Multi Domain Language Model Adaptation using Explicit Semantic Analysis
Kevin Kilgour, Kraft, Florian, Sebastian Stüker, Alex Waibel · 2011
This paper presents an adaptive multi domain language model built from large sources of pre existing human created structured data. The sources’ structure is exploited to create a large array of ngram language models which are dynamically interpolated at decoding time to produce a context dependent language model that continuously adapts itself to the current domain. Because the use of human annotators is expensive and impractical we explore existing sources of human created structured data and how to extract our desired data from them. The language model is evaluated on its performance with a speech recognition system used to decode the Quaero 2009 evaluation data set. Compared to the baseline language model of our Quaero 2009 evaluation system our proposed adaptive language model reduces the WER of the speech recognition system by 0.5% absolute with some shows showing reductions of up to 14.4%. Index Terms: Speech Recognition, Language Model Adaptation, Explicit Semantic Analysis