Posted: 18th September 2013

Three papers by members of the Computational Linguistics Group have been accepted for presentation at the 2013 Conference on Empirical Methods in Natural Language Processing, to be held in Seattle on 19-20 October.

• Adaptor Grammars for Learning Non-Concatenative Morphology, by Jan Botha and Phil Blunsom, extends adaptor grammars to a richer grammatical formalism to propose a method for unsupervised learning of concatenative and non-concatenative morphology, obtaining positive results on Arabic and Hebrew
• Two Recurrent Continuous Translation Models, by Nal Kalchbrenner and Phil Blunsom, introduces a deep learning approach to machine translation that is alignment-free and encompasses an inherent notion of similarity between words, phrases and sentences.
• Prior Disambiguation of Word Tensors for Constructing Sentence Vectors, by Dimitri Kartsaklis and Mehrnoosh Sadrzadeh, shows how to improve a number of tensor-based compositional distributional models of meaning in a variety of tasks by introducing a step of disambiguation prior to composition.