Dieuwke Hupkes bio photo

Dieuwke Hupkes

Postdoc and scientific manager of the ELLIS Amsterdam unit, interested in structures in language

Email Twitter Instagram Github

Visiting address

Room F3.260
Science Park 904
1098 XH Amsterdam

Curriculum Vitae


Research

Overview

Research Interests

I am interested in many different aspects of natural language, a selection:

  • Computational and cognitive models of natural language processing
  • Neurocomputational models of language (processing)
  • Hierarchical compositionality and recursion
  • Learning biases
  • Semantic parsing
  • Recurrent neural networks
  • Statistical parsing
  • Language acquisition.

Diagnostic Classifiers

Currently I am investigating how recurrent neural networks can process hierarchical compositionality. In this paper, some preliminary findings for this project are described, as well as an extensive account of how a task involving hierarchical compositionality can be implemented in a recursive neural network (work of my colleague Sara Veldhoen). The source code of this project is available at https://github.com/dieuwkehupkes/processing_arithmetics.

Previous Projects

POStagging of historical Dutch

Before I started my PhD I was employed in the digital humanities project CREATE, where I worked on improving part of speech tagging of 17th century Dutch, for which I used parallel corpora consisting of diachronic bible texts. This project taught me several interesting things about my own language and how it is to work in the field of digital humanities, but also got me interested in semi supervised learning. This paper, that I presented at LREC in 2016, describes the main findings of the project. It always felt like this project wasn’t quite finished when I left it, so if you are interested in POStagging historical languages, drop me a line!

Compositionality of translation

In my master thesis, I aimed to empirically establish the compositionality of translation according to dependency parses. If you are interested in the results, my thesis can be found here.