Goal-Driven Answer Extraction

We describe the structure and functioning of an answer-extraction system built from the ground up, in only three person-months, using shallow text-processing techniques. Underlying these techniques is the attribution to each question of a goal type serving to characterize the outward form of candidate answers. The goal type is used as a filter during long-answer extraction, essentially a small-scale IR process which returns 250-byte windows rather than whole documents. To obtain short answers, strings matching the goal type are extracted from these windows and ranked by heuristics. TREC-9 performance figures show that our system has difficulty dealing with brief, definition-based questions of the kind most likely to be posed by users. We propose that specialized QA strategies be developed to handle such cases. 1 System Overview In the following we shall refer to our system by the working title xr3 , which stands for eXtraction de R

• Type: Conference
• Year: 2000
• Where: TREC
• Authors: Michael Laszlo, Leila Kosseim, Guy Lapalme
• Keywords: Goal Type, Shallow Text-processing Techniques, Short Answers, TREC 2000, TREC 2008