Click here for Other situated corpora collected in the SLaTe lab

OSU Quake 2004 Dialog Corpus:
A publicly available corpus of collaborative dialog in a virtual world

This small corpus is intended for researchers who want to study collaboration between humans working on a situated task. In this case, two human partners perform a treasure-hunt task in a graphically-rendered virtual world that is portrayed on their computer monitors. The partners spoke to each other in order to coordinate their activity on the task. The problem domain was chosen to simulate the search-and-rescue domain.

Here's a little sample of the interaction in this domain.

The corpus has the following properties:

Two-party dialog about solving a search-and-rescue style task.
The partners could move about in the world independently, so their knowledge of the world diverges.
The partners could see each other and also see objects and events in the world, so they frequently make exophoric reference to items in the world.
The partners could both perform the same actions in the world, so the initiative for completing the task is left open.
Dialog behaviors such as turn-taking and disfluencies happen naturally.

The technical report OSU-CISRC-8/05-TR57 describes the data collection conditions, subject instructions, etc. If you publish results that include the use of this corpus, please cite:

Donna K. Byron and Eric Fosler-Lussier. "The OSU Quake 2004 corpus of two-party situated problem-solving dialogs." In Proceedings of the 15th Language Resources and Evaluation Conference (LREC'06), 2006.
» PDF, BibTeX

The corpus

To date we have transcribed and annotated 5 problem-solving sessions. The data is available to any researchers who want to use it it subject to the terms of use.

Links to tools we used

If you would like to perform a similar experiment, here are some of the tools you will need:

Quake II software. The game engine client/server code and the rendering software are opensource, and they can be downloaded from idsoftware.com. You also need the game pak files on the QuakeII original game CD, which can be purchased on the used market. We used QuakeII because it is opensource, but you could also use NeverWinter Nights, Halo, HalfLife, or Unreal Tournament. These are all game engines that are used by AI researchers to study embodied interaction.
QE Radiant or Quark are the level editing tools we used to build the little quake world. We used wally to make the wall textures.
Transcriber from ldc can be used to align the transcript with the audio stream.
Soundscriber is a tool for transcription that automatically loops each section of audio a number of times. It makes transcribing easier. This distribution only runs on Windows.
Annotation tools: mmax
Audio Alignment: sonic
Our audio recording equipment setup is described in the TR.

Local Resources

Guidelines for transcribing the audio

SLaTe: Speech and LAnguage TEchnologies Lab

OSU Quake 2004 Dialog Corpus: A publicly available corpus of collaborative dialog in a virtual world

The corpus

Links to tools we used

Local Resources

SLaTe: Speech and LAnguage TEchnologies Lab

Dept. of Computer Science & Engineering 580 Dreese Labs The Ohio State University

OSU Quake 2004 Dialog Corpus:
A publicly available corpus of collaborative dialog in a virtual world

Dept. of Computer Science & Engineering
580 Dreese Labs
The Ohio State University