Multi-layer Dialogue Annotation for Automated Multilingual Customer Service

A flexible, platform-independent Java annotation tool, called XDMLTool, takes plain-text dialogue files as input, and yields annotated files in the widely used XML format, which has been used to annotate several hundred call-center dialogues in France, the UK and the US.

Abstract

For the AMITIÉS multilingual human-computer dialogue project [1], we have developed new methods for the manual annotation of spoken dia-logue transcriptions from European financial call centers on multiple levels. We have modified the DAMSL schema [2] to create a dialogue act taxon-omy appropriate to the functions of call center dia-logues. We use a domain-independent framework populated with domain-specific lists to capture the semantics of spoken dialogues. Our new flexible, platform-independent Java annotation tool, called XDMLTool, takes plain-text dialogue files as in-put, and yields annotated files in the widely used XML format. To date, XDMLTool has been used to annotate several hundred call-center dialogues in France, the UK and the US. We present definitions of each tag as well as examples in English and French. These annotation methods are developed for an experimental system that automates finan-cial call centers in Europe. The multi-level annota-tion scheme has been used to develop a prototype triaging application for financial services. 1