IS USED TO SELECT AND PROMOTE EFFECTIVE QUESTION AND ANSWER CAND...
1996) is used to select and promote effective
question and answer candidate. This problem is
transformation rules. We rely on recent work in
particularly acute in the case of story
attribute-efficient relational learning (Khardon et
comprehension due to the rarity of information
al., 1999; Cumby and Roth, 2000; Even-Zohar and
restatement in the single document.
Roth, 2000) to acquire natural representations of
Several recent systems have specifically
the underlying domain features. These
addressed the task of story comprehension. The
During the performance phase, only the narrative
and question are given.
apply
reinforcement to
At the lexical level, an answer to a question is
rule base
generated by applying a series of transformation
rules to the text of the narrative. These
transformation rules augment the original text with
return FAIL
no
one or more additional sentences, such that one of
these explicitly contains the answer, and matches
more
the form of the question.
yes
lookup existing
valid rule
processing
primitive
applicable rule
exists?
On the abstract level, this is essentially a
ops?
time?
yes
yes
process of searching for a path through problem
acting by
inference
search
space that transforms the world state, as described
instantiate
new rule
goal state
by the textual source and question, into a world
generalize against
reached?
state containing an appropriate answer. This
START
process is made efficient by learning answer-
extract current
generation strategies. These strategies store
execute rule in
ABSTRACT
state features &
LEARNING/
domain
procedural knowledge regarding the way in which
compare to goal
REASONING
FRAMEWORK
answers are derived from text, and suggest
appropriate transformation rules at each step in the
match candidate
INTERMEDIATE
lexically pre-
modify raw text
answer-generation process. Strategies (and the
sentence
PROCESSING
process raw text
extract answer
LAYER
procedural knowledge stored therein) are acquired
by explaining (or deducing) correct answers from
RAW
TEXTUAL
lexicalized answer
training examples. The framework’s ability to
raw text, question, (answer)
DOMAIN