Database Reference
In-Depth Information
AverageBloat
MaximumBloat
Lidstone
1.E-20
1.E-16
1.E-12
1.E-08
1.E-04
1.E+00
1.E+12
1.E+10
1.E+08
1.E+06
1.E+04
1.E+02
1.E+00
FIGURE 10.29
: Estimated bloat for various values of
for a specific
estimated index size of 145 MB. The y-axis uses a log scale.
questions with one target type and one or more keywords to match represent
a very common class of information need. In Section 10.2 we described the
subsystem that interprets a natural language query into a semistructured
form, comprising one target answer type and a set of keywords to match. In
Section 10.3 we described a machine learning approach to learning a scoring
function that rewards proximity between instances of the target type and
keyword matches. In Section 10.4 we described those modules of IR4QA that
are responsible for index management and query execution. IR4QA is public-
domain code that is available for non-profit use.
5
10.5.2 Ongoing and Future Work
Since 2004, when we proposed (5) and began initial work on the project
reported here, significant advances have been made by several groups at
building systems for entity search. The RADAR project at CMU
6
,the
Haystack project at MIT
7
and desktop search offerings from several companies
represent entities and relations in personal data (people, places, files, emails,
addresses) in a relational or graphical format and enable type-oriented entity
searches. There are even whole workshops (11; 22; 19) dedicated to ranking
in databases and novel IR systems.
The EntityRank project at UIUC (9) is a recent development that is
especially noteworthy. EntityRank allows multiple atypes in a query, which
are to be collectively bound to form a record-like structure. For example, to
find the customer service phone number of Amazon.com, one may use the
query
ow(amazon customer service #phone)
,where
#phone
is the atype
and
ow
enforces that the selector tokens appear to the left of the mention of the
Search WWH ::
Custom Search