Multilingual NLP and Interpretability

dr. Miryam de Lhoneux

Abstract:

Abstract: In this talk, I give a high-level introduction to my research, with the objective to introduce myself to the group as well as to help identify potential future collaborations. The talk is organised around two main themes: multilingual NLP and interpretability. I give an overview of the current state of multilingual NLP where I give arguments why dependency parsing is ahead of the game, thanks to the Universal Dependencies dataset. I describe successes and shortcomings of transfer learning for universal dependency parsing and what it means for low-resource NLP. I then describe ongoing and future work in NLP for truly low-resource languages. I finally describe ongoing work on interpreting what neural networks learn about language.

Multimodal Face Naming

Tingyu Qu

Abstract:

How to integrate multimodal information for multimodal news analysis is a challenging problem. Faces and names often contain important information in multimodal news. In this work, we focus on aligning faces in news images with names in corresponding captions. In the seminar, I’ll first share some experiences with multimodal news summarization as the motivation. Then I’ll pitch the ideas we have on face-naming. Looking forward to the discussion.

Updated: