SSEmb: A Joint Structural and Semantic Embedding Framework for Mathematical Formula Retrieval

Ruyin Li; Xiaoyu Chen

REVIEW 4 major objections 5 minor 1 cited by

SSEmb: A Joint Structural and Semantic Embedding Framework for Mathematical Formula Retrieval

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SSEmb blends graph structure and surrounding-text semantics to raise formula retrieval accuracy, outperforming all embedding-based baselines on ARQMath-3.

desk verdict A coherent joint structural-semantic formula retriever with a genuinely new augmentation, but the headline 5-point margin is arithmetically overclaimed and the key validity assumption is untested. read the letter →

arxiv 2508.04162 v2 pith:FGEHE3DU submitted 2025-08-06 cs.IR

Ruyin Li , Xiaoyu Chen This is my paper · ORCID

classification cs.IR

keywords mathematicalinformationretrievalformulagraphcontrastivelearningoperatorsubstructuresubstitutionsemanticembeddingSentence-BERTARQMath

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that formula retrieval is best served by a joint embedding of what a formula is (its operator structure) and what it is about (the text around it). To capture structure, SSEmb applies Graph Contrastive Learning to Operator Graphs, with a new augmentation that replaces selected substructure roots with wildcards in a way claimed to preserve mathematical validity. To capture semantics, it encodes the surrounding post text with Sentence-BERT. On the ARQMath-3 formula retrieval task, the combined system outperforms embedding-based baselines by over five percentage points on P'@10 and nDCG'@10, and when fused by reciprocal rank fusion it improves every method it is combined with, reaching state-of-the-art results together with Approach0. If correct, the framework shows that contextual semantics plus structure-aware graph embeddings, rather than structure alone, is the effective recipe.

What carries the argument

Operator Graph (OPG), a compact directed acyclic graph built from the Operator Tree by sharing identical subtrees; Graph Contrastive Learning (GCL) with a novel substructure-substitution augmentation that replaces the root of selected substructures with wildcards from leaves upward; Sentence-BERT encoding of truncated surrounding text; and a weighted cosine-similarity fusion in two-stage ranking. The substitution is the load-bearing novelty: it is meant to create diverse positive views while preserving the computational skeleton of a formula, so the GNN learns structure from global to local scale.

What would settle it

Take the trained StructEmb module and replace substructure substitution with equally frequent random root-attribute substitutions at arbitrary depths; if ARQMath-3 P'@10 and nDCG'@10 do not drop, the hierarchical substitution strategy is not the source of the gain. Alternatively, re-linearize substituted OPGs and parse them as formulas: if most do not parse, mathematical validity is not being preserved.

Watch

Extended reading notes

Core claim

SSEmb's central claim is that structural and semantic cues are complementary for formula retrieval, and that they can be combined in one embedding pipeline. Structurally, formulas are converted to Operator Graphs—compact directed acyclic graphs that share identical subtrees—and embedded by a GNN trained with Graph Contrastive Learning. The paper introduces substructure substitution, a hierarchical augmentation that starts at leaf nodes and proceeds upward, replacing selected substructure roots with wildcards; lower-order substitutions are applied more often to preserve the global skeleton while adding local variation. Semantically, the text of the post containing each formula is truncated to

Load-bearing premise

The claim rests on the assumption that wildcard substitution produces mathematically meaningful formula variants that teach structure rather than destroying it; no measurement of that validity is provided.

Editorial extensions

If this is right

Embedding-based formula retrieval can close most of the gap to matching-based systems like Approach0, at least on ARQMath-3.
Surrounding text matters: removing the SemEmb module drops nDCG'@10 from 0.7343 to 0.6762.
Substructure substitution alone gives competitive structural embeddings, outperforming the embedding baseline TangentCFT2.
Because RRF fusion with SSEmb improves every baseline run, the SSEmb ranking is complementary to existing structural matchers.
Combined with Approach0, SSEmb reaches the best reported result on the ARQMath-3 formula retrieval task.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The 'preserving mathematical validity' claim for substructure substitution is asserted but not measured; if validated, the same augmentation could transfer to other symbolic structures (code snippets, chemical graphs) where subtree replacement is a natural view generator.
The fusion weight alpha=0.5 and the 1024-character truncation suggest the authors implicitly assume a roughly equal balance between structure and text; a per-query adaptive weight could be tested against the fixed scheme.
Because SSEmb uses only the post text and formula tree structure, it could be extended to ARQMath's answer retrieval task without retraining on new modalities, a direction the authors mention.
The gains from adding SSEmb's run to every baseline suggest that its ranking errors are largely independent of existing matchers; a formal diversity analysis of the failure sets would support this inference.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things. The substructure substitution augmentation is a genuinely new idea for operator graphs, and the two-module fusion is clean. But the abstract's 'over 5 percentage points' claim is not actually supported by their own Table 1: nDCG'@10 margin is 4.75pp, not 5. The paper is still worth a serious look.

