Discourse-Based Modeling for AAC

This paper presents a method for an AAC system to predict a whole response given features of the previous utterance from the interlocutor. It uses a large corpus of scripted dialogs, computes a variety of lexical, syntactic and whole phrase features for the previous utterance, and predicts features that the response should have, using an entropy-based measure. We evaluate the system on a held-out portion of the corpus. We find that for about 3.5% of cases in the held-out corpus, we are able to predict a response, and among those, over half are either exact or at least reasonable substitutes for the actual response. We also present some results on keystroke savings. Finally we compare our approach to a state-of-the-art chatbot, and show (not surprisingly) that a system like ours, tuned for a particular style of conversation, outperforms one that is not. Predicting possible responses automatically by mining a corpus of dialogues is a novel contribution to the literature on whole utterance-based methods in AAC. Also useful, we believe, is our estimate that about 3.5-4.0% of utterances in dialogs are in principle predictable given previous context.