Author Archives: mmeteer

There are two corpuses of meetings that have both been annotated with dialog acts, ICSI and AMI.

Questions: How does multiparty dialog differ from the two-party dialog in SWBD? Is it any closer to the task oriented dialogs? What are the challenges of dialog act identification in multiparty dialogs? What other kinds of discourse research is being done on these corpuses?

Primary paper:

The ICSI MRDA corpus is described here:

and can be found here: http://www1.icsi.berkeley.edu/~ees/dadb/
There may be enough in the samples to get a sense of the annotation. The information in the whole corpus is distributed, so you need to combine the .dadb and .trans files (both are csv, it works in excel), and then get rid of all the columns you don’t need.

The AMI corpus is here: http://groups.inf.ed.ac.uk/ami/download/
You need the NXT toolkit: http://groups.inf.ed.ac.uk/nxt/

Work by Harry Bunt identified multiple dimensions to dialog and organized tag sets according to those dimensions.

• What are the dimensions?
• What aspects of dialog do they explain that the DAMSL set does not?
• How well do they work in practice (find someone that has used them)
• Bunt as led the ISO standard effort. How does this compare to his 2006 paper?
  
  465204 {:BC89X8K3} modern-language-association 50 1 101 http://www.cs232f.mmeteer.com/wp-content/plugins/zotpress/
• Who is using the ISO standard and how well is it working?

In this class, we’ll talk about “Dialog Acts”, starting with a paper that came out of a summer workshop at Johns Hopkins University in 1997, which became a standard in the industry. We’ll also read a paper of work that came out of this that considers multiple dimensions of dialog acts.

We’ll also start looking at data actually annotated with dialog acts, using the Penn Switchboard Treebank Discourse Relations data set. The Stolcke paper and the Switchboard data and annotation documentation are available on Latte.

There are a set of 5 unlabeled SWBD dialogs and the same 5 labeled. After reading the paper and the documentation, try to label some of the dialogs and then look at the official labeling. Be prepared to talk about what you find was hard or confusing.

For the next class:

Posts will contain the readings and other information for upcoming classes. The Zotero group (link on the right) contains papers relevant to the course. Those assigned for a class will be referenced in the post for that class.

In the first class, in addition to going over the content and format of the course, we’ll talk about evolution of the study of discourse structure in computational linguistics by looking at a paper written by Bonnie Webber and Aravind Joshi, both distinguished researchers in the field part of the ACL Workshop “Rediscovering 50 years of Discoveries”.