---
title: "Beyond the Token — paper by Yagiz and Horasan (2026)"
description: "The authors call this mismatch Tokenization-Morphology Misalignment (TMM) and propose MAFEX to fix it. 23 references."
authors: ["Muhammet Anil Yagiz", "Fahrettin Horasan"]
year: 2026
venue: "Proceedings of the Second Workshop Natural Language Processing for Turkic Languages (SIGTURK 2026)"
source: "https://aclanthology.org/2026.sigturk-1.19/"
doi: "10.18653/v1/2026.sigturk-1.19"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/yagiz-2026-beyond-token-correcting-tokenization-bias/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/yagiz-2026-beyond-token-correcting-tokenization-bias/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/25/) / 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

**Authors:** Muhammet Anil Yagiz and Fahrettin Horasan · **Year:** 2026 · **Venue:** [Proceedings of the Second Workshop Natural Language Processing for Turkic Languages (SIGTURK 2026)](https://md.research.elamur.ai/free/en/venues/sigturk/) ([Workshop Natural Language Processing for Turkic Languages 2026](https://md.research.elamur.ai/free/en/venues/sigturk/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.sigturk-1.19/) · [PDF](https://aclanthology.org/2026.sigturk-1.19.pdf) · [DOI 10.18653/v1/2026.sigturk-1.19](https://doi.org/10.18653/v1/2026.sigturk-1.19) · [OpenAlex](https://openalex.org/W7140134996)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## 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 University](https://md.research.elamur.ai/free/en/universities/kirikkale-university/) — University, Kırıkkale, Türkiye · 2 works · 256 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works

🧭 Topics (2): [🧩 Morphology](https://md.research.elamur.ai/free/en/topics/morphology/) · 420 works · [✂️ Tokenization & segmentation](https://md.research.elamur.ai/free/en/topics/tokenization/) · 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.

```bibtex
@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}
}
```

- [← previous Beyond the Singular: Revealing the Value of Multiple Generations in Benchmark Evaluation](https://md.research.elamur.ai/free/en/papers/zhang-2026-beyond-singular-revealing-value-multiple/)
- [next → Beyond Transcription: Unified Audio Schema for Perception-Aware AudioLLMs](https://md.research.elamur.ai/free/en/papers/zhang-2026-beyond-transcription-unified-audio-schema/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/25/)

