Leveraging Twitter for Low-Resource Conversational Speech Language Modeling

In applications involving conversational speech, data sparsity is a limiting factor in building a better language model. We propose a simple, language-independent method to quickly harvest large amounts of data from Twitter to supplement a smaller training set that is more closely matched to the domain. The techniques lead to a significant reduction in perplexity on four low-resource languages even though the presence on Twitter of these languages is relatively small. We also find that the Twitter text is more useful for learning word classes than the in-domain text and that use of these word classes leads to further reductions in perplexity. Additionally, we introduce a method of using social and textual information to prioritize the download queue during the Twitter crawling. This maximizes the amount of useful data that can be collected, impacting both perplexity and vocabulary coverage.