What's new: They adapt Graph Contrastive Learning to the OPG representation developed in Song & Chen's earlier work, and instead of generic node/edge dropping they replace whole substructures with wildcards at increasing levels. That's a plausible augmentation for math formulas, and the ablation suggests it beats attribute masking alone and the standard node/edge perturbations. Using Sentence-BERT over 1024 characters of surrounding text, rather than a single sentence, is a sensible extension of MathAMR. The two-stage retrieval is pragmatic. The internal numbers in Tables 1-3 are consistent with each other.

Soft spots: The stress-test note is right. Table 1 shows 0.7343 vs 0.6868, a 4.75pp gain on nDCG'@10. So the joint claim 'over 5 percentage points on P'@10 and nDCG'@10' is false as written; only P'@10 clears that bar. That's a fixable but embarrassing headline error. More substantive: the core premise that substructure substitution preserves mathematical validity is asserted in one sentence, with no examples, no manual checks, no measurement. If the wildcard substitutions often create meaningless graphs, the augmentation is just noise. Also, they don't compare against the closest prior work: Wang & Chen's GCL on formula trees, or Mansouri et al.'s MathAMR contextual embeddings. Those are the two systems that make their claimed contribution legible. And there's no significance testing, no released code/data, and the hyperparameters (p1, p2, p3, alpha) are given without a validation protocol. These are all standard things for a conference submission, but they matter here because the headline margin is only a few points.

But the central conclusion--that fusing structural and semantic embeddings helps and that structured augmentation helps GCL--is plausible and backed by the ablation. The paper is not a mess; it's an incomplete but promising system.

Who it's for: people working on math information retrieval and graph contrastive learning for structured data. A serious referee should get it, but should demand a corrected abstract, a comparison with the two missing baselines, and some evidence that the substitution preserves validity. I'd send it to review.

Referee Report

4 major / 5 minor

Summary. The paper proposes SSEmb, a retrieval framework for mathematical formulas that combines structural embeddings with contextual semantic embeddings. Structurally, formulas are converted to Operator Graphs (OPG) and encoded by a Graph Contrastive Learning model with a proposed substructure-substitution augmentation. Semantically, surrounding post text is encoded with Sentence-BERT. Retrieval is performed by weighted fusion of structural and semantic cosine similarities. The method is evaluated on the ARQMath-3 formula retrieval task, reporting gains over embedding-based baselines and improved results when SSEmb is combined with existing runs by reciprocal rank fusion.

Significance. If the reported results hold, the paper makes a useful contribution to mathematical information retrieval by demonstrating that a joint structural-semantic embedding approach, with a formula-specific graph augmentation, can outperform prior embedding-based systems on a standard benchmark. The ablation study gives some support for the qualitative claim that both the semantic module and the substructure-substitution augmentation contribute to performance. The paper does not provide code or reproducibility artifacts, and several load-bearing claims require additional evidence, so the significance is conditional on those points being addressed.

major comments (4)

