Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

DCG-SQL: Enhancing In-Context Learning for Text-to-SQL with Deep Contextual Schema Link Graph

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that representing a text-to-SQL question together with a pruned, attention-linked database schema graph is what makes in-context demonstration retrieval work, and that this improves SQL generation for both GPT-4-class…

desk verdict A solid, well-ablated retrieval method for text-to-SQL that actually helps small LLMs, but the teacher-model overlap and unvalidated thresholds are the soft spots to probe. read the letter →

arxiv 2505.19956 v2 pith:MTAO4ZFJ submitted 2025-05-26 cs.AI cs.CL

Jihyung Lee , Jin-Seop Lee , Jaehoon Lee , YunSeok Choi , Jee-Hyong Lee This is my paper
classification cs.AIcs.CL
keywords text-to-SQLin-contextlearningdemonstrationretrievalschemalinkingpruninggraphrepresentationcontrastiveSpiderbenchmark
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 argues that previous demonstration retrieval for text-to-SQL fails because it encodes the question alone or encodes the schema superficially, so retrieved examples are no better than random. It proposes a deep contextual schema link graph: first prune the database schema to the tables and columns a question actually needs, then link question tokens to those schema items using attention scores from the pruning model, and encode the resulting graph into an embedding for retrieval. Training the graph encoder with contrastive labels derived from how well a teacher LLM generates the gold SQL when given each candidate demonstration yields a retriever that consistently improves execution and exact-match accuracy across GPT-4-class models and small 3B-8B models. If true, this means the question-schema pair, not just the question, is the right unit of similarity for in-context SQL generation, and that smaller LLMs can be made more competitive by retrieving demonstrations that match the target SQL structure.

What carries the argument

The deep contextual schema link graph. A RoBERTa-large cross-encoder takes the question concatenated with the database schema, where each table and column is marked with [TAB] and [COL] special tokens, and binary classification heads on those special tokens prune irrelevant schema items. Because the pruning model is trained to know which items matter, its attention scores between question tokens and the special token of a relevant schema item become the predicted schema links: a link is added when the attention score exceeds a threshold (0.66 for tables, 0.43 for columns). The resulting graph, with nodes for question tokens and surviving schema items and edges for attention-match and syntactic links, is embedded by a bi-encoder built on a Relation-Aware Transformer initialized from BERT-base. The encoder is trained with contrastive learning where positives and negatives come from ranking candidate demonstrations by the likelihood the gold SQL receives under Llama 3.1-8B when that candidate is supplied as a demonstration, with candidates pre-filtered by tree-edit distance over abstract syntax trees to keep training cheap. At inference, the top-k nearest graph embeddings are retrieved and used as demonstrations with automated chain-of-thought prompts.

What would settle it

A concrete experiment: fix the pruned nodes and the contrastive training, but replace the attention-based question-to-schema edges with random edges during graph-embedding training and retrieval; if execution accuracy on Spider stays near 82 with Llama 3.1-8B, the schema-link edges are not doing the work. Alternatively, relabel the contrastive training using likelihoods from a smaller teacher, such as a 3B model, and if the retriever's gains over random demonstrations vanish, the teacher-likelihood assumption is load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that a graph that prunes irrelevant database schema items and adds edges between question tokens and schema items according to the attention of a schema-pruning classifier is a better representation for retrieving demonstrations than question-text embeddings, masked question embeddings, question-plus-schema text embeddings, or full-schema graphs. On Spider, the method reports execution accuracy of 87.5 with GPT-4, 82.1 with Llama 3.1-8B, and 74.7 with Llama 3.2-3B, consistently above the compared baselines, with a 21.3% relative exact-match gain and a 9.5% relative execution-accuracy gain for Llama 3.1-8B over ACT-SQL. The same retrieval used with DAIL-SQL's prompt template improves DAIL-SQL from 71.2 to 75.6 execution accuracy, so the gains are not tied to the authors' prompt template. The paper also reports that retrieved demonstrations are much closer to the target SQL in tree-edit distance (13.82 vs 27.17 for ACT-SQL and 22.38 for DAIL-SQL), supporting the claim that graph-based retrieval finds useful patterns.

