Text Search-Enhanced with Types and Entities - Text Mining: Classification, Clustering, and Applications - page 245

What is the capital city of Japan 0

WP VBZ DT NN NN IN NNP

1

WHNP VP NP NP

2

3

PP

4

NP

5

SQ

SBARQ 6

FIGURE 10.5 : Stanford Parser output example.

i 1 2 3 4 5 6 7

y i 0 0 0 1 1 2 2

x i

What is the capital city of Japan

Features for x i s
1 WP,1 VBZ,1 DT,1 NN,1 NN,1 IN,1 NNP,1
2 WHNP,1 VP,1 NP,1 NP ,1 NP,1 Null,1 NP,2
3 Null,1 Null,1 Null,1 Null,1 Null,1 PP,1 PP,1
4 Null,1 Null,1 NP,1 NP,1 NP,1 NP,1 NP,1
5 Null,1 SQ,1 SQ,1 SQ,1 SQ,1 SQ,1 SQ,1
6 SBARQ SBARQSBARQSBARQSBARQSBARQSBARQ

↓

FIGURE 10.6 : A multi-resolution tabular view of the question parse
showing tag and num attributes in each cell. capital city is the informer

span with y =1.

interested in associating a binary label (informer token or not) with the leaf
nodes of the parse tree. It is tempting to cast this as a structured prediction
problem where the graph is the parse tree itself, and the edges of the parse
tree express Markov dependencies. A straightforward implementation of this
approach may limit the exploitation of long-range features observable from
the parse tree. For example, we may observe from training data that informer
spans occur more often near the beginning of the question, but are rarely the
first word. In other words, there is a distribution over token positions where
the informer span occurs. Also consider the information latent in the parse
tree at multiple levels of resolution. For example, in Figure 10.5, to label city
as part of the informer span, it may help us more to know that the next chunk
of Japan has POS “PP” at level 3, than to know that the next token of has
POS “IN” at level 1.
To exploit these valuable clues, we will first express the question parse tree