Output details
11 - Computer Science and Informatics
University of Edinburgh
Covariance in Unsupervised Learning of Probabilistic Grammars
<22> Originality: This paper makes novel first use of the Bayesian setting for the problem of grammar induction. The paper derives novel prior distributions for learning the syntax of language, and its results considerably improved the previously best experimental results in unsupervised parsing.
Significance: Solutions for the problem of grammar induction are long sought-after: they have implications both from the engineering perspective and the scientific perspective (for understanding how humans acquire language).
Rigour: The paper includes thorough experiments with six languages, and shows that the results generalize to these languages.