Abstract
Standard Active Learning (AL) often saves money but can accidentally bake social biases into AI models by ignoring underrepresented groups. We introduce D-CALM (Dynamic Clustering-based Active Learning for Mitigating Bias), a framework that solves this by dynamically re-clustering data and focusing annotation efforts on regions where the model is currently struggling the most. This approach leads to better performance and fairer representation across diverse text classification tasks.
Introduction: The Bias Induction Trap
Active Learning is a popular strategy for training machine learning models when labeled data is scarce. Instead of labeling everything, the model "asks" for labels on the examples it finds most confusing. However, this creates a "bias induction" trap: if a model is "confidently wrong" about a minority group (e.g., misclassifying hate speech against a specific demographic because it hasn't seen enough examples), it may never realize it needs more data from that group. Generic AL methods often over-sample from majority groups or specific "hard" regions while completely neglecting others, leading to models that perform poorly on marginalized populations. D-CALM aims to bridge the gap between annotation efficiency and socio-cultural fairness.
Methodology: The D-CALM Framework
The authors propose D-CALM, a framework that combines uncertainty sampling with dynamic clustering to ensure diverse and representative data selection.
The D-CALM loop operates through several key stages:
- Vectorization: Unlabeled text is converted into embeddings using SentenceBERT or Doc2Vec.
- Initial Clustering: The data is partitioned into clusters using KMeans.
- Error Estimation: The model’s accuracy is estimated for each cluster using a small development set.
- Dynamic Allocation: Instead of picking the "top N" most uncertain samples globally, D-CALM allocates the annotation budget to clusters proportional to their error rate. If the model is failing in a specific cluster, it receives more annotation "slots."
- Sub-clustering and Diverse Sampling: Each cluster is further split into sub-clusters, and the most informative (uncertain) sample is picked from each. This prevents "batch redundancy" and ensures the model learns from a wide variety of difficult examples.
Crucially, the clustering and allocation are dynamic—they update every iteration as the model learns and its error surface changes.
Experimental Setup
The researchers evaluated D-CALM on eight datasets covering a wide range of NLP tasks, including hate speech (CONAN, HATE), emotion detection (CARER), linguistic acceptability (CoLA), and general classification (BOOK32, Q-Type). These datasets were chosen to test the model's ability to handle both standard classification challenges and sensitive socio-cultural contexts.
Key Results and Insights
D-CALM consistently outperformed standard AL (TopN) and Random sampling across all eight datasets. In many cases, it achieved higher Macro-F1 scores with fewer samples than competing methods.
Diversity as an Antidote to Bias
One of the most significant findings involves the CONAN dataset (hate speech). Standard AL almost entirely ignored "Persons of Color" (POC) and "Disabled" categories because they were underrepresented. D-CALM doubled the representation of these groups in the training set compared to standard AL, significantly reducing the error rate for these minority groups.
Model Agnosticism
D-CALM works equally well for heavy-duty Transformers (BERT) and traditional models like SVMs. This suggests the benefit comes from the data selection strategy itself, rather than specific model architectures. By forcing the model to sample from high-error regions, D-CALM prevents it from becoming "obsessed" with a specific type of error.
Conclusion and Future Work
While D-CALM offers a robust solution for mitigating bias in active learning, it does come with increased computational costs due to the need for re-clustering in every iteration. Furthermore, its success depends on the quality of initial vector representations. Future work may involve adapting D-CALM for natural language generation tasks, such as counterspeech generation, where maintaining diversity is even more critical.
Summary updated