Morpheus: A Morphology-Aware Neural Tokenizer and Word Embedder for Turkish

In three linesMorpheus is a morphology-aware neural tokenizer for agglutinative Turkish. The model uses differentiable Poisson-binomial dynamic programming to segment morphemes with 1.425 bits-per-character compression and MorphScore macro-F1 of 0.61 (vs ~0.32 for subword tokenizers). Lossless by construction: decode(encode(w)) = w.

Summary generated by Claude — human-verified