Pith. sign in

REVIEW 5 major objections 6 minor 61 references

The paper argues that text matching should be modeled as structured representation alignment: a dual-graph encoder over syntax and topic nodes, trained with node- and graph-level contrastive losses, pushes legal statute matching to 86.7% F1

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

StructCoh, a graph-enhanced contrastive learning framework for text semantic matching, reportedly outperforms prior methods on legal and plagiarism benchmarks, but the reported results are not reproducible from the paper alone.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Plausible architecture, but the headline result is unfalsifiable without code, data, and a specified node-alignment procedure. the 5 major comments →

arxiv 2509.02033 v1 pith:6NO3SYX3 submitted 2025-09-02 cs.CL

StructCoh: Structured Contrastive Learning for Context-Aware Text Semantic Matching

Chao Xue , Ziyuan Gao This is my paper
classification cs.CL
keywords semantic text matchingstructured contrastive learningdual-graph encodergraph isomorphism networklegal document matchingplagiarism detectionhierarchical contrastive losshard negative mining
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper proposes that text semantic matching should be treated as a structural representation alignment problem rather than token-sequence comparison. To do so it introduces StructCoh, a dual-graph encoder that builds a syntactic dependency graph and a topic-interaction graph for each text, fuses them, and trains a hierarchical contrastive objective at both node and graph levels. The paper reports large gains on specialized benchmarks: 86.7% F1 on COLIEE-2023 legal statute matching (+6.2 absolute), 84.5% F1 on CaseLaw retrieval, and 82.4% accuracy on a new plagiarism-detection dataset, plus average accuracy of 85.59 (base) and 88.16 (large) across six GLUE tasks. The intended payoff is a matching model that does not collapse under paraphrases and adversarial edits, where surface lexical overlap is misleading.

Core claim

The paper's central claim is that text semantic matching improves when the comparison happens at two structural levels at once: local semantic units within each text and whole-document structural patterns across texts. StructCoh encodes each text with two graphs—a dependency graph over syntactic tokens and a topic interaction graph over entities/topics—using Graph Isomorphism Network layers initialized from BERT token, POS, and relation embeddings. A cross-graph attention step lets syntactic nodes borrow topic context, and attentive pooling yields a text vector. Training couples a node-level InfoNCE loss over aligned node pairs with a graph-level InfoNCE loss whose negatives are hard samples

What carries the argument

The carrying mechanism is the dual-graph encoder with a hierarchical contrastive objective. The syntactic dependency graph records token-level structural relations such as subject and object; the topic interaction graph groups entities/topics with co-occurrence or similarity edges. Both graphs are propagated through Graph Isomorphism Network layers, fused by cross-graph attention, and pooled into a global text vector. The training objective is two-scale: node-level InfoNCE regularizes aligned semantic units, and graph-level InfoNCE aligns whole-document representations, with negatives mined by structural dissimilarity. The fusion and the two losses are what the ablation study credits for the

Load-bearing premise

The node-level part of the training objective relies on a precomputed list of corresponding nodes between the two texts, built by string, entity, or role-matching heuristics; the paper never says how this list is made, and if it is noisy, the learned representations can lose validity.

