Interspeech'2005 - Eurospeech
This paper presents a robust semantic model for one-stage interpretation of natural speech. Our semantic analysis uses no explicit syntactic and morphologic knowledge, which seems sufficient for narrow application domains. In contrast to previous approaches, our semantic model is embedded in a uniform, hierarchical, stochastic modeling framework together with acoustic-phonetic and lexical knowledge, and semantic representations are computed directly from acoustic observations through a one-stage decoding process. The decoder produces a hierarchical (tree-) structure of words and semantic category symbols by use of the so-called hierarchical language model (HLM). We discuss generation of HLM by mixing rule-based and data-driven language model (LM) generation techniques, namely weighted regular expressions, n-grams and exact LM. Different HLM configurations with varying discounting techniques, n-gram orders and scaling factors are examined. Experiments were conducted with an airport information dialogue application. The evaluation results are based on HLM perplexity and our previously published semantic tree accuracy.
Bibliographic reference. Thomae, Matthias / Fabian, Tibor / Lieb, Robert / Ruske, Günther (2005): "Hierarchical language models for one-stage speech interpretation", In INTERSPEECH-2005, 3425-3428.