Spoken Dialogue Systems (SDS) are used to interact with intelligent agents through natu- ral language. Speech processing errors may cause the system to fail to generate an ap- propriate response. In this paper, we present a novel framework for understanding spoken dialogue in which utterance analysis is esca- lated through a multi-level system according to the feedback retrieved at the syntactic, se- mantic, and contextual/topic level. Analysis is applied incrementally at each level as the system attempts to resolve the uncertainty sur- rounding utterance interpretation. We demon- strate how our multi-level approach can be in- tegrated with other SDS components to im- prove its ability to recognize spoken task com- mands. We evaluate this by comparing the in- terpretation accuracy of utterances from two task domains given as input to an SDS, un- der two experimental conditions: one with the multi-level framework and one without.
@inproceedings{valenti2020multilevel,
title={A Multi-level Framework for Understanding Spoken Dialogue Using Topic Detection},
author={Andrew Valenti and Ravenna Thielstrom and Michael Gold and Felix Gervits and Matthias Scheutz},
year={2020},
booktitle={Proceedings of the 24th Workshop on the Semantics and Pragmatics of Dialogue (SemDial)},
url={https://hrilab.tufts.edu/publications/valenti2020multilevel.pdf}
}