Exploiting end of sentences and speaker alternations in language modeling for multiparty conversations

Abstract

The effective handling of end-of-sentences and speaker alternations, both of which are frequently observed in multiparty conversations, in recurrent neural network language models (RNNLMs) is investigated. This kind of auxiliary information is represented as context cues and feature vectors. The former representation can be inserted directory into a transcription and treated as a word token, while the latter serves as auxiliary input to the neural networks. Experimental comparisons using multiparty conversation data, including the AMI meeting corpus, demonstrated that both representations contribute to improvement of the RNNLMs, and that dealing with the end-of-sentences is important, especially on the multiparty conversations.