SECTION 4 HIGHLIGHTS THE MANAGEMENT OF THE INTER-FORMANCE OF A SYSTEM...
9
.
E1: "in the early 1970s"; Category: TIME
Representing topics in terms of relevant concepts
E2: "Egyptian President Anwar Sadat"; Category: PERSON
E3: "Egypt"; Category: COUNTRY
and relations is important for the processing of ques-
E4: "BW stockpile"; Category: UNKNOWN
2 predicates: P1="validate"; P2="has"
tions asked within the context of a given topic. For
Predicate−Argument Structures
P1: validate
interactive Q/A, however, the ideal topic-structured
Reference 1 (definitional)
Egyptian President X
arguments: A0 = E2: Answer Type: Definition
Reference 2 (metonymic)
PROCESSING
A1 = P2: have
representation would be in the form of question-
arguments: A0 = E3
Reference 3 (part−whole)
A1 = E4
E5: BW program
answer pairs (QUABs) that model the individual
ArgM−TMP: E1: Answer Type: Time
segments of the scenario. We have currently cre-
Reference 4 (relational)
P3: admit
ated two sets of QUABs: a handcrafted set and
QUESTIONS
Definition Pattern: Who is X?
an automatically-generated set. For the manually-
Q1: Who is Anwar Sadat?
Pattern: When did E3 P1 to P2 E4?
created set of QUABs, 4 linguists manually gener-
Q2: When did Egypt validate to having BW stockpiles?
Pattern: When did E3 P3 to P2 E4?
ated 3210 question-answer pairs for each of the 8
Q3: When did Egypt admit to having BW stockpiles?
Pattern: When did E3 P3 to P2 E5?
dialogue scenarios considered in our experiments.
Q4: When did Egypt admint to having a BW program?
In a separate effort, we devised a process for au-
Figure 4: Associating Questions with Answers.
tomatically populating the QUAB for each scenario.
R
Question Generation: In order to automati-
In order to generate question-answer pairs for each
cally generate questions from answer passages, we
subtopic, we first identified relevant text passages in
considered the following two problems:
the document collection to serve as “answers” and
Problem 1: Every word in an answer passage
then generated individual questions that could be an-
can refer to an entity, a relation, or an event. In
order for question generation be successful, we
4
Initially,P Q
contains only the seed relation. Additionalmust determine whether a particular reference
relations can be added with each iteration.is “interesting” enough to the scenario such that
termine the selection of the answer. The predi-
it deserves to be mentioned in a topic-relevant
cate
!
can be substituted by its nominalization
question. For example, Figure 4 illustrates an
of
8
is BW, the same argument is
since
O
<
8
transferred to
answer that includes two predicates and four
. The causality implied by the
entities. In this case, four types of reference are
answer from Figure 5 has two components: (1)
used to associate these linguistic objects with
the effect (i.e. the predicate
8
) and (2) the re-
sult, which eliminates the semantic effect of the
other related objects: (a) definitional reference,
used to link entity (E1) “Anwar Sadat” to a cor-
negative polarity item never by implying the
responding attribute “Egyptian President”, (b)
predicate
!
, obstacle. The questions that are
metonymic reference, since (E1) can be coerced
generated are based on question patterns asso-
into (E2), (c) part-whole reference, since “BW
ciated with causal relations and therefore allow
stockpiles”(E4) necessarily imply the existence
different degrees for the specificity of the resul-
of a “BW program”(E5), and (d) relational ref-
tative, i.e obstacle or deterrent.
erence, since validating is subsumed as part
We generated several questions for each answer
of the meaning of declaring (as determined by
passage. Questions were generated based on pat-
WordNet glosses), while admitting can be de-
terns that were acquired to model interrogations
fined in terms of declaring, as in declaring [to
using relations between predicates and their argu-
be true].
ments. Such interrogations are based on (1) as-
ANSWER
sociations between the answer type (e.g. DATE)
Egyptian Deputy Minister Mahmud Salim states that Egypt’s
enemies would never use BW because they are aware that the
and the question stem (e.g. “when” and (2) the
Egyptians have "adequate means of retaliating without delay".
Predicates: P’1=state; P’2 = never use; P3 = be aware;
relation between predicates, question stem and the
P’4 = have P"4 = "the possesion"
Causality:
words that determine the answer type (Narayanan
P"4 = "the possesion" = nominalization(P’4) = EFFECT(P’2(BW))
and Harabagiu, 2004). In order to obtain these
P’2(BW) = NON−NEGATIVE RESULT(P5); P’5 = "obstacle"
PROCESSING
specialization
Reference: P’1 P’6 = view
predicate-argument patterns, we used 30% (approxi-
mately 1500 questions) of the handcrafted question-
Pattern: Does Egypt P’6 P"4(BW) as a P’5?
Does Egypt view the possesion of BW as an obstacle?
answer pairs, selected at random from each of the 8
Does Egypt view the possesion of BW as a deterrent?
dialogue scenarios. As Figures 4 and 5 illustrate, we
Figure 5: Questions for Implied Causal Relations.
used patterns based on (a) embedded predicates and
Problem 2: We have found that the identifica-
(b) causal or counterfactual predicates.
tion of the association between a candidate an-
swer and a question depends on (a) the recogni-
4 Managing Interactive Q/A Dialogues
tion of predicates and entities based on both the
output of a named entity recognizer and a se-
As illustrated in Figure 1, the main idea of man-
mantic parser (Surdeanu et al., 2003) and their
aging dialogues in which interactions with the Q/A
system occur is based on the notion of predictions,
structuring into predicate-argument frames, (b)
i.e. by proposing to the user a small set of questions
the resolution of reference (addressed in Prob-
that tackle the same subject as her question (as illus-
lem 1), (c) the recognition of implicit rela-
trated in Table 1). The advantage is that the user can
tions between predications stated in the answer.
follow-up with one of the pre-processed questions,
Some of these implicit relations are referential,
8
8
<
as is the relation between predicates
and
that has a correct answer and resides in one of the
QUABs. This enhances the effectiveness of the dia-
illustrated in Figure 4. A special case of im-
logue. It also may impact on the efficiency, i.e. the
plicit relations are the causal relations. Fig-
ure 5 illustrates an answer where a causal re-
number of questions being asked if the QUABs have
lation exists and is marked by the cue phrase
good coverage of the subject areas of the scenario.
because. Predicates – like those in Figure 5 –
Moreover, complex questions, that generally are not
) or negative (like
can be phrasal (like
processed with high accuracy by current state-of-
).
Causality is established between predicates
8the-art Q/A systems, are associated with predictive
and
8
’ as they are the ones that ultimately de-
questions that represent decompositions based on
similarities between predicates and arguments of the
as
,)
:
7
@
B
7
A
, where
?6
:>/
?
6
CB
original question and the predicted questions.
=
D
(,7HG
<
:E
:>FN
%
E
The selection of the questions from the QUABs
E
6
F
BIKJLNM<OQP
MR