Current Projects

!IS_A_LIST Parsing -- We have developed a framework for inducing structured models for parsing. The grammars that are learned allow efficient and accurate inference. We are currently investigating alternative, even better, approaches for inducing structure and estimating models.

!IS_A_LIST Machine Translation -- We are currently building an end-to-end machine translation system that leverages English syntax. As part of this project, we address related issues such as bilingual parsing algorithms and unsupervised transfer models.

!IS_A_LIST Unsupervised Coreference Resolution -- We have developed a model of unsupervised coreference resolution using nonparametric Bayesian techniques.

!IS_A_LIST Word alignment -- Word alignment is an important component of a machine translation system and is also useful for automatically inducing lexicons. We have developed a simple but extremely effective unsupervised training procedure that is competitive with even supervised approaches. Combining this method with a margin-based discriminative approach, we obtain a state-of-the-art word alignment system.

!IS_A_LIST Bitext Mutual Constraint -- Syntactic machine translation systems usually take bitexts with independently generated word alignments and syntactic parses as input. In this project, we aim to improve the quality of alignments and parses by constraining them against each other in a unified model.

!IS_A_LIST Speech Recognition -- Good speech recognition systems require fine grained manual tuning and complicated decoding algorithms. We want to leverage the techniques used in our parsing work to automatically induce latent models for speech recognition. Our goal is to drastically simplify the learning and decoding phases while keeping the performance.

!IS_A_LIST Historical Linguistics -- Current statistical approaches to historical linguistics are based on cognate judgement matrices and focus on inferring the topology and branch lengths of phylogenies. In contrast, our framework is based on phonology and models word mutation explicitly. This allows us to perform reconstruction of ancient word forms.