---
title: "ThinkStruct: RST-Aware Attention for Logical — Tran 2026"
description: "ThinkStruct targets logical-reasoning questions in machine reading comprehension."
authors: ["Nhi Thao Tran", "Tien Le", "Toan Pham", "Quoc Hoang Vu"]
year: 2026
venue: "Proceedings of the 30th Conference on Computational Natural Language Learning"
source: "https://aclanthology.org/2026.conll-main.27/"
doi: "10.18653/v1/2026.conll-main.27"
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/tran-2026-thinkstruct-rst-aware-attention-logical/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/tran-2026-thinkstruct-rst-aware-attention-logical/)

[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/10/) / ThinkStruct: RST-Aware Attention for Logical Reasoning in Machine Reading Comprehension

# ThinkStruct: RST-Aware Attention for Logical Reasoning in Machine Reading Comprehension

**Authors:** Nhi Thao Tran, Tien Le, Toan Pham, Quoc Hoang Vu · **Year:** 2026 · **Venue:** Proceedings of the 30th Conference on Computational Natural Language Learning · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.conll-main.27/) · [PDF](https://aclanthology.org/2026.conll-main.27.pdf) · [DOI 10.18653/v1/2026.conll-main.27](https://doi.org/10.18653/v1/2026.conll-main.27) · [OpenAlex](https://openalex.org/W7166847640)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14)

## Summary
ThinkStruct targets logical-reasoning questions in machine reading comprehension. Earlier graph-based systems split passages into logical units using fixed connective lists, which produces segmentation errors and gives inconsistent answers when the same reasoning is worded differently. The authors instead run an RST parser (Chistova, 2024) over the context and each answer option. This yields Elementary Discourse Units (EDUs) plus a rhetorical tree with nucleus–satellite roles and relation labels such as Condition and Cause-Effect. They turn the tree into an RST-enhanced graph with extra edges and inject its adjacency matrix into the attention of a transformer built on RoBERTa. A contrastive module creates negative samples by flipping directional relations, so the model learns which way a logical link points.

On LogiQA, the model beats the encoder-based baselines by a clear margin. On ReClor the results are mixed, and the contrastive component actually lowers scores there. The work shows that hierarchical discourse structure can support multi-hop reasoning without external augmented data.

## Key points
- **LogiQA test accuracy:** ThinkStruct scores 44.09%, compared with 40.09% for LogiDRS, 38.90% for MERIt and 40.71% for AdaLoGN (a 3.38-point gain over AdaLoGN).
- **ReClor test accuracy:** the full model reaches 60.00%, roughly matching AdaLoGN (60.20%) and edging past DAGN (58.20%). It still trails LogiDRS (60.80%).
- **Ablation on contrastive learning:**
  - On LogiQA, it raises test accuracy from 39.48% to 44.09%.
  - On ReClor, the variant without it does better: 69.18% dev and 61.80% test, versus 61.92% and 60.00% with it.
- **Why the ablation diverges:** ReClor produces sparse RST trees with few asymmetric edges, so valid hard-negative pairs often cannot be built. Its questions also tend to be solvable from surface cues.
- **Comparison with generative LLMs:** these models (LLaMA2/3, Mistral-7B, GPT-3.5, GPT-4) are reported alongside the encoder baselines. GPT-4 scores well above all encoder-based systems.