[Abstract and §3.2, Table 1] The abstract claims SSEmb 'outperforms existing embedding-based methods by over 5 percentage points on P'@10 and nDCG'@10.' Table 1 reports the best embedding baseline, TangentCFT2ED, at nDCG'@10=0.6868 and SSEmb at 0.7343. The difference is 0.0475, i.e. 4.75 percentage points, which is below 5. The P'@10 margin (0.6803 vs 0.6105) exceeds 5, but the joint claim over both metrics is not supported by the paper's own numbers. This is a direct inconsistency in the headline result and must be corrected.
[§2.1, substructure substitution] The paper asserts that the proposed substructure substitution 'preserv[es] mathematical validity' and 'mathematical coherence' of OPGs, and this claim is central to the rationale for why this augmentation is preferable to node dropping or edge perturbation. However, no measurement, qualitative example, or validity check is provided. The ablation shows that removing substructure substitution hurts performance, but it does not establish that the augmented graphs are mathematically meaningful. Without evidence, the reported structural gains could be an artifact of the specific augmentation distribution rather than a robust learning signal. The authors should provide concrete examples of augmented OPGs and, ideally, a quantitative validity check or an analysis of how often substitutions produce degenerate graphs.
[§3.3, Table 2] The ablation table labels are difficult to interpret. Entries such as '- StructEmb-subs', '- StructEmb-attr_node_edge', and '- StructEmb-attr' are not explicitly defined in the text, and the prose describing 'three contrastive experiments' does not clearly map onto the table rows. This makes it hard for a reader to know exactly which component is removed in each row and undermines reproducibility. The authors should rename the rows and state precisely which augmentation modules are included in each configuration.
[§2.3, §3.1] The method depends on several hyperparameters, including the fusion weight α (set to 0.5) and the substitution probabilities p1=0.3, p2=0.005, p3=0.002. No sensitivity analysis is reported for any of these. Since α directly controls the balance between the two modalities and the probabilities control the augmentation distribution, it is important to know how stable the reported results are with respect to reasonable variations. Without this, the reader cannot assess whether the reported improvements are robust or an artifact of a particular hyperparameter choice.

minor comments (5)

[Conclusion] Typo: 'structual' should be 'structural'.
[References] Reference [28] has 'ClEF' instead of 'CLEF'. Please check all references for consistent venue naming.
[Table 3] The table header 'Original RRF' is ambiguous: it appears to show both the original run scores and the RRF-combined scores, but the layout is confusing. Please make clear which columns are original and which are combined.
[Reproducibility] No link to source code or trained models is provided. Including such artifacts, or at least a detailed description of preprocessing steps, would substantially improve reproducibility.
[§2.2] The semantic module truncates surrounding text to 1024 characters. It is not discussed how queries with very little or no surrounding text are handled. This could affect retrieval performance for a subset of queries and should be clarified.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claims are external ARQMath-3 benchmark evaluations; the only self-citation (OPG from [22]) is a representation choice, not a circular reduction.

full rationale

SSEmb is an empirical retrieval system evaluated on the ARQMath-3 benchmark against published baselines. There is no derivation chain in which a predicted quantity is defined in terms of an input or in which a fitted parameter is renamed as a prediction. The structural module uses OPG from the authors' prior work [22] (Song and Chen, with current co-author Xiaoyu Chen), but this is a representation choice, not a theorem whose uniqueness forces the result; the reported gains are measured against external runs. The hyperparameters (p1=0.3, p2=0.005, p3=0.002, alpha=0.5) are stated, and although their tuning is not described, no metric is a mathematical function of them by construction. Flagged for completeness: the Abstract claims 'SSEmb outperforms existing embedding-based methods by over 5 percentage points on P\'@10 and nDCG\'@10,' but Table 1 gives nDCG'@10 margins of 0.7343 - 0.6868 = 0.0475 (4.75pp) for the best embedding baseline; this is an arithmetic/consistency defect in the headline, not circularity. Also, the Section 2.1 assertion that substructure substitution preserves 'mathematical validity' is unmeasured; this is a correctness/robustness concern, not a circular step. Because the central result is an external benchmark comparison with no equation-level equivalence to inputs, the circularity score is low.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The paper contributes a new augmentation and a new integration of existing components. Its central claim rests on imported OPG representation, Graph Contrastive Learning, Sentence-BERT, and ARQMath judgments, plus an unverified assumption that wildcard substitutions preserve mathematical validity. No new physical or conceptual entities such as new forces, particles, or conserved quantities are introduced; the only new construct is the substitution operation itself, which is a method rather than an entity.

