Enrichissement de corpus par approche générative et impact sur les modèles de reconnaissance d'entités nommées
Résumé
Industrial applications of Named Entity Recognition (NER) are usually
confronted with imbalanced corpora. This could harm the performance of
trained models when dealing with unknown data. In this paper we develop
two generation-based data enrichment approaches to improve entity
distribution. We compare the impact of enriched corpora on NER models,
using both non-contextual and contextual embeddings, and a biLSTM-CRF as
entity classifier. The approach is evaluated on a contract renewal
detection task. The results show that the proposed enrichment
significantly improves the model's effectiveness on unkonwn data, while
not degrading the performance on the original test set.