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

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{tran2026thinkstruct,
  title = {{ThinkStruct: RST-Aware Attention for Logical Reasoning in Machine Reading Comprehension}},
  author = {Tran, Nhi Thao and Le, Tien and Pham, Toan and Vu, Quoc Hoang},
  booktitle = {Proceedings of the 30th Conference on Computational Natural Language Learning},
  year = {2026},
  doi = {10.18653/v1/2026.conll-main.27},
  url = {https://doi.org/10.18653/v1/2026.conll-main.27}
}
```

- [← previous The Woman with Spindles in the Luwian-Phoenician Inscriptions from Karatepe](https://md.research.elamur.ai/free/en/papers/baranowski-2026-woman-spindles-luwian-phoenician-inscriptions/)
- [next → This One or That One? A Study on Accessibility via Demonstratives with Multimodal Large Language Models](https://md.research.elamur.ai/free/en/papers/wang-2026-this-one-or-that-one/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/10/)

## References

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

1. Sahil Bakshi and Dipti Misra Sharma. 2021. A transformer based approach towards identification of discourse unit segments and connectives. In Proceedings of the 2nd Shared Task on Discourse Relation Parsing and Treebanking (DISRPT 2021), pages 13– 21. [find](https://scholar.google.com/scholar?q=Sahil+Bakshi+and+Dipti+Misra+Sharma%2E+2021%2E+A+transformer+based+approach+towards+identification+of+discourse+unit+segments+and+connectives%2E+In+Proceedings+of+the+2nd+Shared+Task+on+Discourse+Relation+P)
2. Elena Chistova. 2024. Bilingual rhetorical structure parsing with large parallel annotations. In Findings of the Association for Computational Linguistics ACL 2024, pages 9689–9706, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Elena+Chistova%2E+2024%2E+Bilingual+rhetorical+structure+parsing+with+large+parallel+annotations%2E+In+Findings+of+the+Association+for+Computational+Linguistics+ACL+2024%2C+pages+9689%E2%80%939706%2C+Bangkok%2C+Thailand+) [our card](https://md.research.elamur.ai/free/en/papers/chistova-2024-bilingual-rhetorical-structure-parsing-large/)
3. Yinya Huang, Meng Fang, Yu Cao, Liwei Wang, and Xiaodan Liang. 2021. DAGN: discourse-aware graph network for logical reasoning. CoRR, abs/2103.14349. [find](https://scholar.google.com/scholar?q=Yinya+Huang%2C+Meng+Fang%2C+Yu+Cao%2C+Liwei+Wang%2C+and+Xiaodan+Liang%2E+2021%2E+DAGN%3A+discourse%2Daware+graph+network+for+logical+reasoning%2E+CoRR%2C+abs%2F2103%2E14349%2E)
4. Fangkai Jiao, Yangyang Guo, Xuemeng Song, and Liqiang Nie. 2022. MERIt: Meta-Path Guided Contrastive Learning for Logical Reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3496–3509, Dublin, Ireland. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Fangkai+Jiao%2C+Yangyang+Guo%2C+Xuemeng+Song%2C+and+Liqiang+Nie%2E+2022%2E+MERIt%3A+Meta%2DPath+Guided+Contrastive+Learning+for+Logical+Reasoning%2E+In+Findings+of+the+Association+for+Computational+Linguistics%3A+ACL+2)
5. Diederik P. Kingma and Jimmy Ba. 2017. Adam: A method for stochastic optimization. Preprint, arXiv:1412.6980. [find](https://scholar.google.com/scholar?q=Diederik+P%2E+Kingma+and+Jimmy+Ba%2E+2017%2E+Adam%3A+A+method+for+stochastic+optimization%2E+Preprint%2C+arXiv%3A1412%2E6980%2E)
6. Michael Lambropoulos and Shunichi Ishihara. 2024. Towards an implementation of rhetorical structure theory in discourse coherence modelling. In Proceedings of the 22nd Annual Workshop of the Australasian Language Technology Association, pages 1–11. [find](https://scholar.google.com/scholar?q=Michael+Lambropoulos+and+Shunichi+Ishihara%2E+2024%2E+Towards+an+implementation+of+rhetorical+structure+theory+in+discourse+coherence+modelling%2E+In+Proceedings+of+the+22nd+Annual+Workshop+of+the+Australas)
7. Xiao Li, Gong Cheng, Ziheng Chen, Yawei Sun, and Yuzhong Qu. 2022. Adalogn: Adaptive logic graph network for reasoning-based machine reading comprehension. Preprint, arXiv:2203.08992. [find](https://scholar.google.com/scholar?q=Xiao+Li%2C+Gong+Cheng%2C+Ziheng+Chen%2C+Yawei+Sun%2C+and+Yuzhong+Qu%2E+2022%2E+Adalogn%3A+Adaptive+logic+graph+network+for+reasoning%2Dbased+machine+reading+comprehension%2E+Preprint%2C+arXiv%3A2203%2E08992%2E)
