Computational Linguistics Reading Group

This Friday, I’ll present our Semantic Tagging paper which was accepted to COLING (abstract and preprint available here:
https://arxiv.org/abs/1609.07053).

I’ll also talk about my current research, dealing with multilingual multitask model transfer, and show some preliminary results from a few interesting settings. Most previous work on model transfer deals with sharing model parameters between several tasks in a monolingual setting, or between several languages for a single task. However, parallel data exists for a large proportion of the languages in the world. This can be used to learn massively multilingual word representations. Additionally, corpora in many languages have been annotated with a variety of different tag sets, which can be leveraged by relying on correlations between these tag sets. Using a deep neural network, these language- and task correlations can be exploited in a language agnostic way, without any need for, e.g., handcrafting language- or task-dependent features, resulting in a single model with shared parameters for many languages and tasks.