Document Actions
Professor Alexander Clark visit & seminar, June 20th-23rd
Professor Alexander Clark, from Royal Holloway University of London, will be visiting the LARCA group June 20th-23rd to explore common research interests.
Professor Clark's main interests are related with unsupervised learning of natural language, and its relevance to first language acquisition. He has approached the topic both theoretically and practically: trying to define what a good definition of learnability is, trying to prove that you can learn languages according to various models of learnability, designing algorithms, and writing computer programs that can learn models of language both from synthetic and natural examples.
He will give a talk on Monday 20th, 15:00 (place to be announced) open to everyone.
Title: Unsupervised Learning of Context Free and Context-Sensitive Languages
Abstract:
Learning context free grammars was a central concern in the early days of generative linguistics and machine learning but fell out of favour as it was felt to be impossibly hard. Recently there has again been a great deal of research in this field using heuristic techniques (Klein and Manning, 2004). Here we will discuss a recently developed family of correct efficient algorithms based on distributional learning in the style of Zellig Harris, for learning discrete combinatorial objects such as strings, trees and graphs. We will focus on the case of context free grammars, where there are now a range of algorithms using a variety of different learning models: positive data only, positive and negative data, exact query learning and stochastic learning. All of these algorithms are based on the same representational idea, where the non-terminals of the representation correspond to well defined sets of strings. The most elementary result is one where the sets of strings concerned are the congruence classes of the language; this gives rise to a natural class of languages, the congruential context free languages, that lie half way between context free grammars and regular grammars. We will also discuss the natural extensions to context sensitive representations developed by Ryo Yoshinaka and in particular to the inference of multiple context free grammars: these are rich enough to handle cross-serial dependencies and displaced constituents.
Homepage: http://www.cs.rhul.ac.uk/home/alexc/
Professor Clark's main interests are related with unsupervised learning of natural language, and its relevance to first language acquisition. He has approached the topic both theoretically and practically: trying to define what a good definition of learnability is, trying to prove that you can learn languages according to various models of learnability, designing algorithms, and writing computer programs that can learn models of language both from synthetic and natural examples.
He will give a talk on Monday 20th, 15:00 (place to be announced) open to everyone.
Title: Unsupervised Learning of Context Free and Context-Sensitive Languages
Abstract:
Learning context free grammars was a central concern in the early days of generative linguistics and machine learning but fell out of favour as it was felt to be impossibly hard. Recently there has again been a great deal of research in this field using heuristic techniques (Klein and Manning, 2004). Here we will discuss a recently developed family of correct efficient algorithms based on distributional learning in the style of Zellig Harris, for learning discrete combinatorial objects such as strings, trees and graphs. We will focus on the case of context free grammars, where there are now a range of algorithms using a variety of different learning models: positive data only, positive and negative data, exact query learning and stochastic learning. All of these algorithms are based on the same representational idea, where the non-terminals of the representation correspond to well defined sets of strings. The most elementary result is one where the sets of strings concerned are the congruence classes of the language; this gives rise to a natural class of languages, the congruential context free languages, that lie half way between context free grammars and regular grammars. We will also discuss the natural extensions to context sensitive representations developed by Ryo Yoshinaka and in particular to the inference of multiple context free grammars: these are rich enough to handle cross-serial dependencies and displaced constituents.
Homepage: http://www.cs.rhul.ac.uk/home/alexc/
