PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Active Learning for Text Classification using Deep Neural Networks

Christopher Schröder, Andreas Niekler

arXiv:2008.07267v1cs.CLcs.LG

TL;DR

NLP active learning faces costly annotation and limited coverage of neural approaches in existing surveys. This paper surveys deep-neural-network text classification active learning, its query strategies, and related NLP advances; it finds the area underresearched and experimental comparisons often lacking.

  • Problem

    Expert labeling can be costly for large NLP datasets, while existing surveys provide limited coverage of NN-based active learning and recent neural advances.

    Source: §1 · p. 1

  • Method

    The survey builds a taxonomy of query strategies and reviews NN-based active text classification, recent NLP advances, and prior experiments.

    Source: §1 · p. 2, §6 · p. 9

  • Results

    The survey finds NN-based active text classification is underresearched and often lacks comparability.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The survey identifies areas needing re-evaluation and develops research questions for future work on NN-based active text classification.

    Source: §1 · p. 2, §6 · p. 9

    Prior experiments often use different datasets, limiting comparability, and it remains unclear whether recent neural networks benefit from established query strategies.

    Source: §4.3 · pp. 8–9

Abstract

from arXiv · show

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.

1 Introduction

The introduction frames costly expert labeling as a challenge for NLP and asks whether neural-network advances can improve active learning. It presents a survey of NN-based text-classification active learning, its query-strategy taxonomy, recent NLP advances, and research gaps.

1 Introduction

  • Expert labeling can be time-consuming, costly, and infeasible for large NLP datasets with lengthy documents.Some tasks, such as patent and clinical text classification, may require domain experts.Source: §1 · p. 1
  • Active learning selects which instances a human annotator labels next, then retrains the model in an iterative loop.The survey assumes pool-based batch-mode learning, which suits closed text datasets and reduces retraining operations.Source: §1 · pp. 1–2
  • If neural-network advances transfer to active learning, practitioners could use fewer samples for the same performance or improve performance with the same data.The introduction cites advances in architectures, word representations, and fine-tuned pre-trained language models.Source: §1 · p. 2
  • Neural networks remain less common in active learning amid concerns about small-data training, artificial text-data generation, and unreliable prediction uncertainty.The survey identifies uncertainty as a challenge because many prominent query strategies rely on it.Source: §1 · p. 2
  • The survey builds a query-strategy taxonomy, reviews NN-based active text classification and recent NLP advances, and analyzes experiments to identify gaps and future questions.It also examines whether recent advances have been adopted for active learning.Source: §1 · p. 2

2 Related Work

The section reviews general and NLP-focused active-learning surveys, then identifies a gap: earlier surveys do not cover NN-based active learning for text classification in depth or reflect recent neural and language-model advances.

General active-learning surveys

  • Settles surveys active-learning scenarios and strategies, including variations in labeling costs and query types, and discusses mixed empirical and theoretical findings about effectiveness.The survey also connects active learning to areas such as reinforcement and semi-supervised learning.Source: §2 · p. 2

Uncertainty-based query-strategy surveys

  • Fu, Zhu, and Li organize uncertainty-based strategies around uncertainty for i.i.d. instances versus instance correlation, and compare their algorithmic time complexity.Correlation-based strategies account for feature, label, and structure relationships to reduce redundancy.Source: §2 · p. 3

General active-learning surveys

  • Aggarwal et al. group strategies as heterogeneity-, performance-, or representativeness-based, while also covering non-standard active-learning variations.Their categories include prediction uncertainty or dissimilarity, predicted loss change, and distribution density.Source: §2 · p. 3

NLP-focused active-learning surveys

  • Olsson focuses on disagreement-based query strategies and practical choices such as seed-set selection, scenario choice, and stopping.Source: §2 · p. 3

Survey gap

  • Earlier surveys mention some neural-network applications but do not cover NN-based active learning in depth or incorporate recent architectures and language-model advances.The section names KimCNN [48], word embeddings, and contextualized language models among the developments not adequately covered.Source: §2 · p. 3

3 Active Learning

Active learning (AL) iteratively selects unlabeled instances for expert labeling, while query strategies determine which instances to request. The survey organizes these strategies by input information and examines neural-network (NN) uncertainty and small-data challenges.

3 Active Learning

  • The oracle labels batches selected by the active learner, which retrains its model after each update until a stopping criterion is met.Retraining makes each update at least as expensive as training the underlying model.Source: §3 · p. 3

3.1 Query Strategies

  • Data-based strategies use raw inputs and optionally labeled-pool labels, targeting data uncertainty or representative subsets.Representativeness strategies geometrically compress a set into fewer points that reflect its properties.Source: §3.1 · p. 4
  • Model-based strategies score model-provided quantities other than final predictions, while prediction-based strategies score prediction outputs, especially uncertainty or disagreement.The distinction can be thin; prediction-based uncertainty roughly corresponds to inter-class uncertainty, and model-based uncertainty to intra-class uncertainty.Source: §3.1 · p. 5
  • Ensembles combine primitive strategies and can mix taxonomy classes, so the survey classifies the ensemble concept rather than detailing each hybrid.The survey notes that disagreement among classifiers has been covered in prior surveys.Source: §3.1 · p. 5
  • The taxonomy classifies query strategies by the input information they use and highlights where uncertainty enters; strategies using multiple information types are assigned to the most specific class.Prediction-based takes precedence over model-based, which takes precedence over data-based.Source: §3.1 · pp. 3–5

3.2 Neural-Network-Based Active Learning

  • NN-based AL research addresses uncertainty estimation and the tension between DNNs’ data needs and AL’s small labeled datasets.The survey describes NN overconfidence and the limited feasibility of ensemble and Bayesian approaches on larger datasets, while small-data training risks overfitting and weak generalization.Source: §3.2 · pp. 5–6
  • Prior approaches include expected-gradient and embedding-change selection, core-set sampling, and BADGE, while reported GAN-based synthesis is usually less effective for NLP because of artificial-text quality.BADGE uses k-means++ seeding on final-layer gradients to query by uncertainty and diversity.Source: §3.2 · p. 6

4 Active Learning for Text Classification

The survey connects recent text-classification advances with AL research, then reviews experiments to identify common practices and limitations. It finds prediction-based strategies prevalent, but comparisons across studies and the transfer of findings to modern NNs remain uncertain.

