Language Modeling for limited-data domains

Abstract

With the increasing focus of speech recognition and natural language processing applications on domains with limited amount of in-domain training data, enhanced system performance often relies on approaches involving model adaptation and combination. In such domains, language models are often constructed by interpolating component models trained from partially matched corpora. Instead of simple linear interpolation, we introduce a generalized linear interpolation technique that computes context-dependent mixture weights from features that correlate with the component confidence and relevance for each n-gram context. Since the n-grams from partially matched corpora may not be of equal relevance to the target domain, we propose an n-gram weighting scheme to adjust the component n-gram probabilities based on features derived from readily available corpus segmentation and metadata to de-emphasize out-of-domain n-grams. In scenarios without any matched data for a development set, we examine unsupervised and active learning techniques for tuning the interpolation and weighting parameters. Results on a lecture transcription task using the proposed generalized linear interpolation and n-gram weighting techniques yield up to a 1.4% absolute word error rate reduction over a linearly interpolated baseline language model. As more sophisticated models are only as useful as they are practical, we developed the MIT Language Modeling (MITLM) toolkit, designed for efficient iterative parameter optimization, and released it to the research community.

(cont.) With a compact vector-based n-gram data structure and optimized algorithm implementations, the toolkit not only improves the running time of common tasks by up to 40x, but also enables the efficient parameter tuning for language modeling techniques that were previously deemed impractical.