Experiments from LIMSI at the French Named Entity Recognition Coarse-grained task

Overview & Synopsis

This paper presents the participation of the LIMSI team in the HIPE 2020 Challenge on the Coarse-grained named entity recognition task for French. Our approach jointly predicts the literal and metonymy entities. For this, a Camem-BERT base model and a CRF model were used. We submitted three systems: a joint model using only CamemBERT, a joint model extended with a CRF layer, and a CamemBERT model without joint option. Experimental results show that the second system achieved best results on the literal tags (F1=.814) while the third system performed best (F1=.667) on the metonymy tags. The second system allowed us to obtain our best results on both the dev and test datasets for the literal tags. Nevertheless, we observed a difference on the metonymy tags where our first system obtained best results on the dev dataset (F1=.663) while our third system performed best on the test dataset (F1=.667).