Semi-Supervised Word Sense Disambiguation for Mixed-Initiative Conversational Spoken Language Translation

Lexical ambiguity can cause critical failure in conversational spoken language translation (CSLT) systems due to the wrong sense being presented in the target language. In this paper, we present a framework for improving translation of ambiguous source words that (a) constrains statistical machine translation (SMT) decoding with phrase pair clusters to select a desired sense for translation; (b) automatically predicts the intended sense of an ambiguous source word given its context; and (c) combines the above to define a set of interactive strategies to confirm the intended sense of an ambiguous word and guide the system to the correct translation. The novel use of this framework in a realworld CSLT system distinguishes our approach from the existing work focusing on word sense disambiguation (WSD) for non-interactive, batch-mode SMT. In addition to reporting metrics that evaluate this approach in an interactive spoken language translation system, we also present offline assessments of the component technologies, viz. constrained SMT decoding with sense-specific phrase pair clusters, and automated word sense prediction.