Abstract

IR techniques have proven quite successful at locating within large collections of documents those relevant to a user’s query. Often, however, the user wants not whole documents but brief answers to specific questions: How old is the President? Who was the second person on the moon? When was the storming of the Bastille? Recently, a number of research projects have investigated the computational techniques needed for effective performance at this level of granularity, focusing just on questions that can be answered in a few words taken as a passage directly from a single text (leaving aside, for the moment, the answering of longer, more complex answers, such as stories about events, descriptions of objects, compare&contrast discussions, arguments of opinion, etc.).
The systems being built in these projects exhibit a fairly standard structure: all create a query from the user’s question, perform IR with the query to locate (segments of) documents likely to contain an answer, and then pinpoint the most likely answer passage within the candidate documents. The most common difference of approach lies in the pinpointing. A ‘pure IR’approach would segment each document in the collection into a series of mini-documents, retrieve the segments that best match the query, and return them as answer. The challenge here would be to make segments so small as to be just answer-sized but still large enough to be indexable. A ‘pure NLP’approach would be to match the parse and/or semantic interpretation of the question against the parse and/or semantic interpretation of each sentence in the candidate answer-containing documents, and return the …

Metadata

publication

TREC 52, 53-56, 2000

year

2000

publication date

2000/11/13

authors

Eduard H Hovy, Laurie Gerber, Ulf Hermjakob, Michael Junk, Chin-Yew Lin

link

http://trec.nist.gov/pubs/trec9/papers/webclopedia.pdf

resource_link

http://trec.nist.gov/pubs/trec9/papers/webclopedia.pdf

journal

TREC

volume

52

pages

53-56