Elamur Research⌕
☰
home / papers / 2026 / Beyond the Token: Correcting the Tokenization Bias in XAI via Morphologically-Aligned Projection

Beyond the Token: Correcting the Tokenization Bias in XAI via Morphologically-Aligned Projection

Paper · 2026 · Muhammet Anil Yagiz and Fahrettin Horasan

Summary

Standard attribution methods such as Integrated Gradients, SHAP and DeepLIFT score subword tokens. In agglutinative languages like Turkish, those tokens rarely match morphemes, so credit for a grammatical meaning such as negation gets spread over arbitrary fragments. The authors call this mismatch Tokenization-Morphology Misalignment (TMM) and propose MAFEX to fix it.

MAFEX works in three steps:

  • It segments the input with Zemberek.
  • It builds a binary matrix linking tokens to morphemes and uses it to sum IG scores (50 steps) into morpheme-level scores. The authors prove this linear projection preserves the Completeness axiom.
  • It blends in a causal score, obtained by masking each morpheme, with weight λ = 0.7 on the gradient part.

Evaluation uses TRUST-TR, a hand-built diagnostic set of 20 Turkish sentences, each labeled with one decisive morpheme, across four Turkish models. MAFEX ranks that morpheme first in 91.25% of cases on average. This is far above token-level baselines and a random-grouping control. The result suggests that explanations for morphologically rich languages should be built from linguistic units rather than tokenizer artifacts.

Key points

  • Models tested: BERTurk (dbmdz/bert-base-turkish-cased), BERTurk-Sentiment (savasy/bert-base-turkish-sentiment-cased), Cosmos-BERT (ytu-ce-cosmos/turkish-base-bert-uncased) and the Mistral-based decoder Kumru-2B (vngrs-ai/Kumru-2B).
  • Average F1@1 (top-ranked morpheme equals the gold key morpheme): MAFEX 91.25% (85–95% per model), token IG 13.75%, SHAP and DeepLIFT 8.75% each, Random Grouping 55%. The conclusion cites 16.25% for token-level methods, which is inconsistent with the abstract's 13.75%.
  • The Random Grouping control merges tokens into groups of morpheme-like sizes but with random boundaries. MAFEX beats it by 33.75 points, which supports the claim that the gain comes from true morpheme boundaries rather than mere aggregation.
  • Example: for Gelemedim ("I could not come"), token IG highlights the root Gel, while MAFEX isolates the negation/inability suffix -eme.
  • The benchmark is very small (N = 20), which the authors acknowledge as a limitation. MAFEX is released as an open-source Python package (pip install mafex, github.com/anilyagiz/mafex).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Kırıkkale UniversityKırıkkale · Türkiye2 works · 256 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 2Research subjects and languages the work is about.Morphology · 420 worksTokenization & segmentation · 124 works

Cite

Cite as Muhammet Anil Yagiz and Fahrettin Horasan. 2026. Beyond the Token: Correcting the Tokenization Bias in XAI via Morphologically-Aligned Projection. Proceedings of the Second Workshop Natural Language Processing for Turkic Languages (SIGTURK 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.sigturk-1.19

@inproceedings{yagiz2026beyond,
  title = {{Beyond the Token: Correcting the Tokenization Bias in XAI via Morphologically-Aligned Projection}},
  author = {Muhammet Anil Yagiz and Fahrettin Horasan},
  booktitle = {Proceedings of the Second Workshop Natural Language Processing for Turkic Languages (SIGTURK 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.sigturk-1.19},
  url = {https://doi.org/10.18653/v1/2026.sigturk-1.19}
}

References

23 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Model Coverage: We focus on BERT- [2] M. Sundararajan, A. Taly, and Q. Yan. Axiomatic based encoders and one decoder (Kumru-attribution for deep networks. ICML, 2017. find
  2. [3] S. M. Lundberg and S.-I. Lee. A unified approach to interpreting model predictions. NeurIPS, 2017.
  3. [23] J. DeYoung, et al. Eraser: A benchmark to evaluate rationales and explanations in nlp. ACL, 2020. find
  4. [4] S. J. Mielke, et al. Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP. arXiv, 2021.
  5. [24] A. Madsen, et al. Post-hoc Interpretability for Neural NLP: A Survey. ACM Computing Surveys, 2021. find
  6. [5] K. Oflazer. Two-level description of Turkish morphology. Literary and linguistic computing, 1994.
  7. [6] G. Eryiğit and K. Oflazer. Statistical dependency parsing for Turkish. EACL, 2006. find
  8. [7] J. Bastings, et al. The elephant in the interpretability room. BlackboxNLP, 2020. find
  9. [8] S. Schweter. BERTurk: BERT models for Turkish. Zenodo, 2020. find
  10. [9] S. Yıldırım. Turkish-base-bert-sentiment-cased. HuggingFace Model Hub, 2020. find
  11. [10] YTÜ-CE Cosmos. Turkish-base-bert-uncased. HuggingFace Model Hub, 2023. find
  12. [11] VNGRS-AI. Kumru-2B: A Turkish Decoder Model. HuggingFace Model Hub, 2025. find
  13. [12] P.-J. Kindermans, et al. The (un)reliability of saliency methods. Explainable AI, Springer, 2019. find
  14. [13] N. Elhage, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 2021. find
  15. [14] K. Wang, et al. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. ICLR, 2023. find
  16. [15] K. Bostrom and G. Durrett. Byte Pair Encoding is Suboptimal for Language Model Pretraining. EMNLP, 2020.
  17. [16] L. Xue, et al. ByT5: Towards a token-free future with pre-trained byte-to-byte models. TACL, 2022. find
  18. [17] J. Vig, et al. Investigating gender bias in BERT’s attention heads. NeurIPS, 2020. find
  19. [18] A. Geiger, et al. Causal abstractions of neural networks. NeurIPS, 2021. find
  20. [19] A. A. Akın and M. D. Akın. Zemberek, an open source nlp framework for turkic languages. Structure, 2007. find
  21. [20] A. Ustun, et al. Aya Model: An Instruction Fine-tuned Open-Access Multilingual Language Model. arXiv, 2024. find
  22. [21] A. Vaswani, et al. NeurIPS, 2017. find
  23. Attention is all you need. [22] J. Devlin, et al. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL, 2019.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yagiz-2026-beyond-token-correcting-tokenization-bias/