2 INTERPRETING NUMERICAL ANSWERSSWER QUALIFIES AS A POSITIVE OR NEGA...
4.2 Interpreting numerical answers
swer qualifies as a positive or negative instance of
the adjective in the question, and thus interpret the
The second technique aims at determining
indirect response as a ‘yes’ or a ‘no’. The prob-
whether a numerical answer counts as a positive
abilistic nature of this technique adheres perfectly
or negative instance of the adjective in the ques-
to the fact that indirect answers are intimately tied
tion (category II in our corpus).
up with uncertainty.
Adjectives that can receive a conventional unit
of measure, such as little or long, inherently pos-
5 Evaluation and results
sess a degree of vagueness (Kamp and Partee,
1995; Kennedy, 2007): while in the extreme cases,
Our primary goal is to evaluate how well we can
learn the relevant scalar and entailment relation-
judgments are strong (e.g., a six foot tall woman
can clearly be called “a tall woman” whereas a
ships from the Web. In the evaluation, we thus ap-
five foot tall woman cannot), there are borderline
plied our techniques to a manually coded corpus
cases for which it is difficult to say whether the
version. For the adjectival scales, we annotated
adjectival predication can truthfully be ascribed
each example for its main predication (modifier, or
adverb–modifier bigram), including whether that
to them. A logistic regression model can capture
predication was negated. For the numerical cases,
these facts. To build this model, we gather distri-
we manually constructed the initial queries: we
butional information from the Web.
For instance, in the case of (3), we can retrieve
identified the adjective and the modified entity in
the question, and the unit of measure in the answer.
from the Web positive and negative examples of
age in relation to the adjective and the modified en-
However, we believe that identifying the requisite
tity “little kids”. The question contains the adjec-
predications and recognizing the presence of nega-
tion or embedding could be done automatically us-
tive and the modified entity. The reply contains the
ing dependency graphs.
4
unit of measure (here “year-old”) and the numer-
ical answer. Specifically we query the Web using
3
Otherwise, our model is ruined by references to “young