Conclusions/Final Thoughts - Methods for Mining and Summarizing Text Conversations - page 105

Features Used

Conversational

Semantic

Domain Adaptation

Machine Learning
Techniques

Supervised

Unsupervised

Semi-Supervised

Syntactic

Simple Lexical
(bag of words)

Figure 5.2: The most promising computational approaches.

To review our overall discussion of the material presented in this topic, in Chapter 2 we
described some of the conversation corpora that are available and widely used by researchers. The
fact that there are so many freely available corpora makes this an excellent time to be researching
conversation-related topics, and bodes well for future research. There is also growing agreement in
this research community on the evaluation metrics and annotation standards used. For example, both
the AMI and BC3 corpora contain abstractive and extractive summary annotations that are linked
to one another. This mapping between extracts and abstracts can help researchers build systems
that are increasingly abstractive rather than solely cut-and-paste. And having corpora in different
domains that are annotated similarly to one another can aid work on domain adaptation.
In Chapter 3 , we considered several text mining tasks. We started by covering tasks that can
be performed on any document such as determining what topics are covered in the conversation (i.e.,
topic modeling), as well as detecting what opinions are expressed on those topics (i.e., sentiment and
subjectivity analysis).Then, we focused on tasks that consider unique features of human conversation.
In particular, we have discussed how to determine what dialog acts are expressed in the different
turns, what is the thread structure of a conversation, and what turns are expressing decisions and
action items. For all these tasks, we described and compared both supervised and unsupervised