Word-Phrase-Entity Recurrent Neural Networks for Language Modeling

Interspeech 2016 |

The recently introduced framework of Word-Phrase-Entity language modeling is applied to Recurrent Neural Networks and leads to similar improvements as reported for n-gram language models. In the proposed  architecture, RNN LMs do not operate in terms of lexical items (words), but consume sequences of tokens that could be words, word phrases or classes such as named entities, with the optimal representation for a particular input sentence determined in an iterative manner. We show how auxiliary techniques previously described for n-gram WPE language models, such as token-level interpolation and personalization, can also be realized with recurrent networks and lead to similar perplexity improvements.