Hierarchical Latent Words Language Models for Automatic Speech Recognition

This paper presents hierarchical latent words language models (h-LWLMs) for improving automatic speech recognition (ASR) performance in out-of-domain tasks. Language models called h-LWLM are an advanced form of LWLM that are one one hopeful approach to domain robust language modeling. The key strength of the LWLMs is having a latent word space that helps to efficiently capture linguistic phenomena not present in a training data set. However, standard LWLMs cannot consider that the function and meaning of words are essentially hierarchical. Therefore, h-LWLMs employ a multiple latent word space with hierarchical structure by estimating a latent word of a latent word recursively. The hierarchical latent word space helps us to flexibly calculate generative probability for unseen words. This paper provides a definition of h-LWLM as well as a training method. In addition, we present two implementation methods that enable us to introduce the h-LWLMs into ASR tasks. Our experiments on a perplexity evaluation and an ASR evaluation show the effectiveness of h-LWLMs in out-of-domain tasks.