What would settle it

Replace the unspecified node-alignment set in Eq. (10) with random node pairs on COLIEE-2023, keeping all other settings fixed. If F1 stays near 86.7%, the node-level loss is not doing the claimed work; if it drops toward the 82.2% ablation value, the alignment heuristic is load-bearing and the method's reproducibility depends on publishing it.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Legal matching systems could move from lexical-overlap scoring to structural alignment, which is exactly where the paper reports the +6.2-point F1 gain on COLIEE-2023.
  • Paraphrase-resistant plagiarism detection becomes a graph-alignment task: on SPD-1.0 the reported 82.4% accuracy is 14.7 points above BERT-based baselines.
  • General-domain matching also benefits: StructCoh-large reports 88.16 average accuracy on six GLUE tasks, above the SyntaxBERT-large number in the same table.
  • The ablations imply that all four components—dependency graph, topic graph, fusion layer, and hierarchical contrastive loss—are load-bearing, since removing any one drops COLIEE-2023 F1 by several points.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The node-alignment set in Eq. (10) is left unspecified; if it is noisy, the node-level loss may hurt more than help. A testable variant would build the alignment from attention or graph matching instead of hand-crafted string/entity/role heuristics, or remove the node-level term entirely.
  • The hard-negative threshold in Eq. (12) suggests a curriculum-like effect: negatives are selected dynamically from the batch, so batch composition matters. One could test sensitivity to batch size and negative mining threshold on the legal benchmarks.
  • The same dual-graph structure could transfer to other structure-rich matching tasks such as scientific claim verification or clinical trial comparison, where argument structure rather than surface wording determines the label; the paper does not test these domains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes StructCoh, a graph-enhanced contrastive learning framework for text semantic matching. It constructs a syntactic dependency graph and a topic interaction graph for each input text, encodes them with Graph Isomorphism Networks, fuses the two views with cross-graph attention, and trains the model with a hierarchical contrastive objective that combines node-level alignment (Eq. 10) and graph-level contrast (Eq. 11), plus hard negative mining (Eq. 12). The authors claim strong results on three legal/plagiarism benchmarks, including an 86.7% F1 score and +6.2% absolute gain over Coref-GNN on COLIEE-2023 legal statute matching, as well as improvements on GLUE tasks and TextFlint robustness benchmarks. The paper also introduces a new plagiarism dataset (SPD-1.0) and reports an ablation study on COLIEE-2023.

Significance. If the reported results were reproducible, the paper would make a meaningful contribution: the dual-graph encoder and hierarchical contrastive objective are a reasonable combination of existing ideas, and the claimed gains on legal matching and plagiarism detection would be practically important. The authors also explicitly propose hard negative mining and graph fusion, and the general framing of structural alignment is well motivated. However, the paper is first and foremost an empirical claim, and at present none of the central numbers can be verified from the manuscript: there is no code, no data, no numeric baseline table for the specialized-domain results, and the GLUE statistics in Table 2 are inconsistent with public dataset documentation. The node alignment set M in Eq. (10) is not specified, so even the training objective is not fully defined. These issues preclude a positive assessment of the paper's central claims.