free parameters (8)

fusion weight alpha = 0.5
Controls the balance between structural and semantic similarity in Eq. (3); fixed with no sensitivity analysis or validation split reported.
substitution probabilities p1, p2, p3 = 0.3, 0.005, 0.002
Probabilities for first-, second-, and third-order substructure substitution in Section 2.1; chosen by hand without reported validation.
attribute masking rate = 0.01
Fraction of node attributes masked during graph augmentation; fixed hyperparameter.
node label frequency cutoff = 11
Node labels appearing fewer than 11 times across all formulas are discarded before random initialization.
initial candidate pool size = 500,000
Stage-1 structural retrieval cutoff before semantic re-ranking; affects both efficiency and the recall ceiling.
embedding dimension = 400
Output dimension of the structural GNN encoder.
GNN layers = 2
Number of Graph Isomorphism Network layers in the structural encoder.
learning rate, batch size, epochs = 1e-4, 2560, 25
Training configuration for the contrastive structural encoder.

assumptions (6)

domain assumption The Operator Graph is a faithful and compact representation of formula structure.
Adopted from [22], whose first author is the corresponding author of this paper; all structural modeling is built on this representation.
domain assumption Substructure substitution preserves mathematical validity and creates useful positive views for contrastive learning.
Section 2.1 asserts this to justify the augmentation; no validity checks or view-quality analysis are provided, making it a load-bearing unverified premise.
domain assumption Surrounding text, truncated to 1024 characters, encodes formula semantics relevant for retrieval.
Section 2.2 treats post text as the source of usage and domain semantics; no analysis of noise or truncation effects is given.
domain assumption Graph Contrastive Learning on unlabeled formula graphs yields embeddings whose cosine similarity orders structural relevance.
Section 2.1 applies InfoNCE to OPG views; the link between contrastive geometry and formula similarity is assumed.
domain assumption ARQMath-3 relevance judgments and official run scores are a valid external benchmark.
Section 3 uses 76 queries and official judgments; the paper's claims inherit the benchmark's validity and completeness.
domain assumption Sentence-BERT (all-mpnet-base-v2) transfers to mathematical text without domain adaptation.
Section 2.2 uses a general-purpose pretrained model directly; its applicability to mathematical prose is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SSEmb: A Joint Structural and Semantic Embedding Framework for Mathematical Formula Retrieval." pith.science (2026). https://pith.science/paper/FGEHE3DU

@misc{pith2026250804162,
  author       = {Pith},
  title        = {Pith review of: SSEmb: A Joint Structural and Semantic Embedding Framework for Mathematical Formula Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FGEHE3DU}},
  note         = {Machine review of arXiv:2508.04162}
}

read the original abstract

Formula retrieval is an important topic in Mathematical Information Retrieval. We propose SSEmb, a novel embedding framework capable of capturing both structural and semantic features of mathematical formulas. Structurally, we employ Graph Contrastive Learning to encode formulas represented as Operator Graphs. To enhance structural diversity while preserving mathematical validity of these formula graphs, we introduce a novel graph data augmentation approach through a substitution strategy. Semantically, we utilize Sentence-BERT to encode the surrounding text of formulas. Finally, for each query and its candidates, structural and semantic similarities are calculated separately and then fused through a weighted scheme. In the ARQMath-3 formula retrieval task, SSEmb outperforms existing embedding-based methods by over 5 percentage points on P'@10 and nDCG'@10. Furthermore, SSEmb enhances the performance of all runs of other methods and achieves state-of-the-art results when combined with Approach0.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Syntax Meets Semantics: Understanding Scientific Formulae
cs.IR 2026-08 conditional novelty 6.0 of 10

Formula syntax and textual semantics show weak direct correspondence but strong latent correlation; contrastive learning bridges the gap and lifts retrieval from ~5% to ~58% recall@10.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages · cited by 1 Pith paper

[1]

Cormack, Charles L A Clarke, and Stefan Buettcher

Gordon V. Cormack, Charles L A Clarke, and Stefan Buettcher. 2009. Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 758–759

work page 2009
[2]

Yifan Dai, Liangyu Chen, and Zihan Zhang. 2020. An N-ary Tree-based Model for Similarity Evaluation on Mathematical Formulae. In Proceedings of the 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC) . 2578–2584

work page 2020
[3]

Kenny Davila and Richard Zanibbi. 2017. Layout and Semantics: Combining Representations for Mathematical Formula Search. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). 1165–1168

work page 2017
[4]

Kenny Davila, Richard Zanibbi, Andrew Kane, and Frank Wm. Tompa. 2016. Tangent-3 at the NTCIR-12 MathIR Task. In Proceedings of the 12th NTCIR Conference on Evaluation of Information Access Technologies (NTCIR) . 338–345

work page 2016
[5]

