Egy emBERT próbáló feladat

Dávid Márk Nemeskey

Abstract:

In the last couple of years deep, contextual embeddings have superseded traditional, manually compiled feature sets in most NLP tasks. However, the Hungarian NLP pipelines (e-magyar, magyarlanc) still manual features. In this article, we introduce the emBERT module, which allows the integration of contextual embedding-based classifiers into e-magyar, via the transformers library.

The module provides classifiers for named netity recognition and NP chunking, achieving state-of-the-art performance.