major comments (5)
  1. [Section 3.4, Eq. (10)] The set M of aligned node pairs is never defined beyond the sentence 'string match, entity match, or role alignment heuristics.' Since L_node is one of only two training losses, any implementation must make concrete choices about M (e.g., exact string matching after lemmatization, partial entity overlap, role alignment by dependency labels). The paper gives no algorithm, no examples, no statistics on M (coverage, precision), and no sensitivity analysis. Different choices of M change the training objective and could drive the reported gains for reasons unrelated to argument-structure alignment. This is a load-bearing specification gap.
  2. [Section 4.1 and Table 1] The headline result — 86.7% F1 with +6.2% absolute gain over Coref-GNN on COLIEE-2023 — is not testable from the manuscript. Table 1 lists only '5,000 / 1,000' pairs and the vague task label 'Legal Statute Matching'; it does not say whether this is the official COLIEE statute retrieval task, the entailment task, or a custom re-split. Figure 2 is referenced instead of a numeric baseline table, and no Coref-GNN score or other baseline numbers are given. The metric (F1 over what units? pairs? paragraphs? documents?) is also unspecified. The central empirical claim is therefore unfalsifiable as presented.
  3. [Table 2] The GLUE dataset statistics are inconsistent with public documentation. For example, RTE dev is listed as 5,462 whereas the public dev set has 277 examples; QNLI train/dev/test are listed as 104,744/40,432/5,464 instead of the public 104,743/5,463/5,463. STS-B is a regression task with 5-class labels, not a binary classification task, yet it is listed under #Class=2 and the column header appears as 'SST-B' in Table 3 despite the text referring to STS-B. These errors make it impossible to reconstruct the experimental setup and undermine confidence in all GLUE results.
  4. [Sections 3.6–4.4] All reported results are single numbers with no standard deviations, no multiple seeds, and no statistical significance tests. Table 5 (ablation) similarly reports one run per variant. No code, data, or reproducibility instructions are provided. Given that some GLUE differences are small (e.g., StructCoh-base vs. SyntaxBERT-base on several tasks), the absence of variance and code means the reader cannot determine whether the claimed improvements are meaningful or within run-to-run noise. For an empirical paper whose central contribution is performance, this is a major limitation.
  5. [Section 3.4, Eq. (12)] The hard-negative threshold gamma and the loss weights lambda_node, lambda_graph, temperatures tau_n, tau_g, LDA topic count, and GIN dimensions are all free hyperparameters, but the paper gives no chosen values and no sensitivity analysis. Because the hard negative mining mechanism is central to the claimed improvement, the reader cannot assess how much of the gain depends on fine-tuned thresholding or whether the method is robust to these choices. At minimum, the paper should report the selected hyperparameters and a sensitivity study over gamma and the lambda weights.
minor comments (6)
  1. [Table 3 and Section 4.2] The column header 'SST-B' is inconsistent with the text's 'STS-B'; moreover, Section 4.2 claims superior performance on STS-B, but Table 3 reports no STS-B column. Clarify whether STS-B results are omitted or mislabeled.
  2. [Section 3.2, Eq. (3)] The notation p in Eq. (3) is undefined, and the formula for omega_kl is not fully specified. Define Co-occur, Freq, and p, or remove the undefined symbol.
  3. [References] The reference list is in an inconsistent style and contains corrupted or truncated entries, e.g., [5] 'D. LiangacOS Automation!', [36] is cut off, and [43] does not match the described UER paper. Several self-citations appear tangential to the claims. The manuscript needs a thorough reference cleanup.
  4. [Table 4] The transformation names are inconsistent: 'Appendlrr (Al)' should likely be 'AppendIrr (AI)', and abbreviations such as SA, NW, IA, Al, AP, BT, TT, SN, SW are not all defined in the table or text. Consistency in naming would improve readability.
  5. [Table 6] The case study table is confusing: the header row lists models but the columns contain binary labels and similarity percentages, and it is unclear which model produced which number. Please clarify the presentation.
  6. [Section 1 / Table 1] The paper introduces a new dataset SPD-1.0 but provides no description of how it was constructed, what documents it contains, how plagiarism labels were obtained, or where it can be accessed. A new benchmark claim requires this information.

Circularity Check

0 steps flagged

No significant circularity: the benchmark claims are empirical and the model components are not defined in terms of the target results.

full rationale

StructCoh's central claims are empirical F1/accuracy comparisons on external benchmarks. The methodology is not circular by construction: dual-graph construction (Eqs. 1–3), GIN encoding (Eq. 4), cross-graph fusion (Eqs. 5–6), attentive pooling (Eqs. 7–9), node-level contrast (Eq. 10), and graph-level contrast (Eq. 11) all take text and labels as input and produce representations, but the reported 86.7% F1 on COLIEE-2023 is a measured outcome, not an algebraic consequence of these equations. The alignment set M in Eq. (10) is a training input, not the prediction; even if M is heuristic, the final similarity in Eq. (14) is a cosine over pooled graph embeddings, not a direct readout of M. Hard-negative mining in Eq. (12) uses the model's own similarity, but that is standard iterative hard-negative mining, not a circular reduction of the prediction to the input. Self-citations (e.g., refs. [2], [30], [34], [44]) appear only in related-work or background contexts and are not load-bearing for the core architecture or the benchmark result. The paper does have serious reproducibility and correctness gaps—no numeric baseline table is provided for the COLIEE-2023 +6.2% claim, Table 2 reports impossible GLUE dev-set sizes (e.g., RTE 5,462 vs. the public 277), and Eq. (10)'s M is left unspecified—but these are reporting/reproducibility problems, not circularity. Accordingly, no circular step is identified and the derivation chain is self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 0 invented entities

