The "conversational agent" we are developing is a logically connected set of algorithms and representation structures that enable the tracking of contextual information and management of speaker interaction. It was first developed in the late 1980's in support of a non-speech, mixed-modality interface to the Cyc common sense knowledge base. This computational discourse framework extends formal linguistic theories of discourse semantics toward enabling computational processing of spontaneous language phenomena: extended dialogue segments, attentional focus phenomena, complex dependence relationships between referring expressions, and repair dialogue segments. We have adapted the original context tracking framework to voice-to-voice machine translation, non-dialogue text processing applications, and spoken dialogue and mixed-modality interfaces. The discourse representation and updating algorithms emphasize robustness in the environment of incomplete knowledge representation, imperfect speech recognition or sentence analysis, and temporary misconceptions on the part of the user or the dialogue agent itself. In the first half of this talk I will describe the dialogue agent framework and discuss its implementation to serve spoken dialogue applications.
In the second half of this talk I will describe our work on the extraction and visualization of dialogue topology information from spontaneous human-human dialogue data. We use the label "dialogue topology" to mean the collection of features that can be extracted in lieu of or in advance of standard natural language understanding. The analysis and the data can be used to train adaptive components of the dialogue based systems mentioned above. Furthermore, the analysis of dialogue data is an end in itself, crucial for tasks such as extraction of data from stored video interviews or other dialogues, filtering of spoken documents in a corpus of dialogues, gisting, topic spotting, and other correlates of corpus based text processing. I will discuss the methods and preliminary results of extraction and use of topological data, and the portrayal of dialogue topology with a visualization tool.
Note Regarding Empirical Discourse Research: We are part of an ongoing international research consortium called the Discourse Resource Initiative whose purpose is to coordinate the results of researchers at universities, government laboratories, and commercial companies in the US, Europe, Japan, and elsewhere, to promote sharing of resources related to discourse processing. The DRI's web site is reachable at http://www.georgetown.edu/luperfoy/Discourse-Treebank/dri-home.html