8. Jian Liu, Leyang Cui, Hanmeng Liu, Dandan Huang, Yile Wang, and Yue Zhang. 2020. Logiqa: A challenge dataset for machine reading comprehension with logical reasoning. CoRR, abs/2007.08124. [find](https://scholar.google.com/scholar?q=Jian+Liu%2C+Leyang+Cui%2C+Hanmeng+Liu%2C+Dandan+Huang%2C+Yile+Wang%2C+and+Yue+Zhang%2E+2020%2E+Logiqa%3A+A+challenge+dataset+for+machine+reading+comprehension+with+logical+reasoning%2E+CoRR%2C+abs%2F2007%2E08124%2E) [our card](https://md.research.elamur.ai/free/en/papers/liu-2020-logiqa-challenge-dataset-machine-reading/)
9. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692. [find](https://scholar.google.com/scholar?q=Yinhan+Liu%2C+Myle+Ott%2C+Naman+Goyal%2C+Jingfei+Du%2C+Mandar+Joshi%2C+Danqi+Chen%2C+Omer+Levy%2C+Mike+Lewis%2C+Luke+Zettlemoyer%2C+and+Veselin+Stoyanov%2E+2019%2E+Roberta%3A+A+robustly+optimized+BERT+pretraining+approach%2E+C)
10. Zhengyuan Liu, Ke Shi, and Nancy F Chen. 2025. Transformer-based document-level discourse processing: Exploiting prior language knowledge and hierarchical parsing. Computer Speech & Language, page 101809. [find](https://scholar.google.com/scholar?q=Zhengyuan+Liu%2C+Ke+Shi%2C+and+Nancy+F+Chen%2E+2025%2E+Transformer%2Dbased+document%2Dlevel+discourse+processing%3A+Exploiting+prior+language+knowledge+and+hierarchical+parsing%2E+Computer+Speech+%26+Language%2C+page+101)
11. AI @ Meta. 2024. The llama 3 herd of models. CoRR, abs/2404.19737. [find](https://scholar.google.com/scholar?q=AI+%40+Meta%2E+2024%2E+The+llama+3+herd+of+models%2E+CoRR%2C+abs%2F2404%2E19737%2E)
12. OpenAI. 2023. GPT-4 technical report. abs/2303.08774. [find](https://scholar.google.com/scholar?q=OpenAI%2E+2023%2E+GPT%2D4+technical+report%2E+abs%2F2303%2E08774%2E) [our card](https://md.research.elamur.ai/free/en/papers/openai-2023-gpt-4-technical-report/)
13. CoRR, Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022.… [find](https://scholar.google.com/scholar?q=CoRR%2C+Long+Ouyang%2C+Jeff+Wu%2C+Xu+Jiang%2C+Diogo+Almeida%2C+Carroll+L%2E+Wainwright%2C+Pamela+Mishkin%2C+Chong+Zhang%2C+Sandhini+Agarwal%2C+Katarina+Slama%2C+Alex+Ray%2C+John+Schulman%2C+Jacob+Hilton%2C+Fraser+Kelton%2C+Luke+Mi)
14. Chenxu Wang, Ping Jian, and Zhen Yang. 2025. Thought-path contrastive learning via premise-oriented data augmentation for logical reading comprehension. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 25345– 25352. [find](https://scholar.google.com/scholar?q=Chenxu+Wang%2C+Ping+Jian%2C+and+Zhen+Yang%2E+2025%2E+Thought%2Dpath+contrastive+learning+via+premise%2Doriented+data+augmentation+for+logical+reading+comprehension%2E+In+Proceedings+of+the+AAAI+Conference+on+Artifi)
15. Siyuan Wang, Wanjun Zhong, Duyu Tang, Zhongyu Wei, Zhihao Fan, Daxin Jiang, Ming Zhou, and Nan Duan. 2021. Logic-driven context extension and data augmentation for logical reasoning of text. CoRR, abs/2105.03659. [find](https://scholar.google.com/scholar?q=Siyuan+Wang%2C+Wanjun+Zhong%2C+Duyu+Tang%2C+Zhongyu+Wei%2C+Zhihao+Fan%2C+Daxin+Jiang%2C+Ming+Zhou%2C+and+Nan+Duan%2E+2021%2E+Logic%2Ddriven+context+extension+and+data+augmentation+for+logical+reasoning+of+text%2E+CoRR%2C+abs)
16. Bingsheng Yao and 1 others. 2024. Discursive socratic questioning: Evaluating the faithfulness of language models’ understanding of discourse relations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Bingsheng+Yao+and+1+others%2E+2024%2E+Discursive+socratic+questioning%3A+Evaluating+the+faithfulness+of+language+models%E2%80%99+understanding+of+discourse+relations%2E+In+Proceedings+of+the+62nd+Annual+Meeting+of+th) [our card](https://md.research.elamur.ai/free/en/papers/miao-2024-discursive-socratic-questioning-evaluating-faithfulness/)
17. Weihao Yu, Zihang Jiang, Yanfei Dong, and Jiashi Feng. 2020. Reclor: A reading comprehension dataset requiring logical reasoning. CoRR, abs/2002.04326. [find](https://scholar.google.com/scholar?q=Weihao+Yu%2C+Zihang+Jiang%2C+Yanfei+Dong%2C+and+Jiashi+Feng%2E+2020%2E+Reclor%3A+A+reading+comprehension+dataset+requiring+logical+reasoning%2E+CoRR%2C+abs%2F2002%2E04326%2E) [our card](https://md.research.elamur.ai/free/en/papers/yu-2020-reclor-reading-comprehension-dataset-requiring/)