The central claims rest on several unstated modeling choices: the node alignment heuristic, hard-negative threshold gamma, loss weights, and graph construction details. Without these numbers, the reported results cannot be reproduced.

free parameters (5)
  • gamma
    Hard-negative sampling threshold in Eq.(12), 'a tunable threshold'; no value is reported.
  • lambda_node / lambda_graph
    Loss weights in Eq.(13), tuned on the validation set; values not reported.
  • tau_n / tau_g
    Temperature parameters in Eqs.(10) and (11); values not reported.
  • number of LDA topics
    Topic interaction graph construction in Section 3.2 requires choosing topic count; not specified.
  • GIN layers and hidden dimension
    Architecture hyperparameters in Section 3.3 (L, d); values not listed.
axioms (3)
  • domain assumption Semantic equivalence often manifests through isomorphic or similar substructures
    Invoked in Section 1 to justify graph alignment; this is a domain belief about legal/plagiarism text.
  • domain assumption Dependency parsing and LDA topic modeling yield accurate graphs
    The dual-graph construction in Section 3.2 assumes these tools produce useful structures.
  • domain assumption Aligned node pairs M can be recovered by heuristics
    Eq.(10) and Section 3.4 assume a meaningful set M exists; no concrete method is given.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of StructCoh: Structured Contrastive Learning for Context-Aware Text Semantic Matching." pith.science (2026). https://pith.science/paper/6NO3SYX3

@misc{pith2026250902033,
  author       = {Pith},
  title        = {Pith review of: StructCoh: Structured Contrastive Learning for Context-Aware Text Semantic Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NO3SYX3}},
  note         = {Machine review of arXiv:2509.02033}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text semantic matching requires nuanced understanding of both structural relationships and fine-grained semantic distinctions. While pre-trained language models excel at capturing token-level interactions, they often overlook hierarchical structural patterns and struggle with subtle semantic discrimination. In this paper, we proposed StructCoh, a graph-enhanced contrastive learning framework that synergistically combines structural reasoning with representation space optimization. Our approach features two key innovations: (1) A dual-graph encoder constructs semantic graphs via dependency parsing and topic modeling, then employs graph isomorphism networks to propagate structural features across syntactic dependencies and cross-document concept nodes. (2) A hierarchical contrastive objective enforces consistency at multiple granularities: node-level contrastive regularization preserves core semantic units, while graph-aware contrastive learning aligns inter-document structural semantics through both explicit and implicit negative sampling strategies. Experiments on three legal document matching benchmarks and academic plagiarism detection datasets demonstrate significant improvements over state-of-the-art methods. Notably, StructCoh achieves 86.7% F1-score (+6.2% absolute gain) on legal statute matching by effectively identifying argument structure similarities.

Figures

Figures reproduced from arXiv: 2509.02033 by Chao Xue, Ziyuan Gao.

Figure 1
Figure 1. Figure 1: An example of the effectiveness of StructCoh in structural understanding in legal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparison of StructCoh and baselines. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 51 canonical work pages

  1. [1]

    COLIEE 2023 Summary: Legal Information Extrac- tion/Entailment and Court Judgment Prediction

    R. Kano, M. Yoshioka, Y . Li et al. "COLIEE 2023 Summary: Legal Information Extrac- tion/Entailment and Court Judgment Prediction". In Proceedings of the 16th International Competition on Legal Information Extraction/Entailment, 2023

    work page 2023

  2. [2]

    Question calibration and multi-hop modeling for temporal question answering,

    C. Xue, D. Liang, P. Wang, and J. Zhang, “Question calibration and multi-hop modeling for temporal question answering,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 19 332–19 340

    work page 2024

  3. [3]

