Database Reference
In-Depth Information
objective that makes sure the first page lists different persons sharing
the name.
•
Recognizing that the query is a disease name triggers a canned response
from structured records about causes, symptoms and cures.
•
A navigational query that matches businesses in certain broad sectors
triggers a purpose-differentiated first response, e.g., with links for
downloading software, booking tickets, contacting service staff, etc.
Entities and relations form complex networks in our mind, and yet, search
engines seem limited to the paradigms of entering the information need into
a small text box, and getting the response in the form of a ranked list of
URLs with snippets. Many research systems have tried to get past this
simplistic interface, but its simplicity and convenience frequently trump a
more thoughtful design. It appears that any enhancement to the query input
interface must be evolutionary, and allow a fallback to the rudimentary text-
box whenever desired.
However, even the smallest hint of type information in the query helps
immensely. Informal study of Web search query logs reveals many sessions
of 3-8 queries where some words remain fixed, such as
Nikon Coolpix
, while
others come and go, such as
weight
,
light
,
heavy
,
gm
,
oz
, etc. Clearly, the
user wishes to determine the weight of a given camera, and is trying hard to
express this information need through a “telegraphic” Web query. We have
built a prototype metasearch tool where there are
two
query boxes. In one,
the user enters the type of the answer desired, such as
city
. In the other, the
user enters ordinary words to be matched, such as
India
,
Australia
,
cricket
.
This is an approximate representation of the question “In which cities are
cricket matches being played between India and Australia?” Informally, we
have found improvements to response quality if the user takes the trouble of
separating the uninstantiated answer type from words to be matched. For
one thing, responses are not page URLs, but instances of type
city
.
10.1.1 Guessing Answer Types
In the area of
question answering
(QA), queries are expected to be relatively
coherent questions already, such as “What is the height of Mount Everest?” A
large-scale search engine would largely, if not completely, ignore the valuable
prepositions and articles that give away the type (here,
height
) of the desired
answer. In the first part of this article (Section 10.2), we will present a
technique to extract the answer type (also called
atype
for short) from a
well-formed question. The atypes are provided to the system as a directed
acyclic graph (DAG) of types, edges representing transitive “is-a” relations,
e.g., Einstein is-a physicist is-a scientist.
Search WWH ::
Custom Search