JSALT 2015 Research
Research Group: Continuous Wide-Band Machine Translation
Continuous space models (CSMs, a.k.a., neural networks, deep learning, connectionism), are rapidly revolutionizing related fields of speech recognition and computer vision, and are beginning to make a big impact on language processing. These models allow for learning highly general characteristics language, ranging from the syntax and semantics of morphemes and words, to phrasal structure, all the way to sentence and discourse semantics. This workshop will develop CSM techniques for machine translation. Specifically, the project aims to learn representations of words, syntax and discourse in one language to inform the translation of this text into another language. It will develop local models of individual word translation, including morphological decomposition, and structural global models over the range of complex decisions required in translating a sentence, most notably reordering. A key focus will be on the document context, such that text in neighboring sentences can affect the translation decisions in the sentence, both in a shallow manner as well as based on linguistic models of discourse, such as Rhetorical Structure Theory. This is an opportunity to use and contribute to leading open source toolkits in neural network modeling and machine translation.
- Trevor Cohn (University of Melbourne)
- Chris Dyer (Carnegie Mellon University)
- Jacob Eisenstein (Georgia Tech)
- Kaisheng Yao (Microsoft Research)