Swiss chocolate at CAp 2017 NER challenge: partially annotated data and transfer learning

In this paper, we describe several deep learning approaches for Named Entity Recognition (NER) for the case where only little annotated data is available. We show that training a separate network per entity greatly improves the precision of a network and transfer learning on a different language or a partially annotated corpus increases the F1-score by up to 7 points. Our transfer learning system was evaluated in the CAp 2017 competition for Named Entity Recognition on French tweets, where it achieved 5th place, obtaining an F1-score of 50.05.