ActiveTigger ![]()
An open source collaborative text annotation software for computational social sciences
É. Schultz, J. Boelaert, A. Morin, A. Claesson, E. Bonutti D’Agostini, A. Chatelain, É. Ollion
CREST – Groupe ENSAE-ENSAI – Institut Polytechnique de Paris – CSS@IPP · JADT 2026, Palermo
Introducing ActiveTigger
An open source tool to collaboratively annotate text corpora and train classifiers with active learning to speed up annotation.
Prototype ~2020 from research practices (J. Boelaert & É. Ollion) \(\rightarrow\) refacto & public v1.0.0 released in May 2026
- Document-level annotation of large corpora of short texts
- Sensible defaults first, customization later – avoid overwhelming non-experts
- End-to-end integration in a single platform
What are the main use cases ?
- Annotate quickly a large corpus – fine-tune models from the interface, use active learning to accelerate
- Collaboratively stabilize a codebook – multiple users + inter-annotator agreement
- Identify and retrieve documents of interest – BERTopic, regex filters, projections
- Teach NLP – hands-on experience of a full supervised pipeline, no code required
Demo
How prominent is gender-related research in French social science publications?
Main features (1): annotate & represent
- Projects: CSV / Parquet / XLSX import, stratified train/validation/test split, shared across users
- Annotation: multiclass & multilabel schemes, integrated codebook, history, filters (label, user, regex, prediction), comments
- Text representations: sentence-transformers embeddings, fastText, DFM, regex features
- Exploration: tabular view, 2D projections (UMAP / t-SNE), BERTopic topic models
Main features (2): train, evaluate, predict
- Selection strategies: random, sequential, and active learning (max entropy, target-label probability…)
- Two families of classifiers:
- Quick models: scikit-learn classifiers on features, cross-validated, trained in seconds
- BERT models: Hugging Face fine-tuning, key hyperparameters exposed, loss curves
- Evaluation: precision / recall / F1 per class, macro F1, confusion matrices, misclassified examples – computed separately on train / validation / test
- Inference & export: predict on the full dataset or external files; export annotations, features, models, predictions
- (Experimental) generative panel: prompt external LLM APIs, compare with human annotations
Does it hold up? Validation
Reproducibility – fine-tuning benchmark on a published stance-detection dataset (Luo et al. 2021, global warming, 3 labels):
- Original BERT-base baseline: accuracy 71% [0.64–0.77]
- ActiveTigger, same parameters: accuracy 0.69–0.72, within the reported confidence interval
- The dataset ships with ActiveTigger so anyone can reproduce
Robustness – stress tests on a fresh cloud GPU server (13 min install):
- 10 simultaneous projects with BERT trainings: queue handled correctly, delay < 0.8s
- 10 users annotating the same project: all operations OK, delay < 0.4s
Limitations & roadmap
Current limitations for an instance
- Single-GPU deployments, a few dozen users
- Exports don’t yet capture full project state
Roadmap
- Task management refactoring \(\rightarrow\) multi-GPU, larger communities
- Stable evaluation framework (reference datasets, versioned benchmarks)
- Generative panel co-designed with the social science community
- Full project archiving for re-executable annotation campaigns
Takeaways
- ActiveTigger = end-to-end, collaborative, open source annotation platform for social sciences
- Encodes supervised-learning best practices in the interface
- Active learning as the bridge between human annotation and automation
- Validated, in production, and driven by a research community
Try it: request an account or deploy your own instance - emilien.schultz@ensae.fr
