Contextual Language Model Adaptation for Conversational Agents

Abstract

Statistical language models (LM) play a key role in Automatic Speech\nRecognition (ASR) systems used by conversational agents. These ASR systems\nshould provide a high accuracy under a variety of speaking styles, domains,\nvocabulary and argots. In this paper, we present a DNN-based method to adapt\nthe LM to each user-agent interaction based on generalized contextual\ninformation, by predicting an optimal, context-dependent set of LM\ninterpolation weights. We show that this framework for contextual adaptation\nprovides accuracy improvements under different possible mixture LM partitions\nthat are relevant for both (1) Goal-oriented conversational agents where it's\nnatural to partition the data by the requested application and for (2) Non-goal\noriented conversational agents where the data can be partitioned using topic\nlabels that come from predictions of a topic classifier. We obtain a relative\nWER improvement of 3% with a 1-pass decoding strategy and 6% in a 2-pass\ndecoding framework, over an unadapted model. We also show up to a 15% relative\nimprovement in recognizing named entities which is of significant value for\nconversational ASR systems.\n