Natural Language Processing at the Sub-Word Level

点击翻译按钮获取中文摘要

Processing language involves processing words. However, words are not indivisible abstract atoms - they are made of smaller units, such as morphemes, syllables, letters, and phonemes. In computational linguistics, words are analyzed on distinct levels: phonetic, phonological, orthographic, morphological, etc. These levels of representation underlie such important tasks as morphological parsing, speech synthesis and recognition, spell-checking, and stemming. Since different levels are strongly inter-related, the understanding of the interactions between them is crucial to advancing the state of the art in word-oriented applications. My objective is to investigate such interactions, and develop algorithms for alignment and conversion between levels. In particular, I will focus on the tasks of grapheme-to-phoneme conversion and transliteration. My general approach to the problem will be to incorporate linguistic knowledge into advanced machine-learning techniques, which, given sufficient training data, substantially outperform rule-based approaches. However, the latter often achieve impressive accuracy without any need for training data. They also tend to perform well across different domains. I will explore the ways of combining the two paradigms, by exploiting linguistic understanding for guiding and informing machine learning approaches, as well as for pre- and post-processing their training data. The long term goals of this research programme are acquiring deeper understanding of the dependencies between various representations, and leveraging the novel insights in order to advance the state of the art in natural language processing. I am confident that the implementation of the planned projects will result in successful applications in other areas of natural language processing.