    Academic Plagiarism Detection via Contrastive Credibility Analysis

    S. J. Pan, J. Yang, C. Zhou et al. "Academic Plagiarism Detection via Contrastive Credibility Analysis". Transactions on Asian and Low-Resource Language Information Processing

    work page

  4. [4]

    Not all parameters are created equal: Smart isolation boosts fine-tuning performance,

    Y . Wang, D. Liang, and M. Peng, “Not all parameters are created equal: Smart isolation boosts fine-tuning performance,” arXiv preprint arXiv:2508.21741, 2025

    arXiv 2025

  5. [5]

    Cqg: A simple and effective controlled generation framework for multi-hop question generation,

    Z. Fei, Q. Zhang, T. Gui, D. LiangacOS Automation!, S. Wang, W. Wu, and X.-J. Huang, “Cqg: A simple and effective controlled generation framework for multi-hop question generation,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2022, pp. 6896–6906

    work page 2022

  6. [6]

    Dabert: Dual attention enhanced bert for semantic matching,

    S. Wang, D. Liang, J. Song, Y . Li, and W. Wu, “Dabert: Dual attention enhanced bert for semantic matching,” in Proceedings of the 29th International Conference on Computational Linguistics, 2022, pp. 1645–1654

    work page 2022

  7. [7]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    J. Devlin, M.-W. Chang, K. Lee et al. "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding". In Proceedings of NAACL-HLT, pp. 4171-4186, 2019

    work page 2019

  8. [8]

    Local and global: temporal question answering via information fusion,

    Y . Liu, M. L. Di Liang, F. Giunchiglia, X. Li, S. Wang, W. Wu, L. Huang, X. Feng, and R. Guan, “Local and global: temporal question answering via information fusion,” in Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, 2023, pp. 5141–5149

    work page 2023

  9. [9]

    Syntax-Aware Graph Neural Networks for Semantic Role Labeling

    L. Wang, Y . Liu, X. Huang et al. "Syntax-Aware Graph Neural Networks for Semantic Role Labeling". In Findings of EMNLP, pp. 3918-3932, 2022

    work page 2022

  10. [10]

    SimCSE: Simple Contrastive Learning of Sentence Embeddings

    T. Gao, X. Yao, D. Chen. "SimCSE: Simple Contrastive Learning of Sentence Embeddings". In Proceedings of EMNLP, pp. 6894-6910, 2021

    work page 2021

  11. [11]

    Time-aware multiway adaptive fusion network for temporal knowledge graph question answering,

    Y . Liu, D. Liang, F. Fang, S. Wang, W. Wu, and R. Jiang, “Time-aware multiway adaptive fusion network for temporal knowledge graph question answering,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    work page 2023

  12. [12]

    Corpus-based and knowledge-based measures of text semantic similarity

    Rada Mihalcea, Courtney Corley, and Carlo Strapparava. Corpus-based and knowledge-based measures of text semantic similarity. AAAI, 2006

    work page 2006

  13. [13]

    Siamese recurrent architectures for learning sentence similarity

    Jonas Mueller and Aditya Thyagarajan. Siamese recurrent architectures for learning sentence similarity. In AAAI, 2016

    work page 2016

  14. [14]

    Convolutional neural network architectures for matching natural language sentences

    Baotian Hu, Zhengdong Lu, Hang Li, and Qingcai Chen. Convolutional neural network architectures for matching natural language sentences. NeurIPS, 2014

    work page 2014

  15. [15]

    BERT: Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2019

    2019

  16. [16]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In EMNLP, 2019

    work page 2019

  17. [17]

    Artificial Intelligence and Law: Risks, Research, and Limits

    Anne-Laure Collomb, Alexis Conesa-Bes, and Louis Lautman. Artificial Intelligence and Law: Risks, Research, and Limits. Artificial Intelligence and Law, 28, 2020

    work page 2020

  18. [18]

