Reproduce “AutoIntent: AutoML for Text Classification” on OpenResearch

About this paper

Ilya Alekseev, Roman Solomatin, Darina Rustamova, Denis Kuznetsov

AutoIntent, developed by researchers from the Moscow Center for Advanced Studies and affiliated universities, provides an end-to-end automated machine learning framework for text classification, particularly for intent detection. It achieves competitive accuracy on standard datasets while significantly reducing training time and demonstrates superior performance in out-of-scope detection and few-shot learning scenarios.

Key insights

• Hierarchical optimization across distinct Embedding, Scoring, and Decision modules effectively manages the complexity of NLP AutoML, allowing for efficient exploration of the search space without combinatorial explosion.
• Separating the scoring and decision-making processes enables flexible strategies for multi-label classification and dedicated confidence threshold tuning for robust out-of-scope detection.
• Leveraging pre-computed transformer embeddings combined with a diverse array of downstream classifiers allows for robust performance in few-shot learning scenarios and offers a spectrum of computational efficiencies.

Summary by alphaXiv.