Reproduce “MorphNAS: Differentiable Architecture Search for Morphologically-Aware Multilingual NER” on OpenResearch

About this paper

Prathamesh Devadiga, Omkaar Jayadev Shetty, Hiya Nachnani, Prema R

Researchers at PES University, Bengaluru, India, developed MorphNAS, a differentiable neural architecture search framework specifically for Named Entity Recognition in morphologically rich languages. The framework enhances DARTS by integrating morphologically-aware preprocessing to guide the search, achieving a weighted F1-score of 0.9331 on Kannada test data and a 0.9380 validation F1-score during Hindi search.

Key insights

• Incorporating linguistically-aware preprocessing directly into a differentiable architecture search framework can effectively guide the discovery of architectures optimized for morphologically rich languages.
• Automating the design of neural architectures through a DARTS-based approach provides an efficient alternative to manual model adaptation for diverse and complex linguistic environments.
• The bi-level optimization in DARTS, concurrently optimizing network weights and architectural parameters, streamlines the exploration of specialized micro-architectural elements for specific NLP tasks.

Summary by alphaXiv.