Papers
Topics
Authors
Recent
Search
2000 character limit reached

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

Published 17 Jun 2026 in cs.CL and cs.AI | (2606.18717v1)

Abstract: Turkish is agglutinative: meaning is carried by morphemes, yet the subword tokenizers that drive modern LLMs split words by corpus statistics, fragmenting semantically loaded suffixes and -- in the case of WordPiece and rule-based analyzers -- failing to decode their output back to the original text. This paper presents \textbf{Morpheus}, a neural morpheme-boundary model for Turkish that is at once a lossless, morphology-aware tokenizer and a word-embedding producer. A differentiable Poisson-binomial dynamic program turns per-character boundary probabilities into soft morpheme memberships during training and exact segments at inference, with no string normalization, so $\mathrm{decode}(\mathrm{encode}(w)) = w$ holds by construction. Because the model is neural, the same forward pass that tokenizes also emits a structured word embedding. Among reversible tokenizers -- the only ones valid for generation -- Morpheus attains the lowest bits-per-character ($1.425$), roughly doubles the gold morphological alignment of the subword family (MorphScore macro-F1 $0.61$ vs.\ ${\sim}0.32$), and uses ${\sim}19\%$ less GPU memory than 64K-vocabulary subword tokenizers. As an embedder, frozen Morpheus vectors lead on lexical retrieval (root-family MAP $0.85$) and same-root verification (ROC-AUC $1.00$), surpassing the multilingual retriever BGE-M3 and BERTurk; on context- and inflection-dependent tasks (NER, case/number probing) the heavier contextual encoders remain ahead -- a trade-off we attribute to Morpheus's root-centric geometry. Code: https://github.com/lonewolf-rd/TurkishMorpheus; model: https://huggingface.co/lonewolflab/Morpheus-TR-50K; interactive demo: https://huggingface.co/spaces/lonewolflab/morpheus-tr-demo.

Authors (1)

Summary

  • The paper introduces Morpheus, a neural model that unifies lossless morphology-aware tokenization with root-level word embeddings for Turkish.
  • It leverages a deep, differentiable architecture using multi-scale convolutions and self-attention to achieve 100% reversible tokenization and robust morphological alignment.
  • Empirical results demonstrate improved tokenization fidelity, efficient GPU usage, and superior lexical retrieval compared to traditional subword and rule-based methods.

Morpheus: Morphology-Aware Neural Tokenization and Embedding for Turkish

Motivation and Context

Turkish's agglutinative morphology introduces complexity for NLP systems: single roots generate hundreds of distinct surface forms via ordered morphemic affixation. For modern LMs, tokenization and embedding are bottlenecks. Corpus-driven subword models such as BPE, WordPiece, and Unigram fragment semantically loaded morphemes and often fail reversibility, either through lossy normalization or diacritic stripping. Rule- and dictionary-based systems such as Zemberek or TurkishTokenizer offer improved alignment but either lack scalability, fail to decode surface forms, or remain dictionary-bound. Meanwhile, standard representation learning decouples tokenization (meaningless token IDs) and embedding (large, context-heavy encoders), which duplicates the modeling of morphological signals.

Methodology

Morpheus unifies three functionalities in a standalone neural model:

  • Lossless Morphology-Aware Tokenization: Given a Turkish word as a character sequence, Morpheus predicts morpheme boundaries using a deep, surface-preserving differentiable mechanism, ensuring decode(encode(ww)) = ww by construction.
  • Structured, Morpheme-Derived Embedding: The same forward pass produces an embedding per word, organized at the root-family level.
  • Neural Differentiability and OOV Generalization: The pipeline, trained end-to-end, accommodates any Turkish string, including nonce forms and rare surface fusions.

The segmentation process centers on a Poisson-binomial dynamic program that softly assigns character memberships to morphemes given per-character boundary probabilities. The mechanism supports end-to-end gradient flow so the boundary detector is directly shaped by both supervised (Morfessor-derived) labels and various distributional objectives (skip-gram, contrastive root-family learning, character-level MLM). The architecture employs multi-scale convolutions, RoPE-augmented self-attention, and segment-wise attention pooling in a 320-dimensional latent space.

Data are sourced from carefully cleaned, register-diverse Turkish corpora (Wikipedia, news, academic, colloquial). Preprocessing emphasizes strict surface preservation and comprehensive coverage of rare morphology.

