Mining Text Conversations - Methods for Mining and Summarizing Text Conversations - page 67

Figure 3.11: Graphical representation of the HMM-like model learned from Twitter data in an unsu-
pervised way. The states were labeled by a human annotator. Transition between states are shown only if
their probability is greater than 0.15 [source [ Ritter et al. , 2010 ]].
by a Reaction (p=0.16), a Comment (p=0.18), or a Question (p=0.34), but typically not by an Answer
(p < 0.15)(which is why it is not shown in the graph).
Current and Future Trends in Dialogue Act Modeling for Text Conversations Current research in
dialogue act modeling is pushing for semi-supervised and unsupervised approaches and for more
sophisticated sentence features. As we already mentioned, another open area for fruitful research
is how to integrate the various mining tasks covered in this chapter. For example, an interesting
open question, already partially explored in Ritter et al. [ 2010 ], is how dialogue act modeling could
benefit from topic modeling and vice-versa; or, alternatively, how dialogue act modeling could benefit
from extracting a finer grain conversational structure (see Section 3.4.5 ), as very recently explored
in Joty et al. [ 2011 ].