Learning Belief Networks for Language Understanding

This paper is about learning Belief Networks (BNs) for spoken language understanding. The BNs are used to infer the communicative goal of a user's information-seeking query in a restricted domain. We assume that a restricted domain generally has a finite number of communicative goals. The problem is formulated as N binary classifications (one per goal), and each is performed by a BN. This formulation allows for the identification of queries with multiple goals, as well as queries with out-ofdomain goals. The BN topologies are automatically learnt according to the Minimum Description Length (MDL) principle. We aim to learn the least complex topologies that can best model the available data set. These enhanced topologies are compared with a pre-defined, basic topology. Experiments with the ATIS-3 corpus shows that the enhanced topology improves goal identification accuracies from 83.7% to 91.5% when a single output goal is evaluated, and from 66.0% to 83.1% when multiple output goals are evaluated.