Séminaire de Recherche en Linguistique

Core NLP technologies like morphosyntactic parsers, entity classifiers and coreference resolvers are pervasive in language technology companies. A key challenge going forward, is to improve these technologies for languages other than English. This is so that we can reach the majority of the world\\\'s population in their own language. One recurring objective is to make the accuracy of technology X for language Y as good as it is for English. This remarkably hard, due in part to a lack of resources, but also inherent assumptions built into our models that make sense for English, but not for most languages. In this talk, I will highlight some research efforts to improve coverage of our multilingual syntactic analyzers at Google. One such effort is to characterize restricted classes of non-projective structures that occur empirically across many languages, but can be parsed with efficient inference algorithms. We show that extensions to standard transition-based parsing algorithms can lead to state-of-the-art accuracies for highly non-projective languages while maintaining asymptotic linear run-time.