    Learning deep representations of fine-grained scientific knowledge

    Nanyun Peng, Hoifung Poon, Chris Quirk, Kristina Toutanova, and Wen-tau Yih. Learning deep representations of fine-grained scientific knowledge. In ACL, 2018

    work page 2018

  19. [19]

    Graph neural networks for natural language processing: A survey

    Qipeng Guo, Xinyan Xiao, Tengfei Ma, et al. Graph neural networks for natural language processing: A survey. In IJCAI, 2020

    work page 2020

  20. [20]

    Encoding sentences with graph convolutional networks for semantic role labeling

    Diego Marcheggiani and Ivan Titov. Encoding sentences with graph convolutional networks for semantic role labeling. In EMNLP, 2017. StructCoh 15

    work page 2017

  21. [21]

    Graph convolutional networks for text classifica- tion

    Luyu Yao, Chengsheng Mao, and Yuan Luo. Graph convolutional networks for text classifica- tion. In AAAI, 2019

    work page 2019

  22. [22]

    Connecting the dots: Document- level relation extraction with edge-oriented graphs

    Fenia Christopoulou, Makoto Miwa, and Sophia Ananiadou. Connecting the dots: Document- level relation extraction with edge-oriented graphs. In EMNLP, 2020

    work page 2020

  23. [23]

    Hierarchical graph network for document-level relation extraction

    Zhongcheng Li, Yicheng Zou, Rui Liu, and Yan Song. Hierarchical graph network for document-level relation extraction. In NLPCC, 2022

    work page 2022

  24. [24]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020

    2020

  25. [25]

    DABERT: Dual attention enhanced BERT for semantic matching,

    S. Wang, D. Liang, J. Song, Y . Li, and W. Wu, “DABERT: Dual attention enhanced BERT for semantic matching,” in Proceedings of the 29th International Conference on Computa- tional Linguistics.Gyeongju, Republic of Korea: International Committee on Computational Linguistics, Oct. 2022, pp. 1645–1654. [Online]

    work page 2022

  26. [26]

    SimCSE: Simple contrastive learning of sentence embeddings

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. SimCSE: Simple contrastive learning of sentence embeddings. In EMNLP, 2021

    work page 2021

  27. [27]

    Aligning visual regions and textual concepts for semantic- grounded image representations

    Yikang Yan, Qianqian Xie, et al. Aligning visual regions and textual concepts for semantic- grounded image representations. In CVPR, 2021

    work page 2021

  28. [28]

    Asynchronous deep interaction network for natural language inference,

    D. Liang, F. Zhang, Q. Zhang, and X.-J. Huang, “Asynchronous deep interaction network for natural language inference,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, pp. 2692–2700

    work page 2019

  29. [29]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR, 2017

    work page 2017

  30. [30]

    Dual path modeling for semantic matching by perceiving subtle conflicts,

    C. Xue, D. Liang, S. Wang, J. Zhang, and W. Wu, “Dual path modeling for semantic matching by perceiving subtle conflicts,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

    work page 2023

  31. [31]

    Graph Attention Networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, et al. Graph Attention Networks. In ICLR, 2018

    work page 2018

  32. [32]

    Yuhao Zhang, Peng Qi, and Christopher D. Manning. Graph Convolution over Pruned Dependency Trees Improves Relation Extraction. In EMNLP, 2018

    work page 2018

  33. [33]

    Content diffusion prediction with deep learning and graph convolutional network

    Zhilin Liu, Yichong Xu, et al. Content diffusion prediction with deep learning and graph convolutional network. In WSDM, 2018

    work page 2018

  34. [34]

    Local and global: Text matching via syntax graph calibration,

    L. Li, Q. Liao, M. Lai, D. Liang, and S. Liang, “Local and global: Text matching via syntax graph calibration,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 11 571–11 575

    work page 2024

  35. [35]

