Summarizing Text Conversations - Methods for Mining and Summarizing Text Conversations - page 98

work also focuses on detecting disfluencies such as filled pauses, false starts and repairs in order to
increase summary readability and informativeness.
Murray and Carenini [ 2008 ] have developed extractive summarization techniques for spoken
and written conversations, with a focus on meetings and email. The primary goal in this research
is to identify a common feature set that would yield good summarization performance in both
spoken and written conversations. The features included speaker/participant dominance, lexical
cohesion, participant-based term weights, centroid similarity scores, turn-taking and other structural
characteristics of multi-party conversation. The participant-based term weights are based on the
intuition that certain words will tend to associate more with some participants than others, owing to
varying roles (e.g., industrial designer vs. financial expert) and generally different areas of interest and
expertise. When rating the features according to the F statistic (basically the ability of an individual
feature to discriminate the positive and negative classes) they find that the feature rankings are very
similar for the two domains. This is shown in Figures 4.4 and 4.5 (source: [ Murray and Carenini ,
2008 ]). We do not discuss the individual features in detail here, but merely note that the most
important feature subsets are very similar in the two domains.

0.25 0.09

manual
ASR

0.08

0.2

0.07

0.06

0.15

0.05

0.04

0.1

0.03

0.02

0.05

0.01

0 0

feature ID (see key) feature ID (see key)

Figure 4.4: F statistics, meetings. Figure 4.5: F statistics, emails.

The highest AUROC scores for the extractive classifiers are approximately 0.85 for meetings
and 0.75 for emails. Most significantly, they compare the conversation-features approach to a speech-
specific system for meetings and an email-specific system for emails and find that the conversation-
features approach performs just as well in each domain. That is, there was no bonus to using domain-
specific features and one could instead rely on general conversation features. This finding is similar to
those of Penn and Zhu [ 2008 ], who find that “avant-garde,” domain-specific features often provide
little or no performance improvement over more general features (see Section 4.3.2 ).
Sandu et al. [ 2010 ] also use general conversation features and try to leverage the large amount
of available labeled meeting data to improve summarization results in the less-resourced domain of
emails. This is the general problem of domain adaptation , where one tries to adapt a system developed domain

adaptation

in a “source” domain to data in a “target” domain. The authors found that domain adaptation
techniques were helpful when no labeled email data was available.