Load-bearing premise

The whole pipeline rests on using the likelihood that Llama 3.1-8B assigns to the gold SQL as the measure of whether a demonstration is useful, and if that teacher signal does not transfer to other target models, the retrieved demonstrations will not be the useful ones.

Editorial extensions

If this is right

  • Demonstration retrieval for text-to-SQL should be evaluated on joint question-schema representations rather than question embeddings alone.
  • Small LLMs can benefit substantially from better retrieval: on Spider, execution accuracy for Llama 3.2-3B rises to 74.7 compared with 66.0 for ACT-SQL.
  • The graph retriever transfers to other prompt templates, since using it with DAIL-SQL's prompt improves execution accuracy from 71.2 to 75.6.
  • The method is efficient at inference: it needs one LLM call and lightweight retrieval modules totaling about 0.5B trainable parameters, with 1.3 seconds retrieval latency.
  • The approach generalizes to Spider variants and the BIRD dataset, where questions mask or rename schema words, because the linking is contextual rather than lexical.

Reading between the lines

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

  • Beyond the paper, a direct test of the mechanism would be to randomize or corrupt the attention-based edges while keeping the pruned nodes; if retrieval accuracy stays high, the schema-link edges are not the carrier of the gain.
  • Beyond the paper, the teacher-likelihood labeling means the retriever inherits the SQL-generation preferences of Llama 3.1-8B, so an ensemble of teachers or a stronger teacher might yield demonstrations that transfer even better to small target models.
  • Beyond the paper, because the pruning and linking labels are derived from the target SQL without explicit schema-linking annotations, the same recipe could be applied to other schema-structured generation tasks such as code generation from API documentation or data-to-text generation.
  • Beyond the paper, a testable extension would be to vary only the teacher LLM used for likelihood labels and measure which teacher produces the most useful demonstrations for a fixed target model and prompt template.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes DCG-SQL, a demonstration-retrieval method for in-context text-to-SQL. It constructs a 'Deep Contextual Schema Link Graph' by pruning irrelevant schema items with a RoBERTa-based classifier, linking question tokens to schema items via attention scores above hand-set thresholds, and embedding the resulting graph with a Relation-Aware Transformer trained by contrastive learning. Demonstration usefulness labels are generated by scoring candidate demonstrations with the likelihood of the gold SQL under Llama 3.1-8B (Eq. 1, Section 3.2). At inference, top-k nearest graph embeddings are retrieved and fed to an LLM with automated chain-of-thought prompts. Experiments are reported on Spider and its variants, plus BIRD, across six LLMs of different sizes, with claims of consistent EX/EM improvements, transfer to DAIL-SQL, and lower retrieval latency.

Significance. If the central claim is fully supported, the paper would make a solid empirical contribution: it tackles the schema-awareness problem in demonstration retrieval, provides a graph representation that jointly encodes question and schema, and reports gains across both hyper-scaled and small LLMs. The strengths include public code release, a broad evaluation across six LLMs and four Spider-derived benchmarks plus BIRD, the transferability experiment in Table 7, and the efficiency comparison in Table 9. However, the significance is conditional on validating the teacher-model proxy that generates the training signal for the retriever and on resolving the inconsistency in the 'consistent improvement' claim.

