Information Technology Reference
In-Depth Information
have shown that the above method can significantly outperform several non-learning
baselines, including BM25, in terms of several different measures such as error rate,
R-precision, and P@ k .
14.2.2 Quantity Consensus QA
Quantity search is an important special case of entity search. A quantity may be a
unitless number or have an associated unit like length, mass, temperature, currency,
etc. TREC-QA track 2007, 2006, and 2005 have 360, 403 and 362 factoid queries,
of which as many as 125, 177, and 116 queries seek quantities. As against “spot
queries” seeking unique answers like date of birth, there is uncertainty about the
answer for quantity consensus queries (e.g., “driving time from Beijing to Shanghai”
or “battery life of iPhone 4”). To learn a reasonable distribution over an uncertain
quantity, the user may need to browse thousands of pages returned by a regular
search engine. This is clearly not only time consuming but also infeasible for users.
In [ 1 ], learning-to-rank technologies are applied to solve the problem of quantity
consensus QA. This is, however, not an easy task. The difficulty lies in that the an-
swer should be derived from all the documents relevant to the query through deep
mining, and simple ranking of these documents cannot provide such information.
To tackle this challenge, evidence in favor of candidate quantities and quantity in-
tervals from documents (or snippets) is aggregated in a collective fashion, and these
intervals instead of original documents (or snippet) are ranked.
For this purpose, one first finds candidate intervals related to the quantity consen-
sus query. This is done by first processing each snippet and finding the quantity (in-
cluding unit) that it contains. Then these quantities are put into the x -axis, and their
merits are evaluated using a merit function. The function basically considers the
number of quantities falling into an interval, and the relevance of the corresponding
snippets to the query. The intervals with top merits are selected as candidate inter-
vals and passed onto the learning-to-rank algorithm. Note that each selected interval
is associated with several snippets whose quantities fall into this interval.
Second, one needs to extract a set of features as the representation of an candidate
interval. Specifically in [ 1 ], the following features are extracted:
Whether all snippets associated with the interval contain some query word.
Whether all snippets associated with the interval contain the minimum IDF query
word.
Whether all snippets associated with the interval contain the maximum IDF query
word.
Number of distinct words found in snippets associated with the interval.
Number of words that occur in all snippets associated with the interval.
One minus the number of distinct quantities mentioned in snippets associated
with the interval, divided by the length of the interval.
Number of snippets associated with the interval, divided by the total number of
snippets retrieved for the query.
Search WWH ::




Custom Search