[Paper Review] A Survey of Active Learning for Text Classification using Deep Neural Networks

Christopher Schröder, Andreas Niekler|arXiv (Cornell University)|Aug 17, 2020

Machine Learning and AlgorithmsComputer Science104 references60 citations

TL;DR

This paper surveys neural-network–based active learning for text classification, presenting a taxonomy of query strategies, reviewing recent NN advances in NLP, and identifying gaps and open questions for future research.

ABSTRACT

Natural language processing (NLP) and neural networks (NNs) have both undergone significant changes in recent years. For active learning (AL) purposes, NNs are, however, less commonly used -- despite their current popularity. By using the superior text classification performance of NNs for AL, we can either increase a model's performance using the same amount of data or reduce the data and therefore the required annotation efforts while keeping the same performance. We review AL for text classification using deep neural networks (DNNs) and elaborate on two main causes which used to hinder the adoption: (a) the inability of NNs to provide reliable uncertainty estimates, on which the most commonly used query strategies rely, and (b) the challenge of training DNNs on small data. To investigate the former, we construct a taxonomy of query strategies, which distinguishes between data-based, model-based, and prediction-based instance selection, and investigate the prevalence of these classes in recent research. Moreover, we review recent NN-based advances in NLP like word embeddings or language models in the context of (D)NNs, survey the current state-of-the-art at the intersection of AL, text classification, and DNNs and relate recent advances in NLP to AL. Finally, we analyze recent work in AL for text classification, connect the respective query strategies to the taxonomy, and outline commonalities and shortcomings. As a result, we highlight gaps in current research and present open research questions.

Motivation & Objective

• Provide a taxonomy of active learning query strategies applicable to text classification.
• Survey how deep neural networks and modern NLP representations affect active learning.
• Analyze recent experiments and datasets used in NN-based active learning for text classification.
• Identify gaps, limitations, and open research questions to guide future work in this area.

Proposed method

• Develop a taxonomy of query strategies based on input information (data, model, prediction, and random baselines).
• Relate neural-network uncertainties, small-data challenges, and transfer learning to active learning design.
• Survey recent NN-based text classification advances (embeddings, LMs, transformers) and map them to active learning.
• Review and synthesize experimental setups across studies, including datasets, models, and query strategies.
• Highlight commonalities, limitations, and open research questions in NN-based AL for text classification.

Experimental results

Research questions

• RQ1 What are the main categories of query strategies for AL in text classification, especially for deep neural networks?
• RQ2 How do recent NLP advances (embeddings, language models, transfer learning) influence active learning effectiveness and practices?
• RQ3 What are the common experimental patterns and gaps in NN-based AL for text classification, and what open questions remain?

Key findings

• Uncertainty-based and disagreement-based prediction strategies dominate NN-based AL for text classification, but their effectiveness with modern NN architectures is mixed.
• NN uncertainty estimation remains challenging and can limit AL performance, especially on small labeled datasets.
• Transfer learning and contextualized language models reduce data needs, enabling more effective AL in small-data regimes.
• Recent experiments show variability and limited comparability across datasets, indicating a need for standardized benchmarks and broader evaluations.
• Advances in NLP representations may enhance AL, but practical adoption requires careful handling of uncertainty and dataset characteristics.