Meeting Recorder Project: Dialog Act Labeling Guide

Abstract

Dialog act annotation potentially provides a means to aid information retrieval and summarization of meeting data. This work presents an in-depth view of the annotation methods of both the dialog act annotation and adjacency pair labeling schemes used for the Meeting Recorder data. Additionally, detailed descriptions of the individual tags within the Meeting Recorder Dialog Act tagset are provided. Issues such as utterance segmentation as well as numerous examples from the meeting data are found within this work.

The ICSI Meeting Recorder Dialog Act (MRDA) Corpus

“…Each type of annotation is described in detail in [7]. Segmentation methods were developed based on separating out speech regions having different discourse functions, but also paying attention to pauses and intonational grouping.…”

“…In addition, meetings that occur naturally involve real topics, debates, issues, and social dynamics that should generalize more readily to other real meetings than might data collected using artificial scenarios. Thus meetings pose interesting challenges to descriptive and theoretical models of discourse, as well as to researchers in the speech recognition community [4,7,9,13,14,15].…”

Meeting Modelling in the Context of Multimodal Research

“…As background knowledge for this box, the meeting agenda gives an indication of several tasks and topics that will appear in the meeting. A concrete choice for an annotation scheme for CAs is based on the functional aspect and might be based on an adaptation of some dialogue act annotation scheme such as MRDA [21], SWBD-DAMSL [22] or the IPA of Bales [23].…”

A Comparison of Language Models for Dialog Act Segmentation of Meeting Transcripts

“…Transcribers were instructed to focus on transcribing words as quickly as possible; there was not a focus on consistency or conventions for marking punctuation. Hence, instead of using the inconsistent first-pass punctuation, it was decided to employ special DA segmentation marks from the MRDA annotation project [4]. In this annotation pass, labelers carefully annotated both dialog acts and their boundaries, using using a set of segmentation conventions for the latter.…”