Mining Text Conversations - Methods for Mining and Summarizing Text Conversations - page 64

deliver

deliver

request

request

E1

E1 E1

request
commit

commit

E2 E3

E2 E3 E2 E3

commit

commit

E4 E5

E4 E5 E4 E5

request
deliver

deliver

request
deliver

deliver

(a)

(b)

(c)

Figure 3.9: Key operations of the collective classification algorithm, assuming a restricted tagset with
only three tags, request , commit and deliver . (a) Email thread with five emails initiated by email E1. (b)
Step1 - labels for each email message are initialized by applying to each message the three classifiers
Content request , Content commit and Content deliver . (c) As part of one iteration in Step2, the labels of
E2 are revised by applying the three classifiers Context request , Context commit and Context deliver . The
contextual information taken into account by the three classifiers is grayed in the Figure.

￿ Question marks are not always intended to signal a question. Due to the informality of email,
and because people strive for conciseness, sometimes question marks are used to express un-
certainty.

￿ Some questions can be phrased declaratively, e.g., “I am wondering if......”.

￿ Some questions are rhetorical in nature; they are not intended to be answered.

Based on these observations, a more sophisticated approach to question detection is proposed
(but not thoroughly tested) by Shrestha and McKeown. The task is framed as a binary classification
problem, where each utterance is classified as being or not-being a question. In their very preliminary
experiments, they only consider a few sentence features, including the length of the sentence and
simple lexical features (i.e., POS tags of the constituent words and bigrams). In future work, this set
could be expanded with additional lexical and syntactic features.
Once all the questions have been detected, one needs to identify the corresponding answers
for each question. This task can also be rather challenging in email. Due to the asynchronous nature
of email and the use of quotation, an answer to a question may not appear in the email that directly
replies to the email in which the question was originally posed. For instance, in the sample email
thread in Figure 3.10 , the answer A 21 to question Q 2 appears in Email-2, a direct reply to the initial