Liangcai Gao, Zhuoren Jiang, Yue Yin, Ke Yuan, Zuoyu Yan, and Zhi Tang. 2017. Preliminary Exploration of Formula Embedding for Mathematical Information Retrieval: Can Mathematical Formulae be Embedded like A Natural Language? arXiv preprint arXiv:1707.05154 (2017)

work page Pith review arXiv 2017
[6]

Xuan Hu, Liangcai Gao, Xiaoyan Lin, Zhi Tang, Xiaofan Lin, and Josef B. Baker

work page
[7]

Andrew Kane, Yin Ng, and Frank Tompa. 2022. Dowsing for Answers to Math Questions: Doing Better With Less. InProceedings of the 13th Conference and Labs of the Evaluation Forum (CLEF) . 40–62

work page 2022
[8]

Giovanni Yoko Kristianto, Goran Topic, and Akiko Aizawa. 2016. MCAT Math Retrieval System for NTCIR-12 MathIR Task. In Proceedings of the 12th NTCIR Conference on Evaluation of Information Access Technologies (NTCIR) . 323–330

work page 2016

Show all 30 references

[9]

Kriste Krstovski and David M. Blei. 2018. Equation Embeddings. arXiv preprint arXiv:1803.09123 (2018)

2018 arXiv
[10]

Pavan Kumar, Arun Agarwal, and Chakravarthy Bhagvati

P. Pavan Kumar, Arun Agarwal, and Chakravarthy Bhagvati. 2012. A Structure Based Approach for Mathematical Expression Retrieval. In Proceedings of the 6th Multi-Disciplinary International Workshop on Artificial Intelligence (MIW AI). 23–34

2012
[11]

Matt Langsenkamp, Bryan Amador, and Richard Zanibbi. 2022. Expanding Spatial Regions and Incorporating IDF for PHOC-Based Math Formula Retrieval at ARQMath-3. In Proceedings of the 13th Conference and Labs of the Evaluation Forum (CLEF). 63–82

2022
[12]

Oard, and Richard Zanibbi

Behrooz Mansouri, Vít Novotný, Anurag Agarwal, Douglas W. Oard, and Richard Zanibbi. 2022. Overview of ARQMath-3 (2022): Third CLEF Lab on Answer Retrieval for Questions on Math. In Proceedings of the 13th Conference and Labs of the Evaluation Forum (CLEF) . 1–27

2022
[13]

Behrooz Mansouri, Douglas Oard, and Richard Zanibbi. 2022. DPRL Systems in the CLEF 2022 ARQMath Lab: Introducing MathAMR for Math-Aware Search. In Proceedings of the 13th Conference and Labs of the Evaluation Forum (CLEF) . 83–103

2022
[14]

Oard, and Richard Zanibbi

Behrooz Mansouri, Douglas W. Oard, and Richard Zanibbi. 2022. Contextualized Formula Search Using Math Abstract Meaning Representation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM). 4329–4333

2022
[15]

Oard, Jian Wu, C

Behrooz Mansouri, Shaurya Rohatgi, Douglas W. Oard, Jian Wu, C. Lee Giles, and Richard Zanibbi. 2019. Tangent-CFT: An Embedding Model for Mathematical Formulas. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (...

2019
[16]

Miller and Abdou S

Bruce R. Miller and Abdou S. Youssef. 2003. Technical Aspects of the Digital Library of Mathematical Functions. Annals of Mathematics and Artificial Intelligence 38, 1–3 (2003), 121–136

2003
[17]

Jozef Mišutka and Leo Galamboš. 2008. Extending Full Text Search Engine for Mathematical Content. Towards Digital Mathematics Library (2008), 55–67

2008
[18]

Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748 (2018)

2018 arXiv
[19]

Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCN...

