Database Reference
In-Depth Information
What kind of an animal is Winnie the Pooh
start
start
0
1
0
1
2
What, kind,
of, an, is,
Winnie, the,
Pooh
What, kind,
of, an
is, Winnie,
the, Pooh
animal
animal
FIGURE 10.4 : 2- and 3-state transition models.
completeness. Consider these question pairs:
1a What country is the largest producer of wheat?
1b Who is the largest producer of wheat?
2a Which president was the winner of the tenth election?
2b Name the winner of the tenth election
In 1b and 2b, for want of better informers, we would want producer and winner
to be flagged as informers, but in 1a and 2a, country and president would be
more useful informers.
The i
1context
of winner in 2a and 2b. Any 2-state model that depends on positions i
±
1 context of producer is identical in 1a and 1b, as is the i
±
±
1
to define features will fail to distinguish between 1a and 1b, or 2a and 2b,
and might mark both country and producer in 1a, and president and winner .
From Figure 10.3 , we see that generating features from parts of the question
that are not informer tokens can reduce accuracy. Therefore, we would like to
identify the single most likely informer span. By design, the BIO model will
never annotate more than one contiguous segment.
The tree kernels used by Zhang et al. exploited the property that questions
with similar parse trees are likely to have the informer span over similar token
positions. We will therefore use the parse tree of the question to generate
features that will be used in a conditional HMM. Unlike Zhang et al. , we will
pay attention only to selected parts of the parse tree.
10.2.2.1
Parse tree and multiresolution feature table
Figure 10.5 shows a sample parse tree output by the Stanford Lexicalized
Parser (23). The tree has been organized in levels. The non-terminal symbols
of the tree follow the Penn Treebank tag convention (see http://www.cis.
upenn.edu/ ~ treebank/ for details).
To employ max-margin methods for labeling (40), a suitable kernel has to
be defined between two such trees. In case of a general CRF, we would be
 
Search WWH ::




Custom Search