    A Structure-aware Graph Neural Network for Legal Case Retrieval

    Jie Wang, Haoteng Tang, Chaojun Xiao, Minghui Qiu, and Jun Huang. A Structure-aware Graph Neural Network for Legal Case Retrieval. In AAAI, 2022

    work page 2022

  36. [36]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang

    work page

  37. [37]

    Resolving word vagueness with scenario-guided adapter for natural language inference,

    Y . Liu, M. Li, D. Liang, X. Li, F. Giunchiglia, L. Huang, X. Feng, and R. Guan, “Resolving word vagueness with scenario-guided adapter for natural language inference,” arXiv preprint arXiv:2405.12434, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    Graph convolutional networks for text classification

    Lei Yao, Chengsheng Mao, and Yuan Luo. Graph convolutional networks for text classification. In Proceedings of the 2019 AAAI Conference on Artificial Intelligence

    work page 2019

  39. [39]

    Heterogeneous graph transformer for graph-to-text generation

    Zhitao Hu, Bowen Tan, and Jiebo Luo. Heterogeneous graph transformer for graph-to-text generation. In NeurIPS, pages 1-12, 2020

    work page 2020

  40. [40]

    Robust lottery tickets for pre-trained language models,

    R. Zheng, R. Bao, Y . Zhou, D. Liang, S. Wang, W. Wu, T. Gui, Q. Zhang, and X. Huang, “Robust lottery tickets for pre-trained language models,”arXiv preprint arXiv:2211.03013

    Pith/arXiv arXiv

  41. [41]

    Modeling coreference relations with graph neural networks

    Yujia Wu, Yiming Yang, and Dinghan Shen. Modeling coreference relations with graph neural networks. In EMNLP, pages 1-11, 2021

    work page 2021

  42. [42]

    Semantics-aware BERT for language understanding

    Hongming Zhang, Zhao Yan, and Yangqiu Song. Semantics-aware BERT for language understanding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 8755–8770, 2020. 16 C. Xue and Z. Gao

    work page 2020

  43. [43]

    Using UER for pretraining tasks on different languages

    Zhang Xia, Jie Wang, and Lili Xu. Using UER for pretraining tasks on different languages. In Proceedings of the 2021 Conference on Computational Linguistics, pages 1-9, 2021

    work page 2021

  44. [44]

    Improving semantic matching through dependency-enhanced pre-trained model with adaptive fusion,

    J. Song, D. Liang, R. Li, Y . Li, S. Wang, M. Peng, W. Wu, and Y . Yu, “Improving semantic matching through dependency-enhanced pre-trained model with adaptive fusion,” inFindings of the Association for Computational Linguistics: EMNLP 2022. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics, Dec. 2022, pp. 45–57. [Online]

    work page 2022

  45. [45]

    Searching for optimal subword tokenization in cross-domain ner,

    R. Ma, Y . Tan, X. Zhou, X. Chen, D. Liang, S. Wang, W. Wu, T. Gui, and Q. Zhang, “Searching for optimal subword tokenization in cross-domain ner,”arXiv preprint arXiv:2206.03352, 2022

    Pith/arXiv arXiv 2022

  46. [46]

    Enhanced LSTM for natural language inference

    Qian Chen, Xiaodan Zhu, Zhenhua Ling, Si Wei, and Hui Jiang. Enhanced LSTM for natural language inference. In Proceedings of the 2016 Annual Conference on Neural Information Processing Systems, pages 1557-1567, 2016

    work page 2016

  47. [47]

    Unleashing poten- tial of evidence in knowledge-intensive dialogue generation,

    X. Wu, J. Yang, T. Li, S. Zhang, Y . Du, L. Chai, D. Liang, and Z. Li, “Unleashing poten- tial of evidence in knowledge-intensive dialogue generation,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    work page 2025

  48. [48]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998–6008

    work page

  49. [49]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. In Neural Computation, volume 9(8), pages 1735–1780, 1997

    work page 1997

  50. [50]

