Summarizing Text Conversations - Methods for Mining and Summarizing Text Conversations - page 100

rather than raw data such as numeric measurements (in their case, turbine readings, weather stations,
and intensive-care unit monitors). In either case, one is looking for messages (or patterns) in the
input, selecting the most critical messages, structuring and combining them in a coherent fashion,
and finally generating text to describe the body of selected messages.
To give a concrete example of how an abstractive conversation summarizer such as that
of Murray et al. [ 2010 ] compares with an extractive system such as Murray and Carenini [ 2008 ],
we can again consider the sample conversation presented at the beginning of this chapter. For the
purpose of this example, let us assume that a three-sentence summary of the conversation is required.
Based on features such as sentence position, sentence length, and term-weights, the following three
sentences could be selected by the extractive system:
￿ From Erica : The purpose of your conversation with Mr. Murdock is to discuss the above in
more detail and to more fully brief you on the purpose of these dinners.
￿ From Joannie : Would it be possible to schedule during the next quarter?
￿ From Erica : Given the time difference involved I will not be able to contact him before the
appointed time tomorrow, and therefore I'd very much appreciate if we could go ahead with
the call as planned.
In contrast, the abstractive system would first populate the ontology with participant instances
( Joannie, Erica, Sherri) and entities (e.g., phone call, office, vacation). Sentences would then be
analyzed using dialogue act classifiers. For example, the following sentence is determined to be a
subjective sentence:
Joannie : Erica, Due to the fact that Jeff is unable to attend on July 19, I believe it would be
better to reschedule the call for sometime next quarter.
The abstractor would then look for patterns of similar sentences and combine similar sentences
into a message. For example, sentences where Joannie is making negative-subjective comments about
the phone call could be combined into a single message with Joannie as the message source and phone
call as the message target. A subset of messages would be selected for the final summary.The summary
text could look something like the following.
Erica mentioned an action item concerning the phone call. Joannie then expressed some negative
opinions about the scheduling. Finally, Sherri asked a question about the upcoming dinners.
In both summary cases, the summary sentences can be linked to the original document sen-
tences. This allows the reader a better understanding of the context and the surrounding sentences.
Moreover, the summary essentially serves as a gateway for the reader to more systematically browse
through the original sentences. For instance, the reader may be very interested in browsing through
all the sentences expressing negative opinions, sentences that represent action items, or sentences
that describe decisions made. While the underlying conversational data are unstructured, the sum-