## 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](https://scholar.google.com/scholar?q=Model+Coverage%3A+We+focus+on+BERT%2D+%5B2%5D+M%2E+Sundararajan%2C+A%2E+Taly%2C+and+Q%2E+Yan%2E+Axiomatic+based+encoders+and+one+decoder+%28Kumru%2Dattribution+for+deep+networks%2E+ICML%2C+2017%2E)
2. [3] S. M. Lundberg and S.-I. Lee. [A unified approach to interpreting model predictions](https://md.research.elamur.ai/free/en/papers/lundberg-2017-unified-approach-interpreting-model-predictions/). NeurIPS, 2017.
3. [23] J. DeYoung, et al. Eraser: A benchmark to evaluate rationales and explanations in nlp. ACL, 2020. [find](https://scholar.google.com/scholar?q=%5B23%5D+J%2E+DeYoung%2C+et+al%2E+Eraser%3A+A+benchmark+to+evaluate+rationales+and+explanations+in+nlp%2E+ACL%2C+2020%2E)
4. [4] S. J. Mielke, et al. [Between words and characters: A brief history of open-vocabulary modeling and tokenization in NLP](https://md.research.elamur.ai/free/en/papers/mielke-2021-between-words-characters-brief-history/). arXiv, 2021.
5. [24] A. Madsen, et al. Post-hoc Interpretability for Neural NLP: A Survey. ACM Computing Surveys, 2021. [find](https://scholar.google.com/scholar?q=%5B24%5D+A%2E+Madsen%2C+et+al%2E+Post%2Dhoc+Interpretability+for+Neural+NLP%3A+A+Survey%2E+ACM+Computing+Surveys%2C+2021%2E)
6. [5] K. Oflazer. [Two-level description of Turkish morphology](https://md.research.elamur.ai/free/en/papers/oflazer-1993-two-level-description-turkish-morphology/). Literary and linguistic computing, 1994.
7. [6] G. Eryiğit and K. Oflazer. Statistical dependency parsing for Turkish. EACL, 2006. [find](https://scholar.google.com/scholar?q=%5B6%5D+G%2E+Eryig%CC%86it+and+K%2E+Oflazer%2E+Statistical+dependency+parsing+for+Turkish%2E+EACL%2C+2006%2E)
8. [7] J. Bastings, et al. The elephant in the interpretability room. BlackboxNLP, 2020. [find](https://scholar.google.com/scholar?q=%5B7%5D+J%2E+Bastings%2C+et+al%2E+The+elephant+in+the+interpretability+room%2E+BlackboxNLP%2C+2020%2E)
9. [8] S. Schweter. BERTurk: BERT models for Turkish. Zenodo, 2020. [find](https://scholar.google.com/scholar?q=%5B8%5D+S%2E+Schweter%2E+BERTurk%3A+BERT+models+for+Turkish%2E+Zenodo%2C+2020%2E)
10. [9] S. Yıldırım. Turkish-base-bert-sentiment-cased. HuggingFace Model Hub, 2020. [find](https://scholar.google.com/scholar?q=%5B9%5D+S%2E+Y%C4%B1ld%C4%B1r%C4%B1m%2E+Turkish%2Dbase%2Dbert%2Dsentiment%2Dcased%2E+HuggingFace+Model+Hub%2C+2020%2E)
11. [10] YTÜ-CE Cosmos. Turkish-base-bert-uncased. HuggingFace Model Hub, 2023. [find](https://scholar.google.com/scholar?q=%5B10%5D+YT%C3%9C%2DCE+Cosmos%2E+Turkish%2Dbase%2Dbert%2Duncased%2E+HuggingFace+Model+Hub%2C+2023%2E)
12. [11] VNGRS-AI. Kumru-2B: A Turkish Decoder Model. HuggingFace Model Hub, 2025. [find](https://scholar.google.com/scholar?q=%5B11%5D+VNGRS%2DAI%2E+Kumru%2D2B%3A+A+Turkish+Decoder+Model%2E+HuggingFace+Model+Hub%2C+2025%2E)
13. [12] P.-J. Kindermans, et al. The (un)reliability of saliency methods. Explainable AI, Springer, 2019. [find](https://scholar.google.com/scholar?q=%5B12%5D+P%2E%2DJ%2E+Kindermans%2C+et+al%2E+The+%28un%29reliability+of+saliency+methods%2E+Explainable+AI%2C+Springer%2C+2019%2E)
14. [13] N. Elhage, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 2021. [find](https://scholar.google.com/scholar?q=%5B13%5D+N%2E+Elhage%2C+et+al%2E+A+mathematical+framework+for+transformer+circuits%2E+Transformer+Circuits+Thread%2C+2021%2E)
15. [14] K. Wang, et al. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. ICLR, 2023. [find](https://scholar.google.com/scholar?q=%5B14%5D+K%2E+Wang%2C+et+al%2E+Interpretability+in+the+wild%3A+a+circuit+for+indirect+object+identification+in+GPT%2D2+small%2E+ICLR%2C+2023%2E)
16. [15] K. Bostrom and G. Durrett. [Byte Pair Encoding is Suboptimal for Language Model Pretraining](https://md.research.elamur.ai/free/en/papers/bostrom-2020-byte-pair-encoding-is-suboptimal/). EMNLP, 2020.
17. [16] L. Xue, et al. ByT5: Towards a token-free future with pre-trained byte-to-byte models. TACL, 2022. [find](https://scholar.google.com/scholar?q=%5B16%5D+L%2E+Xue%2C+et+al%2E+ByT5%3A+Towards+a+token%2Dfree+future+with+pre%2Dtrained+byte%2Dto%2Dbyte+models%2E+TACL%2C+2022%2E)
18. [17] J. Vig, et al. Investigating gender bias in BERT’s attention heads. NeurIPS, 2020. [find](https://scholar.google.com/scholar?q=%5B17%5D+J%2E+Vig%2C+et+al%2E+Investigating+gender+bias+in+BERT%E2%80%99s+attention+heads%2E+NeurIPS%2C+2020%2E)
19. [18] A. Geiger, et al. Causal abstractions of neural networks. NeurIPS, 2021. [find](https://scholar.google.com/scholar?q=%5B18%5D+A%2E+Geiger%2C+et+al%2E+Causal+abstractions+of+neural+networks%2E+NeurIPS%2C+2021%2E)
20. [19] A. A. Akın and M. D. Akın. Zemberek, an open source nlp framework for turkic languages. Structure, 2007. [find](https://scholar.google.com/scholar?q=%5B19%5D+A%2E+A%2E+Ak%C4%B1n+and+M%2E+D%2E+Ak%C4%B1n%2E+Zemberek%2C+an+open+source+nlp+framework+for+turkic+languages%2E+Structure%2C+2007%2E)
21. [20] A. Ustun, et al. Aya Model: An Instruction Fine-tuned Open-Access Multilingual Language Model. arXiv, 2024. [find](https://scholar.google.com/scholar?q=%5B20%5D+A%2E+Ustun%2C+et+al%2E+Aya+Model%3A+An+Instruction+Fine%2Dtuned+Open%2DAccess+Multilingual+Language+Model%2E+arXiv%2C+2024%2E)
22. [21] A. Vaswani, et al. NeurIPS, 2017. [find](https://scholar.google.com/scholar?q=%5B21%5D+A%2E+Vaswani%2C+et+al%2E+NeurIPS%2C+2017%2E)
23. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). [22] J. Devlin, et al. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL, 2019.

*Updated 2 November 2026*