Mining Text Conversations - Methods for Mining and Summarizing Text Conversations - page 68

3.4.3 DECISION DETECTION
Related to the problem of dialogue act classification is the task of decision detection. With decision
detection, the goal is to identify sentences in a conversation that relate to a decision-making process .
What separates this task from dialogue act classification is that sentences or utterances may be

decision
process

relevant to a decision without being performative decision dialogue acts. Consider the following two
sentences:

1. We need to decide on the type of chip.

2. Okay, let's go with the simple chip.

Both sentences are relevant to a decision-making process, but only Sentence 2 is a performative
Decision dialogue act in that the utterance “performs the act” of making a decision. The dialogue act
type for Sentence 1 would likely be Statement or Inform , but we still want to capture the fact that it
is relevant to a decision process.
Decision sentences are more common in meeting and email conversations than in blogs and
discussion fora, since the latter venues tend to be informal and less goal-oriented. Decision sentences
are particularly frequent in meetings, since meetings tend to feature a cohesive group of participants
working on some type of joint venture. This is not necessarily the case even with email conversations.
Another reason that meetings have been the focus of much decision detection research is the
presence of relevant annotations for the AMI and ICSI corpora. As described in Chapter 2 , the
abstract summaries for these meetings contain subsections for describing decisions made during the
meeting. Since there are also abstract-extract links, we can determine which utterances in the meeting
relate to decision processes, thereby providing labeled data for supervised decision detection. Figure
3.12 shows example links between utterances from an AMI meeting and the decision subsection of
the gold-standard abstract.
An example of such a supervised approach is the work of Hsueh and Moore [ 2007 ]
and Hsueh et al. [ 2007 ]. The authors train maximum entropy classifiers on this labeled AMI data,
using lexical, prosodic, topical and contextual features. They also predict decision segments at two
levels of granularity: utterance level and topic segment level. They found that this full set of features
yielded the best precision but suffered from lower recall than a simple baseline using only unigram
features.
Fernandez et al. [ 2008 ] also worked on decision detection in the AMI corpus, but at a finer-
grained level. Rather than just classifying utterances as relevant to a decision process, they created a
hierarchy of decision types. The main three types are issue , resolution and agreement . Issue utterances
are those that introduce or describe the topic to be decided on. Resolution utterances are those

issue

that contain the final adopted decision. There are two subclasses of resolution utterances; proposal

resolution

utterances propose the adopted decision, while restatement utterances confirm or restate the adopted
decision. Agreement utterances are those that signal agreement with the adopted decision. agreement

Again considering Figure 3.12 , we could consider S1 to be an issue sentence since it introduces
the topic to be decided on (whether they should continue with production), S2 to be a resolution