers have built up a stream-corpus which is a huge corpus of timestamped web documents that can be processed chronologically. Hence it is possible to simulate a real time system. The documents come from newswires, blogs, forums, review, memetracker... . In addition, a set of target entities, coming from wikipedia or from twitter, has been selected for their ambiguity or unpopularity. And last but not least, more than 60,000 documents have been annotated so that systems can train on it. The train period starts on documents published from october 2011 until february 2012, and the test period starts from february 2012 to february 2013. The KBA track is divided in two tasks: CCR (Cumulative Citation Recommendation) and SSF (Streaming Slot Filling). CCR task is to filter out documents worth citing in a profile of an entity (e.g., wikipedia or freebase article). SSF task is to detect changes on given slots for each of the target entities. This article is focused only on CCR task. In CCR task, the system is to filter out documents relative to target entities out from the stream-corpus. The system must also be able to give the usefulness