2019
[20]

Tetsuya Sakai and Noriko Kando. 2008. On Information Retrieval Metrics Designed for Evaluation with Incomplete Relevance Assessments. Information Retrieval 11 (2008), 447–470

2008
[21]

Sandip Sarkar, Dipankar Das, Partha Pakray, and David Pinto. 2022. Formula Retrieval Using Structural Similarity. In Proceedings of the 13th Conference and Labs of the Evaluation Forum (CLEF) . 138–146

2022
[22]

Yujin Song and Xiaoyu Chen. 2021. Searching for Mathematical Formulas Based on Graph Representation Learning. In Proceedings of the 14th Conference on Intelligent Computer Mathematics (CICM) . 137–152

2021
[23]

Abhinav Thanda, Ankit Agarwal, Kushal Singla, Aditya Prakash, and Abhishek Gupta. 2016. A Document Retrieval System for Math Queries. In Proceedings of the 12th NTCIR Conference on Evaluation of Information Access Technologies (NTCIR). 346–353

2016
[24]

Pei-Syuan Wang and Hung-Hsuan Chen. 2025. The Effectiveness of Graph Contrastive Learning on Mathematical Information Retrieval. In Proceedings of the 1st International Workshop on Graph-Based Approaches in Information Retrieval (IRonGraphs). 60–72

2025
[25]

Yuehan Wang, Liangcai Gao, Simeng Wang, Zhi Tang, Xiaozhong Liu, and Ke Yuan. 2015. WikiMirs 3.0: A Hybrid MIR System Based on the Context, Structure and Importance of Formulae in a Document. InProceedings of the 15th ACM/IEEE- CS Joint Conference on Digital Libraries (JCDL) . 173–182

2015
[26]

Lafferty

Michihiro Yasunaga and John D. Lafferty. 2019. TopicEq: A Joint Topic and Mathematical Equation Model for Scientific Texts. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI) . 7394–7401

2019
[27]

Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph Contrastive Learning with Augmentations. In Proceedings of the 34th International Conference on Neural Information Processing Systems (NeurIPS). 5812–5823

2020
[28]

Wei Zhong, Yuqing Xie, and Jimmy Lin. 2022. Applying Structural and Dense Semantic Matching for the ARQMath Lab 2022, ClEF. In Proceedings of the 13th Conference and Labs of the Evaluation Forum (CLEF) . 147–170

2022
[29]

Wei Zhong and Richard Zanibbi. 2019. Structural Similarity Search for Formulas Using Leaf-Root Paths in Operator Subtrees. In Proceedings of the 41st European Conference on Information Retrieval (ECIR) . 116–129

2019
[2013]

In Proceedings of the 13th ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL)

WikiMirs: A Mathematical Information Retrieval System for Wikipedia. In Proceedings of the 13th ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL) . 11–20

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Cormack, Charles L A Clarke, and Stefan Buettcher

Gordon V. Cormack, Charles L A Clarke, and Stefan Buettcher. 2009. Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. In Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR) . 758–759

work page 2009

[2] [2]

Yifan Dai, Liangyu Chen, and Zihan Zhang. 2020. An N-ary Tree-based Model for Similarity Evaluation on Mathematical Formulae. In Proceedings of the 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC) . 2578–2584

work page 2020

[3] [3]

Kenny Davila and Richard Zanibbi. 2017. Layout and Semantics: Combining Representations for Mathematical Formula Search. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). 1165–1168

work page 2017

[4] [4]

Kenny Davila, Richard Zanibbi, Andrew Kane, and Frank Wm. Tompa. 2016. Tangent-3 at the NTCIR-12 MathIR Task. In Proceedings of the 12th NTCIR Conference on Evaluation of Information Access Technologies (NTCIR) . 338–345

work page 2016

[5] [5]

Liangcai Gao, Zhuoren Jiang, Yue Yin, Ke Yuan, Zuoyu Yan, and Zhi Tang. 2017. Preliminary Exploration of Formula Embedding for Mathematical Information Retrieval: Can Mathematical Formulae be Embedded like A Natural Language? arXiv preprint arXiv:1707.05154 (2017)

work page Pith review arXiv 2017

[6] [6]

Xuan Hu, Liangcai Gao, Xiaoyan Lin, Zhi Tang, Xiaofan Lin, and Josef B. Baker

work page

[7] [7]

Andrew Kane, Yin Ng, and Frank Tompa. 2022. Dowsing for Answers to Math Questions: Doing Better With Less. InProceedings of the 13th Conference and Labs of the Evaluation Forum (CLEF) . 40–62

work page 2022

[8] [8]

Giovanni Yoko Kristianto, Goran Topic, and Akiko Aizawa. 2016. MCAT Math Retrieval System for NTCIR-12 MathIR Task. In Proceedings of the 12th NTCIR Conference on Evaluation of Information Access Technologies (NTCIR) . 323–330

work page 2016