The Monroe Corpus

Home | People | Projects | Publications | Resources

General | Participants | Tasks | Setup | Dialogue stats | Processing | Transcripts and annotations

Introduction

This page describes the creation of the Monroe corpus at the University of Rochester. Sections are lifted in their entirety from this thesis.

A natural first step in the process of trying to model human dialogue behavior is the examination of appropriate human-human dialogues. We needed a corpus of human-human dialogues that were complex enough and long enough to give the speakers scope for demonstrating a variety of ways of managing initiative, participating in conversational games, etc. We also hoped to study dialogues in the same domain, and about the same type of tasks, as the version of TRIPS we were developing at the time. Existing corpora did not meet these needs, so we collected a new corpus of mixed-initiative human-human dialogues in a complex domain, the Monroe corpus. We here describe the collection, transcription and alignment of this data.

Related work

Publicly available spoken dialogue corpora include the ATIS collections [#!dahletal94!#,#!hemphilletal90!#,#!hirschmanetal93!#], the TRAINS corpus [#!heemanandallen95!#], the HCRC Maptask corpus [#!andersonetal91!#], the Switchboard corpus [#!godfreyetal90!#], the CallHome corpus [#!callhome!#], and the CallFriend corpus [#!callfriend!#]. The ATIS, TRAINS and Maptask corpora are microphone speech; Switchboard, CallHome and CallFriend are telephone speech.

The ATIS, TRAINS, and Maptask corpora are collections of task-oriented dialogues in relatively simple domains. The ATIS project involved the creation of a series of corpora about air traffic reservations. Users interacted with a simulated system in a Wizard of Oz study, or with working dialogue systems.

The TRAINS corpus, collected at the University of Rochester in 1993, consists of 98 dialogues involving 34 speakers. In each conversation, one participant was given a task to complete, and the other acting as assistant. The tasks involved transporting oranges to orange juice factories, making orange juice, and moving the orange juice. There were time constraints. Participants were not allowed to see each other.

The Maptask corpus is a collection of 128 dialogues involving 64 speakers. Each dialogue participant was given a map. For each dialogue, one participant's map had a route drawn on it and the other's did not. Furthermore, the maps were not identical. The task involved the drawing of the route on the map on which the route was not drawn to begin with. Half of the participants were allowed to see each other, and half were not.

The CallHome, Switchboard, and CallFriend corpora do not include task-oriented dialogues. In the Switchboard and CallFriend corpora, people were chosen at random and assigned a topic to discuss. The CallFriend and Switchboard corpora basically involve each speaker taking turns to talk on the topic, with the other one listening; the speakers do not have enough in common, either in terms of shared history or shared task, to communicate freely. With the CallHome corpus, the issue is different: the speakers were friends or family, and frequently had so much shared history that following the conversation can be difficult.

Table 1: Basic corpus statistics
Corpus Number Number Average Vocabulary
  of dialogues of turns dialogue length size (approx.)
TRAINS-93 98 5900 4.5 min. 860
Maptask 128 - 7 min. 1675
Switchboard I 2430 - 6 min. -
Monroe 20 4794 20 min. 1550

Table 1 shows the number of turns, number of words, and average dialogue length for the TRAINS, Maptask, Monroe and Switchboard corpora.

The Monroe corpus is different from these other corpora in several ways:

Because of these differences, the Monroe dialogues are longer and may exhibit more use of mixed-initiative and different types of discourse contributions.

Dialogue collection

In this section, we describe how the Monroe corpus was collected. We give information about:

Conclusions

The Monroe corpus is a useful addition to existing collections of task-oriented spoken dialogues. It is a much smaller corpus than the SWITCHBOARD, ATIS and Maptask corpora (about the same amount of time as the TRAINS corpus). However, the tasks involved are considerably more complex than those used in previous collections. As a result, one sees a variety of types of problem-solving. Also, because subjects were often given an extra task, there is the opportunity to observe re-planning. Finally, the size of individual tasks leads to innovative use of reference and organization of subtasks. The dialogues also exhibit more complexity than previous collections. Because the participants were strongly encouraged to be equal partners in the conversation and planning, there is more mixed-initiative behavior. Also, because there are many alternative ways to solve tasks, there are a variety of argumentation acts, including explanations, motivations, justifications, and elaborations.

About this document ...

This document was generated using the LaTeX2HTML translator Version 98.1 release (February 19th, 1998)

Copyright © 1993, 1994, 1995, 1996, 1997, Nikos Drakos, Computer Based Learning Unit, University of Leeds.

The command line arguments were:
latex2html -no_navigation -split 0 monroe.tex.

The translation was initiated by Amanda Stent on 2001-12-11


Amanda Stent
2001-12-11