Parsing and Grammatical Induction as Constraint Solving

点击翻译按钮获取中文摘要

Motivation: Roughly 7,000 languages are spoken in the world, including about 60 distinct indigenous languages in Canada alone. Many of them are under-studied and poorly supported by computational and computerizedsystems: the number of linguists who can study them is insufficient and most of their efforts pour into mainstream languages. Efficient methods that automatically induce grammars that can be incorporated into modern technological and educational tools would greatly impact the Grammar Induction field and related Natural Language Processing (NLP) areas, while significantly increasing survival chances for endangered languages, with consequent socio-economic potential.State-of-the-art: Most grammar induction models involve probabilities (so that learning a grammar amounts to selecting a model from a pre-specified model family) or statistical models of machine learning. With respect to precision, reliability and explanatory power, such models are inferior to rule and inference-based models, but they can often overcome the inflexibilities of classical NLP methods through extensive search on voluminous data sets, and can achieve impressive results when what matters is only to simulate linguistic understanding in the Turing test sense. A few models that use linguistic information from one language for the task of describing another language rather than addressing induction in full, are usually restricted to specific tasks, e.g. disambiguating the other language and may require parallel corpora.Position of the research within the state-of-the-art: In contrast, with support from my previous NSERC grant, my students and I developed a novel model for inducing unknown grammars from known ones, which works for more than just specific tasks and needs neither a pre-specified model family, nor parallel corpora, nor any of the typical models of machine learning. Our model makes it possible to generate an under-studied language's grammar using representative and correct input sentences in that language, together with its lexicon relevant to that input, all of which is parsed with respect to the correct grammar of a well-studied language. We call our model the Womb Grammar Model (WGM) because it can generate new grammars given appropriate input, much as human wombs can generate all races.Proposed research: The present proposal focuses on solidifying our experimental WGM results into a full-blown and executable computational linguistic theory of parsing and of grammatical inference (long term goal), with particular focus on semantic and pragmatic interpretation (medium-term). We shall fine-tune this theory by interleaving its development with test-bed short-term concrete applications to grammar induction (in particular, around linguistic diversity preservation in Canada), to parsing per se, and to linguistic experimentation (as testing alternative linguistic constraints).