EUROSPEECH 2001 Scandinavia
In this paper we present a novel method for creating language models for Spoken Dialogue Systems (SDS). The idea is based on combining the linguistic structure and the limited requirements for training data of grammar-based models with the robustness of stochastic models regarding spontaneous speech. Our algorithm requires a set of sentences as input, in order to train a Hidden Markov Model (HMM). Classes containing words or phrases with semantic-syntactic similarities are formed automatically and simultaneously with the construction of the HMM. The states and observations of the HMM correspond to the word/phrase classes and words/phrases respectively. The resulting HMM incorporates grammatical structure provided by large context dependencies as well as coverage of ungrammatical spontaneous sentences provided by statistical estimations. The HMM is transformed to a Stochastic Finite-State Network (SFSN), which allows for variable history sizes with no specific upper limit. We used data from 3 different SDSs to evaluate the algorithm. The experiments carried out, resulted in precision and recall values regarding the classes formed, of 0.97 and 0.76 in average, respectively. There was also a reduction of perplexity (16.15% in average) compared to bigrams and a gain in recognition performance (keyword accuracy) of 6.2% compared to grammar-based models and 5.4% compared to bigrams.
Bibliographic reference. Georgila, Kallirroi / Fakotakis, Nikos / Kokkinakis, George (2001): "Efficient stochastic finite-state networks for language modelling in spoken dialogue systems", In EUROSPEECH-2001, 247-250.