Pierre Lison, Aliaksandr Hubin, Jeremy Barnes, Samia Touileb
5 min
Named Entity Recognition (NER) models typically rely on large, hand-annotated datasets. When moving to a new domain where such data is unavailable, performance often drops significantly. This paper addresses the challenge of training NER models in the absence of target-domain labels by using a weak supervision framework.
The authors collect a wide range of "labelling functions"—including out-of-domain neural models, gazetteers (lists of known entities), and heuristic rules—to automatically annotate target-domain text. Because these sources are noisy and often disagree, the authors use a hidden Markov model (HMM) to aggregate these predictions. The HMM estimates the reliability of each labelling function in an unsupervised manner, producing a probabilistic "gold standard" that is then used to train a final sequence labelling model.
The proposed method was evaluated on the CoNLL 2003 dataset and a collection of financial news articles from Reuters and Bloomberg. By aggregating diverse labelling functions, the model achieved an improvement of approximately 7 percentage points in entity-level F1 scores compared to a standard out-of-domain neural NER model. This demonstrates that combining multiple weak signals can effectively bootstrap a model to perform well in a new domain without manual annotation.
This research provides a scalable way to deploy NER systems in specialized fields (like finance or medicine) where expert-annotated data is expensive or impossible to obtain. By leveraging existing knowledge bases and pre-trained models as labelling functions, practitioners can adapt their systems to new contexts with minimal effort, significantly reducing the barrier to entry for high-quality information extraction.
Named Entity Recognition (NER) performance often degrades rapidly when applied to target domains that differ from the texts observed during training. When in-domain labelled data is available, transfer learning techniques can be used to adapt existing NER models to the target domain. But what should one do when there is no hand-labelled data for the target domain? This paper presents a simple but powerful approach to learn NER models in the absence of labelled data through weak supervision. The approach relies on a broad spectrum of labelling functions to automatically annotate texts from the target domain. These annotations are then merged together using a hidden Markov model which captures the varying accuracies and confusions of the labelling functions. A sequence labelling model can finally be trained on the basis of this unified annotation. We evaluate the approach on two English datasets (CoNLL 2003 and news articles from Reuters and Bloomberg) and demonstrate an improvement of about 7 percentage points in entity-level F 1 scores compared to an out-of-domain neural NER model.
Alex: That's a key insight the paper builds in directly. They add what you might call a consistency constraint: once the model settles on an identity for a word in a document, it holds that identity throughout. This is actually quite powerful, because it means a strong signal early on—where multiple rules agree—can anchor the label for every later appearance of that same word, even in sentences where the rules are less certain.
Sam: So the model is doing two things at once. It's learning which rules to trust, and it's using the structure of the document itself—the fact that names don't change meaning halfway through—to sharpen its guesses.
Alex: Exactly. And the output of all this isn't a final answer—it's a set of automatically generated training labels. Those labels, imperfect as they are, get used to train a proper machine learning model. So the weak supervision isn't the end product; it's the scaffolding that makes the end product possible without a single human annotation.
Sam: Did it actually work? When they tested this, did the automatically generated labels produce a model that performed meaningfully better?
Alex: The paper suggests it does. In tests on news articles, the approach produced a notable improvement in identifying named entities compared to models that were simply trained on data from a different domain and applied without any adaptation. The gap was meaningful enough to suggest the method is doing real work, not just adding noise.
Sam: Though I'd imagine this has limits. If all your rules share the same blind spot—say, none of them know anything about a particular industry's jargon—the model just inherits that gap.
Alex: That's a fair point, and it's worth being clear about. The model's ceiling is set by the rules you bring in. If every rule misses the same category of entity, no amount of clever weighting will recover it. The method reduces the cost of getting started, but it doesn't eliminate the need for domain knowledge when you're designing those initial rules.
Sam: So it's less "the machine figures everything out on its own" and more "a domain expert writes a handful of rough rules, and the machine makes the most of them."
Alex: That's the right framing. It shifts the human effort from the slow, expensive work of labeling thousands of documents to the faster work of writing a few informed heuristics. For an organization entering a new field with no existing labeled data, that's a meaningful practical difference. Thanks for listening to ResearchPod.