Named-Entity Recognition: Resurrecting Old School Machine Learning in the Era of Deep Learning

titleNamed-Entity Recognition: Resurrecting Old School Machine Learning in the Era of Deep Learning
start_date2024/12/20
schedule11h
onlineyes
location_infovisioconférence Big Blue Button
summaryIn this talk, I will show that we can bridge old-school methods (finite-state automaton and k-means) with neural networks to achieve SOTA results. First, I will present my EMNLP 2024 paper [1] on discontinuous named-entity recognition, an overlooked setting in the literature. SOTA methods are based on complex pipelines with intricate neural architectures. I will show that using finite-state automaton, we can build a word tagging method that achieves competitive experimental results while being 40x-50x faster than SOTA. Unlike previous attempts to use work tagging in this setting, the proposed approach guarantees well-formedness of predictions. Second, I will present our COLING 2025 paper [2] on few-shot learning for named-entity recognition. Many approaches in this setting are based on variants of nearest neighbor classification. Unfortunately, they cannot leverage unlabeled data. We propose a novel approach for semi-supervised few-shot learning based on joint k-means and subspace selection. For named-entity recognition, a difficulty arises from the fact that most words are tagged with O (outside a mention): when we include a large amount of unlabeled data, the model can easily collapse to assigning tag O for all words. To prevent this issue, we include a ratio-constraint in the fine-tuning step. [1] A fast and sound tagging method for discontinuous named-entity recognition (Caio Corro) https://arxiv.org/abs/2409.16243 [2] Few-shot domain adaptation for named-entity recognition via joint constrained k-means and subspace selection (Ayoub Hammal, Benno Uthayasooriyar, Caio Corro) https://arxiv.org/abs/2412.00426
responsiblesBawden

Workflow history

from state (1)to statecommentdate
submittedpublished2024/12/19 16:40 UTC
speakers
Corro
Event #2479402 - created on 2024/12/19