[Paper Review] Deep Active Learning for Sequence Labeling Based on Diversity and Uncertainty in Gradient

Yekyung Kim|arXiv (Cornell University)|Nov 27, 2020

Machine Learning and Algorithms23 references4 citations

TL;DR

This paper proposes W-BADGE, a novel active learning method for sequence labeling that combines uncertainty and diversity via weighted sampling of gradient embeddings. By prioritizing informative, diverse, and cost-effective samples based on sequence length, W-BADGE consistently outperforms uncertainty-only and diversity-only baselines across NER and slot-filling tasks, achieving near-full model performance with only 15% of the training data on the ATIS dataset.

ABSTRACT

Recently, several studies have investigated active learning (AL) for natural language processing tasks to alleviate data dependency. However, for query selection, most of these studies mainly rely on uncertainty-based sampling, which generally does not exploit the structural information of the unlabeled data. This leads to a sampling bias in the batch active learning setting, which selects several samples at once. In this work, we demonstrate that the amount of labeled training data can be reduced using active learning when it incorporates both uncertainty and diversity in the sequence labeling task. We examined the effects of our sequence-based approach by selecting weighted diverse in the gradient embedding approach across multiple tasks, datasets, models, and consistently outperform classic uncertainty-based sampling and diversity-based sampling.

Motivation & Objective

• To address the data efficiency challenge in sequence labeling tasks like NER and slot-filling, which require costly token-level annotation.
• To mitigate sampling bias in batch active learning caused by uncertainty-only sampling, which tends to select redundant, similar samples.
• To integrate both predictive uncertainty and structural diversity in unlabeled data using gradient embeddings for improved query selection.
• To enhance cost-effectiveness by weighting sample selection based on sequence length, favoring shorter, cheaper-to-annotate sequences.
• To empirically validate the proposed method across multiple datasets, models, and sequence labeling tasks.

Proposed method

• The method computes gradient embeddings for each unlabeled sequence from the final layer of the model to capture predictive uncertainty and representation diversity.
• It applies the BADGE algorithm to select a diverse batch of samples by maximizing the distance between gradient embeddings in the embedding space.
• A novel weighting scheme is introduced that scales sample selection probability inversely with sequence length, favoring shorter sequences to reduce annotation cost.
• The approach integrates uncertainty and diversity by selecting samples that are both uncertain (high gradient norm) and geometrically distant in the gradient space.
• Active learning proceeds iteratively: the model is retrained from scratch after each round using newly labeled data, ensuring no overfitting to initial data.
• The method is evaluated on CoNLL 2003 (NER), ATIS (slot-filling), and FMTOD (multilingual task-oriented) datasets with BiLSTM and BiLSTM encoder-decoder architectures.

Experimental results

Research questions

• RQ1 Can combining uncertainty and diversity in gradient embeddings improve active learning performance for sequence labeling tasks?
• RQ2 Does weighted sampling based on sequence length enhance cost-effectiveness without sacrificing performance?
• RQ3 How does the proposed method compare to uncertainty-only (e.g., BALD) and diversity-only (e.g., Coreset) baselines across different datasets and models?
• RQ4 To what extent can active learning reduce the required labeled data while maintaining high F1 performance?
• RQ5 Is the proposed method robust across different sequence labeling architectures and task types?

Key findings

• W-BADGE consistently outperformed uncertainty-only sampling (e.g., MNLP, BALD) and diversity-only sampling (e.g., Coreset, BADGE) across all datasets and models.
• On the ATIS dataset, W-BADGE achieved 98–99% of the F1 score of a model trained on the full dataset using only 15% of the labeled data.
• In the CoNLL 2003 NER task, W-BADGE significantly outperformed BALD, which previously held state-of-the-art performance in this task.
• The performance gap between W-BADGE and baselines was most pronounced in early training rounds, indicating faster learning efficiency.
• W-BADGE showed strong generalization, performing well on both the ATIS (slot-filling) and FMTOD (multilingual) datasets, even with varying model architectures.
• The weighted sampling strategy effectively reduced annotation cost by favoring shorter sequences, improving cost-effectiveness without degrading performance.