    Transferring from formal newswire domain with hypernet for twitter pos tagging,

    T. Gui, Q. Zhang, J. Gong, M. Peng, D. Liang, K. Ding, and X.-J. Huang, “Transferring from formal newswire domain with hypernet for twitter pos tagging,” in Proceedings of the 2018 conference on empirical methods in natural language processing, 2018, pp. 2540–2549

    work page 2018

  51. [51]

    Bilateral multi-perspective matching for natural language sentences

    Zhiguo Wang, Wael Hamza, and Radu Florian. Bilateral multi-perspective matching for natural language sentences. In NeurIPS, pages 4144–4152, 2017

    work page 2017

  52. [52]

    Comateformer: Combined attention transformer for semantic sentence matching,

    B. Li, D. Liang, and Z. Zhang, “Comateformer: Combined attention transformer for semantic sentence matching,” arXiv preprint arXiv:2412.07220, 2024

    Pith/arXiv arXiv 2024

  53. [53]

    Compare, compress and propagate: Enhance semantic matching in sentences

    Yi Tay, Anh Tuan Luu, and Siu Cheung Hui. Compare, compress and propagate: Enhance semantic matching in sentences. In ACL, pages 1377–1386, 2017

    work page 2017

  54. [54]

    Tablebench: A comprehensive and complex benchmark for table question answering,

    X. Wu, J. Yang, L. Chai, G. Zhang, J. Liu, X. Du, D. Liang, D. Shu, X. Cheng, T. Sunet al., “Tablebench: A comprehensive and complex benchmark for table question answering,”arXiv preprint arXiv:2408.09174, 2024

    Pith/arXiv arXiv 2024

  55. [55]

    DistilBERT, a distilled version of BERT: Smaller, faster, cheaper and lighter

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. DistilBERT, a distilled version of BERT: Smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019

    Pith/arXiv arXiv 1910

  56. [56]

    ALBERT: A Lite BERT for self-supervised learning of language representations

    Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. ALBERT: A Lite BERT for self-supervised learning of language representations. In ICLR, pages 1–13, 2020

    work page 2020

  57. [57]

    SyntaxBERT: Improving pre-trained trans- formers with syntax information

    Guangyuan Bai, Jiasen Zhang, and Yue Zhang. SyntaxBERT: Improving pre-trained trans- formers with syntax information. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2035–2044, 2021

    work page 2021

  58. [58]

    Lam, Makoto Kishida, Lesley Sterling, Makoto Miwa

    Rabelo, Rafael, Monica S. Lam, Makoto Kishida, Lesley Sterling, Makoto Miwa. A Combined Approach for Question Answering in the Legal Domain. In ICAIL, 2019

    work page 2019

  59. [59]

    Adaptive multi-attention network incorporating answer information for duplicate question detection,

    D. Liang, F. Zhang, W. Zhang, Q. Zhang, J. Fu, M. Peng, T. Gui, and X. Huang, “Adaptive multi-attention network incorporating answer information for duplicate question detection,” in Proceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval, 2019, pp. 95–104

    work page 2019

  60. [60]

    Plagiarism Detec- tion and Beyond: Academic Text Rewriting, Large-scale Corpus, and Paraphrase Modeling

    Pan, Liang, Qian Li, Tong Xiao, Zhirui Zhang, Changliang Li, Jingbo Zhu. Plagiarism Detec- tion and Beyond: Academic Text Rewriting, Large-scale Corpus, and Paraphrase Modeling. In ACL, 2021

    work page 2021

  61. [61]

    Progressive Mastery: Customized Cur- riculum Learning with Guided Prompting for Mathematical Reasoning

    Wu, Muling, Qian, Qi, Liu, Wenhao, Wang, Xiaohua, Huang, Zisu, Liang, Di, Miao, LI, Dou, Shihan, Lv, Changze, Wang, Zhenghua, et al. Progressive Mastery: Customized Cur- riculum Learning with Guided Prompting for Mathematical Reasoning . arXiv preprint arXiv:2506.04065, 2025

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.