semisupervised coreference resolution
Final Report Abstract
We achieved good results on a number of semi-supervised research problems and published them in the top venues of our field. In particular: - semi-supervised mining of co-reference chains (enabling a more data-driven approach to co-reference resolution) - semi-supervised learning of embeddings that are more appropriate for semantic tasks like co-reference resolution than standard embeddings (e.g. synonyms and antonyms are not similar) - semi-supervised induction of fine-grained types of entities (which are a constraint on possible co-reference). We published several data resources that will be helpful for future research on co-reference resolution, most importantly: http://cistern.cis.lmu.de/corefresources. By April 2018, we will have three PhD theses that have been successfully completed with support of this project. The main surprise was the advent of deep learning. This fundamental re-orientation of natural language processing research required substantial adaptation of the original work program.
