Named Entity Recognition without Labelled Data: A Weak Supervision Approach

Named Entity Recognition (NER) performance often degrades rapidly when\napplied to target domains that differ from the texts observed during training.\nWhen in-domain labelled data is available, transfer learning techniques can be\nused to adapt existing NER models to the target domain. But what should one do\nwhen there is no hand-labelled data for the target domain? This paper presents\na simple but powerful approach to learn NER models in the absence of labelled\ndata through weak supervision. The approach relies on a broad spectrum of\nlabelling functions to automatically annotate texts from the target domain.\nThese annotations are then merged together using a hidden Markov model which\ncaptures the varying accuracies and confusions of the labelling functions. A\nsequence labelling model can finally be trained on the basis of this unified\nannotation. We evaluate the approach on two English datasets (CoNLL 2003 and\nnews articles from Reuters and Bloomberg) and demonstrate an improvement of\nabout 7 percentage points in entity-level $F_1$ scores compared to an\nout-of-domain neural NER model.\n