Summarizing Text Conversations - Methods for Mining and Summarizing Text Conversations - page 81

summary for Mr. Skilling might detail the time and purpose of the phone call, whereas the summary
for Mr. Murdock might simply confirm that Skilling will participate. This demonstrates there is
no such thing as a single best summary for a given conversation. There are many other reasons
why multiple summaries of a single conversation might differ, such as ambiguity inherent in the
conversation. If we ask two people to create summaries of this email exchange, they may disagree on
whether a decision was actually made regarding the phone call. Furthermore, summaries can vary
according to explicitly provided information needs. A summary might not be generated generically
but rather in response to a user query, in which case a good summary could be focused on a particular
date, time or name.
Summaries of this conversation might also vary in granularity. We could generate a concise,
decision-based summary that reads, It was decided that Mr. Skilling and Mr. Murdock will speak on
the phone tomorrow , or we could generate a summary that describes the decision process , in this case
emphasizing that Erica and Joannie disagreed about the need for the phone call. The decision process
in this example might seem trivial and unnecessary to summarize, but in many real-world cases it is
the decision process that is critical to understand. In fact, automatic summarization has previously
been touted for its use in conducting corporate decision audits [ Murray et al. , 2009 ]. While an

decision
audit

important decision is likely to be well known and disseminated within an organization, the decision
process might quickly be forgotten. If it turns out that the decision was ill-advised, reconstructing
the decision process may be in the interest of the organization, in order to determine responsibility
and accountability.
One can imagine many ways to summarize this conversation, based on decisions, agendas,
action items, opinions or some combination thereof. Keep these possibilities in mind as we discuss
specific work in a variety of conversational domains.

4.2 SUMMARIZATION FRAMEWORK AND BACKGROUND

In discussing automatic summarization of conversations, we describe summarization approaches
and systems according to three aspects:

￿ Assumptions and Inputs. Assumptions can mean assumptions about the nature and format of

Assumptions
and Inputs

the data or assumptions about an end user's information needs, to give just two examples.
Inputs can mean upstream modules such as preprocessing and information extraction.

￿ Measuring Informativeness. This describes how a given approach or particular system deter-

Measuring
Informa-
tiveness

mines salience for a conversation, and constitutes the heart of the summarization pipeline.

￿ Outputs and Interfaces. Outputs can refer to the modality of the summary (e.g., textual vs.

Outputs
and
Interfaces

visual) and more specifically to the structure of the produced summary (e.g., extractive vs.
abstractive text). Interfaces can refer to the manner in which the summary is meant to be used
by an end user.