Invited Talks

Grammar, Meaning & Annotation

Johan Bos, University of Groningen

What is the role of computational grammars in semantic annotation? In the Parallel Meaning Bank, grammar plays a pivotal role. This has good sides, and bad sides. It is good, because annotation is ensured to be carried out in a systematic, consistent and efficient way. But it can also be counterproductive, as linguistic input can be full of surprises. In such cases the grammar is a showstopper. Well, you might say, why not bypass the grammar in such cases? Sure, but annotating meanings from scratch is not straightforward when the targets are expressive semantic representations, such as the Discourse Representation Structures from Discourse Representation Theory used in the Parallel Meaning Bank. I present a new notation for these meaning representations: without variables, without explicit recursion, and without reliance on grammar.

Parsing Typologically Diverse Languages

Miryam de Lhoneux, University of Copenhagen

This talk is about parsing typologically diverse languages. I first argue that the Universal Dependencies (UD) dataset is the best multilingual dataset that we currently have and allows us to ask general questions that are relevant for multilingual NLP. I then ask the question of how well our current parsers generalize across languages and the question of how we evaluate that.

I subsequently ask the question of how accurate our parsers currently are for truly low-resource languages. I explain recent developments in cross-lingual learning that are great at leveraging data from related languages and that improve parsing accuracy for low-resource languages. I show that for low-resource languages for which we do not have a high-resource related language, our parsers are currently highly inaccurate. Since such cases represent the majority of world languages, we might want to shift our focus on these. I finally suggest that we may find answers in the use of typological information, discuss work that has tried to do that and highlight what more can be done.