Background: Corpora and Evaluation Methods - Methods for Mining and Summarizing Text Conversations - page 24

the action performed by the utterance. For example, the utterance Is that right? has the dialogue act
type Backchannel-Question because it is a question and it is also steering the conversation back to the
previous speaker. The most frequent type of dialogue act in meeting corpora is Statement , where a
speaker is simply providing information. So utterances can be classified as dialogue acts, where each
is assigned its most likely dialogue act type .
In the discussion that follows, we often use utterance and dialogue act interchangeably, where
we are referring to a sentence-like unit in speech. In other cases we will refer to specific dialogue act
types such as Yes-No-Question , Agreement or Statement .
AMI and ICSI Meeting Annotation Both the AMI and ICSI corpora were initially annotated for
topic segmentation (see Section 1.4 )and dialogue acts. Note that the MRDA (Meeting Recording
Dialogue Act) corpus refers to dialogue act annotations of the ICSI meetings. These annotations
are described in detail by Shriberg et al. [ 2004 ]. The two meeting corpora were then annotated for
summarization. More specifically, for the summarization annotation, annotators were asked to write
abstractive summaries of each meeting and to extract the meeting dialogue acts that best convey
or support the information in the abstractive summary. As described in Chapter 1 , an abstractive
summary is a high-level summary using novel text to synthesize and describe information from the
document.
Annotators used a graphical user interface (GUI) to browse each meeting, enabling them to
view previous human annotations consisting of a written transcription synchronized to the meeting
audio, and topic segmentation. The annotators were first asked to build a textual summary of the
meeting aimed at an interested third-party, using four sub-headings for their abstract. For the ICSI
meetings, those four headings are:
￿ general abstract: “why are they meeting and what do they talk about?”;
￿ decisions made by the group;
￿ progress and achievements;
￿ problems described
For the AMI meetings, the summary sections vary slightly:
￿ general abstract;
￿ decisions;
￿ actions;
￿ problems.
The maximum length for each section of the abstract is 200 words, and while it was mandatory
that each general abstract section contained text, it was permitted that for some meetings the other