Key Results and Numerical Evaluation

Tokenization Integrity

  • Reversibility: Morpheus achieves 100% roundtrip reconstruction over >30,000 inflected wordforms, matching corpus-based BPE/Unigram/Byte-level methods. Both TurkishTokenizer (95.4%) and WordPiece (58.2%) are non-reversible due to canonicalization and diacritic stripping, rendering them invalid for generative use.
  • Surface Fidelity: In curated OOV evaluation, Morpheus exhibits 0% corruption when matching segment boundaries to surface forms; unlike rule-based systems, no morph string is lost or rewritten.
  • Morphological Alignment: Morpheus attains a MorphScore macro-F1 of 0.61 (UD_Turkish-Kenet), nearly double the subword baselines, and comparable to TurkishTokenizer (0.65), but with perfect surface string fidelity (no normalization-induced mismatch).

Downstream Language Modeling and Efficiency

  • Bits-per-Character (BPC): On parameter-equalized 58M GPT LLMs, Morpheus yields the lowest BPC (1.425) among reversible tokenizers, slightly surpassing BPE and Unigram.
  • Token Fertility: Morpheus emits 1.73 tokens per word, higher than subword schemes (~1.5). This deliberate trade-off optimizes morphological faithfulness at the expense of sequence length.
  • GPU Memory: Morpheus demands ~19% less GPU memory during autoregressive generation than 64K-vocab subword configurations.

Embedding Quality

  • Root-Level Retrieval and Deduplication: Frozen Morpheus embeddings outperform both BERTurk and BGE-M3 in lexical retrieval (root-family MAP 0.85 vs. 0.80/0.49) and same-root verification (ROC-AUC 1.00 vs. 0.98/0.70).
  • Context/Inflection-Sensitive Tasks: Morpheus trails heavily contextual encoders in number/case morphological probing (0.59/0.22 vs. BERTurk's 0.95/0.89) and NER (macro-F1 0.48 vs. 0.79). The explicit root-centric training objective explains this: variance across inflectional features is collapsed, and lack of contextualization limits token-level entity resolution.

Implications and Discussion

Practical Impact

Morpheus provides a single, easily deployable component to replace multi-stage Turkish text pipelines:

  • For text generation, it is one of the few reversible and morphology-aligned choices, critical for high-fidelity open-ended generation tasks.
  • In lexical retrieval, deduplication, and dictionary-style applications, its root-clustered embeddings and lossless segmentation offer significant efficiency and faithfulness gains.
  • The architecture lowers inference memory and features fast encoding/decoding throughput, narrowing practical gaps with established baselines.

Theoretical Insights

The integration of morph-boundary detection and embedding in one mechanism exposes an information-theoretical advantage in agglutinative language modeling. Coupling segmentation with representation allows for mutual reinforcement—morpheme boundary learning and word geometry are both optimized under end-task supervision, rather than separately and potentially adversarially.

Trade-Offs

Morpheus's uniform treatment of OOV/nonce forms via a neural model requires shipping a neural artifact (PyTorch checkpoint) rather than a static vocab table. Higher token counts modestly impact autoregressive speed. Its embedding geometry, favoring the root level, makes it ill-suited to tasks requiring fine-grained inflection or context comprehension, such as high-accuracy NER.

Future Directions

  • Extending Morpheus's method to other agglutinative or morphologically rich low-resource languages could universalize its architecture.
  • Hybrid or hierarchical contextualization strategies may bridge the root-level geometry with deeper, inflectionally nuanced representations.
  • Investigating integration with RAG pipelines: use Morpheus for the lexical/keyword retrieval index and dense contextual encoders for semantic resolution.

Conclusion

Morpheus addresses longstanding deficiencies in Turkish NLP by providing a lossless, morphology-aware tokenizer that simultaneously delivers high-quality, root-structured embeddings in a single neural architecture. Among reversible tokenizers, Morpheus achieves the strongest scores in both intrinsic (BPC, morphological F1) and lexical downstream metrics, at a measured, explicit cost in sequence length and dense contextualization. Its deployment streamlines Turkish text pipelines for both generative LMs and information retrieval, offering a robust, theoretically principled, and practically validated alternative to frequency-driven or rule-based approaches. The decoupling of lossless morph segmentation from surface-destructive normalization measurably advances the state of Turkish tokenization and word representation.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.