Information Technology Reference
In-Depth Information
•
Features corresponding to the merit of the interval.
Third, with the above feature representation, standard Ranking SVM [
7
,
8
]is
used to combine these features and produce a ranked list of all the candidate inter-
vals.
The above method has been tested on TREC-QA data, and the experimental re-
sults show significant improvement over several non-learning baselines in terms
of both NDCG and MAP. Indirect comparison with the TREC participants also
suggests that the proposed method is very competitive: ranked the second-best in
TREC-QA 2007 data, and ranked five out of 63 teams on TREC-QA 2004 data.
14.2.3 Non-factoid QA
In [
14
], an answer ranking engine for non-factoid questions built using a large online
community-generated question-answer collection (Yahoo! Answers) is proposed.
Through the construction of the engine, two issues are investigated. (i) Is it possible
to learn an answer ranking model for complex questions from noisy data? (ii) Which
features are most useful in this scenario?
In the proposed engine, there are three key components, the
answer retrieval
component
uses unsupervised information retrieval models, the
answer ranking
component
uses learning-to-rank technologies, and the
question-to-answer trans-
lation model
uses class-conditional learning techniques. Here we mainly introduce
the learning-to-rank technology used in the answer ranking component.
First of all, given a set of answers, one needs to extract features to represent their
relevance to the question. In [
14
], the following features are extracted.
•
Similarity features:
BM25 and TF-IDF on five different representations of ques-
tions and answers: words, dependencies, generalized dependencies, bigrams, and
generalized bigrams.
•
Translation features:
the probability that the question
Q
is a translation of the an-
swer
A
, computed using IBM's model 1 [
2
], also on five different representations.
•
Density and frequency features:
same word sequence, answer span, overall
matches, and informativeness.
•
Web correlation features:
web correlation and query log correlation.
With the above features, the learning-to-rank method proposed in [
13
], called
Ranking Perceptron, is employed to learn the answer ranking function. The basic
idea of Ranking Perceptron is as follows. Given a weight vector
w
the score for
a candidate answer
x
is simply the inner product between
x
and
w
, i.e.,
f(x)
=
w
T
x
. In training, for each pair
(x
u
,x
v
)
, the score
f(x
u
−
x
v
)
is computed. Given a
margin function
g(u,v)
and a positive rate
τ
,if
f(x
u
−
x
v
)
≤
g(u,v)τ
, an update
is performed:
w
t
+
1
w
t
=
+
(x
u
−
x
v
)g(u, v)τ,
(14.1)
where
g(u,v)
=
(
u
−
1
v
) and
τ
is found empirically using a validation set.
