ISCA Archive ICSLP 2002
ISCA Archive ICSLP 2002

Semantic inference: a data-driven solution for NL interaction

Jerome R. Bellegarda

In sufficiently limited domains, natural language interaction is possible even in the absence of actual natural language understanding. This is particularly true for goal-directed command and control, where the understanding task can essentially be cast as an Nway classification problem. (Data-driven) semantic inference is an approach to such tasks which in principle allows for unrestricted command/query formulation. It relies on a latent semantic analysis framework, whereby each unconstrained word string is automatically mapped onto the intended action through a semantic classification against the set of supported concepts. The objective of this paper is to compare this approach with other like-minded Nway classification methods, such as based on finite-state grammars or nearest-neighbor techniques. All experiments are conducted in the context of a desktop user interface control task involving 113 different actions. Results illustrate some of the performance and robustness benefits of semantic inference.