Conclusions/Final Thoughts - Methods for Mining and Summarizing Text Conversations - page 106

Figure 5.3: Integration of different text mining tasks.
machine learning methods. We also noticed that often synchronous vs. asynchronous conversations
present different challenges. For example, synchronous conversations, especially written ones like
chats, often require disentanglement to determine reply-to relationships between turns. In contrast,
an interesting problem for asynchronous conversation with quotation is to extract a finer level
conversational structure. Another issue we discussed is how each mining task can rely on the others;
for example, work on action item detection can use predicted dialogue act labels as an input feature
to a statistical classifier. Current work is exploring other beneficial dependencies among tasks; for
instance, between both topic modeling and conversational structure.
In Chapter 4 , we gave an overview of many existing systems for summarizing meetings, emails,
blogs and forums. We examine these systems in the light of three considerations: assumptions and
inputs , measures of informativeness , and outputs and interfaces . We hope that having the overview
structured in those terms will help researchers who are considering building a summarization system
in a particular domain and are uncertain of the options and requirements. We also highlight systems
that have been designed to work on conversations in multiple modalities and genres. Finally, we
discuss early work on abstractive systems that attempt to glean a deeper understanding of the
conversation and generate new text to describe it. We use a detailed case study to show a relatively
simple abstractive system can be built and how it compares with a standard extractive system.
As stated in the introduction, conversations are fundamental to the human experience, and we
live in a technological age where conversations are more prevalent than ever.They span modalities and
can grow to include hundreds or even thousands of people. And more often than not, conversations
now exist in a lasting record that can be analyzed, mined, condensed and visualized. We hope that this
topic provides a glimpse of the many possibilities for summarizing and mining such conversations
and gives inspiration for insightful new approaches.