Grammar-based context-specific statistical language modelling

This paper shows how we can combine the art of grammar writing with the power of statistics by bootstrapping statistical language models (SLMs) for Dialogue Systems from grammars written using the Grammatical Framework (GF) (Ranta, 2004). Furthermore, to take into account that the probability of a user’s dialogue moves is not static during a dialogue we show how the same methodology can be used to generate dialogue move specific SLMs where certain dialogue moves are more probable than others. These models can be used at different points of a dialogue depending on contextual constraints. By using grammar generated SLMs we can improve both recognition and understanding performance considerably over using the original grammar. With dialogue move specific SLMs we would be able to get a further improvement if we had an optimal way of predicting the correct language model.