simon-clmtd/neural-transducer-inflection

Neural Transducer Inflection Models

This repository contains fine-tuned neural string transducer models (neural_transducer / package trans) for morphological inflection generation, trained with imitation learning on the CoNLL-SIGMORPHON 2017 Shared Task 1 benchmark.

Repository Contents

Each subfolder contains a complete, self-contained model bundle:

• eng_100/ & eng_1000/: English inflection (100 vs. 1,000 examples)
• deu_100/ & deu_1000/: German inflection (100 vs. 1,000 examples)
• fra_100/ & fra_1000/: French inflection (100 vs. 1,000 examples)
• ita_100/ & ita_1000/: Italian inflection (100 vs. 1,000 examples)

Each bundle directory includes:

• best.model: PyTorch state dict checkpoint
• best.model.json: Complete training parameters and architecture metadata
• vocabulary.pkl: UniMorph character and feature vocabularies
• sed.pkl & sed.pkl.json: Stochastic Edit Distance parameters and metadata
• bundle_manifest.json: Verification hashes and benchmark accuracy scores

Training Setup

• Architecture: Transition-based neural string transducer (Makarov & Clematide, 2020)
• Encoder: 1-layer BiLSTM (hidden dimension 200, output locked dropout 0.3)
• Decoder: 1-layer LSTM (hidden dimension 200)
• Optimizer: AdamW (learning rate 0.001) with ReduceLROnPlateau scheduler
• Seed: 42

Interactive Space Demo

Try the interactive demo space: simon-clmtd/morphological-inflection