4.1 Recent Advances in Text Classification

  • Recent text representations progress from static word embeddings to context-dependent language-model vectors, with BERT-style pretraining and fine-tuning and alternative approaches such as XLNet.Contextual representations handle word meaning in context, unlike static embeddings; the survey describes XLNet as avoiding BERT’s masking approach.Source: §4.1 · pp. 6–7
  • XLNet has a slight edge over BERT in test error rate, ULMFiT follows closely, and KimCNN remains a strong contender in NN text classification.ULMFiT also outperformed prior state of the art on several text-classification datasets when trained with 100 labeled examples.Source: §4.1 · p. 7

4.2 Text Classification for Active Learning

  • Modern AL text-classification studies use CNNs, LSTMs, GRUs, transformer representations, and FastText.zip, with reported advantages varying by model and setting.Transformer representations yielded consistently higher scores as SVM input features; LSTM and GRU significantly outperformed an SVM baseline on ThucNews.Source: §4.2 · p. 7
  • FastText.zip with entropy-based uncertainty sampling was reported as a strong, computationally cheap baseline, and its 5%–40% surrogate datasets supported ULMFiT accuracy close to full-data training.The authors of that study described the strategy as effective and robust.Source: §4.2 · pp. 7–8

4.3 Commonalities and Limitations of Previous Experiments

  • A significant majority of surveyed query strategies are prediction-based, particularly prediction-uncertainty and disagreement strategies.The survey’s experiment overview organizes publications by datasets, models, and query-strategy classes.Source: §4.3 · p. 8
  • Limited dataset overlap weakens comparisons across studies, and the absence of classic datasets in the large-scale study leaves its results difficult to relate to prior work.The survey also notes that it remains unclear whether past findings generalize to DNNs or whether modern NNs benefit from the same query strategies.Source: §4.3 · pp. 8–9

5 Open Research Questions

The survey identifies unresolved challenges in uncertainty estimation, representation evaluation, small-data fine-tuning, and comparable experiments, alongside the difficulty of selecting an effective query strategy.

Uncertainty Estimates in Neural Networks

  • Uncertainty-based strategies remain widely used, but neural networks still face inaccurate uncertainty estimates or limited scalability.The survey identifies these as challenges for uncertainty-based selection.Source: §5 · p. 9

Representations

  • No AL-specific systematic evaluation compares word embeddings and language models with neural networks, and both remain scarcely used.The survey notes that sparse adoption may reflect slow uptake or practical issues that have not been investigated.Source: §5 · p. 9

Small Data DNNs

  • Pre-trained language models and fine-tuning alleviate the small-data problem to some degree, but the minimum data needed for successful fine-tuning remains unknown.The survey frames this as an open question for DNN-based AL.Source: §5 · p. 9

Comparable Evaluations

  • Experiments often use disjoint dataset combinations, reducing or eliminating comparability between recent and past work.Comparability is needed to verify whether insights from shallow neural-network AL apply to DNN-based AL.Source: §5 · p. 9

Learning to Learn

  • Because the optimal query strategy depends on data, model, task, and iteration, meta-learning may learn a selection policy or query strategy.The survey cites approaches that learn optimal selection or query strategies as a whole [6] [49].Source: §5 · p. 9

6 Conclusions

The survey organizes query strategies and reviews DNN-based active learning for text classification, finding that uncertainty-based strategies remain prevalent while adoption and evaluation remain limited. It identifies future research directions for connecting recent NLP advances to active learning.

6 Conclusions

  • The survey classifies query strategies by whether instance selection relies on data-based, model-based, or prediction-based information.It categorizes strategies used in active learning for text classification within this taxonomy.Source: §6 · p. 9
  • Uncertainty-based query strategies are the most widely used class, including when analysis is restricted to neural networks.Source: §6 · p. 9
  • Language-model representations offer finer-grained context-specific representations and handle out-of-vocabulary words, while their adoption in active learning remains limited.The survey reports that fine-tuning-based transfer learning alleviates the small-data problem to some degree but also lacks adoption.Source: §6 · p. 9
  • The surveyed DNN-based active learning research is underresearched and often lacks comparability, despite promising results from early adoptions [104] [90].Source: §6 · p. 9
  • The survey argues that DNN adoption could improve classification with the same data or reduce labeling effort while maintaining performance.It identifies research directions intended to advance DNN-based active learning.Source: §6 · p. 9

A.1 Datasets

The appendix catalogs widely used text-classification datasets and describes their classification types and training and test set sizes.

A.1 Datasets

  • The dataset table lists short IDs and names alongside classification type and training and test set sizes.When no predefined split exists, the training count represents the full dataset size.Source: Table 2 · p. 16
  • The collection includes binary, multi-class, and multi-class multi-label datasets.Examples listed include customer and movie reviews, news, and Reuters datasets.Source: §A.1 · p. 15, Table 2 · p. 16
  • The table notes that 20newsgroups excludes duplicates and that some dataset counts reflect documents or reduced positive/negative labels.Source: §A.1 · p. 16

Related papers

Loading 2008.07267v1…