major comments (4)
  1. [Section 3.2, Eq. (1)] The training signal for the retrieval model is the likelihood of the gold SQL under Llama 3.1-8B, score(x, si) = P_sLLM(SQLx | Gx; si). This assumes that a high likelihood under Llama 3.1-8B is a valid proxy for demonstration usefulness for all target LLMs, including GPT-4, GPT-3.5, DeepSeek-Coder, and Llama 3.2-3B. The paper provides no validation of this proxy: no teacher ablation, no comparison of teacher rankings with rankings induced by other target models, and no analysis of ranking agreement. The cross-model gains reported in Tables 2 and 3 are suggestive but do not isolate this dependency, because the same graph construction and retrieval module are used everywhere; the gains could come from schema pruning or prompt design rather than from teacher-ranked retrieval. Since Llama 3.1-8B is also one of the evaluation models, the largest gains on that model may partly reflect a self-referential advantage. The Limitations section does not acknowledge this dependency, which is the most load-bearing assumption in the pipeline.
  2. [Section 4.1, schema linking thresholds] The attention thresholds for schema linking, tau_tab=0.66 and tau_col=0.43, are introduced without a validation procedure or sensitivity analysis. The graph structure, and therefore every graph embedding used for retrieval, depends directly on these thresholds. If they were tuned on the Spider development set, the reported results on Spider-DK, Spider-Realistic, Spider-Syn, and BIRD may inherit selection bias. Please report how the thresholds were selected, provide a sensitivity analysis over a range of threshold values, and state the resulting average graph size (number of nodes and edges) for the chosen thresholds.
  3. [Section 4.4 and Table 3] The claim that the method 'consistently achieves higher execution accuracy and exact match accuracy compared to existing approaches' is not literally supported by the reported numbers. On Spider-DK with GPT-4, DCG-SQL obtains EX 71.6, which is lower than both ACT-SQL (72.0) and ASTRES (72.3). Other close comparisons in Table 3 are within about one point (e.g., GPT-3.5 Spider-DK EX 69.0 vs. ASTRES 68.8). Because the tables report single-run results without error bars or significance tests, it is difficult to determine whether such differences are meaningful. Please either qualify the consistency claim, explain the Spider-DK GPT-4 result, or provide statistical support for the claimed improvements.
  4. [Section 3.1, schema pruning supervision] The schema pruning classifier is trained 'with the question and the schema as input, and the SQL to indicate relevant schema items,' but the conversion from a gold SQL query to binary relevance labels for tables and columns is not specified. It is unclear whether columns used only in joins, values, or subqueries are labeled positive, and whether negative examples are mined from the same database. This matters because the pruning model determines which schema items enter the graph, and Table 6 reports pruning F1 on the training distribution rather than on downstream retrieval quality. Please clarify the labeling procedure and report the effect of pruning quality on final execution accuracy (e.g., by comparing against an oracle-pruned graph).
minor comments (5)
  1. [Abstract and Introduction] The abstract contains a malformed code link ('https://github.com/jjklle/DCG-SQL}{https://github.com/jjklle/DCG-SQL'), and the introduction has a duplicated sentence fragment ('encoded through a unified representation.'). There is also a typo: 'randomly choosen' in Table 1 and the text.
  2. [Equation (1), Section 3.2] The likelihood computation is underspecified: it is not stated whether the probability is token-level normalized, whether a length penalty is applied, or whether the prompt format used to compute the likelihood matches the inference prompt format. Since LLM likelihoods can be sensitive to these choices, please provide the exact serialization of Gx and the decoding/tokenization details.
  3. [Table 10] The numeric formatting in Table 10 is inconsistent: '38.40' and '39.38' use two decimals, while other entries use one decimal. Please unify the formatting and double-check the reported values.
  4. [Section 4.5, Table 5] The four retrieval baselines in Table 5 are described only in prose. A precise specification of how 'masked question text embedding' and 'existing graph representation' are constructed, including any hyperparameters, would improve reproducibility.
  5. [References and Appendix] The spaCy reference is listed as 'To appear'; please update it with the appropriate version or technical report. Also, the automated-CoT templates in Appendix C.2 are a meaningful design choice, and a short ablation separating the retrieval effect from the CoT template effect would strengthen the causal claims.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the central claim is grounded in held-out execution accuracy, with a mild self-referential teacher-labeling choice that is a robustness risk rather than a circular reduction.

full rationale

The derivation chain is not circular. The schema-pruning classifier and graph encoder are trained on the Spider training set, and all headline results are measured on the held-out Spider dev set (and BIRD dev in Appendix A) with execution and exact-set-match accuracy, so the final numbers are not entailed by the training objectives. The only self-referential element is in Section 3.2, where demonstration usefulness labels are computed as score(x, s_i) = P_sLLM(SQL_x | G_x; s_i) using Llama 3.1-8B, which is also one of the evaluation models. This is a training-label choice, not an equation-level identity: the evaluation metric (execution accuracy of generated SQL) is distinct from the likelihood objective, and the retriever must generalize to held-out anchors through graph embeddings. The paper does not ablate the teacher model, and the limitations section does not mention this dependency, so the cross-model transfer claim (to GPT-4, GPT-3.5, DeepSeek, Llama 3.2-3B) rests on an unvalidated proxy; that is a correctness/robustness gap, not circularity. No load-bearing self-citations appear: prior-work citations are external, and the graph/schema components are isolated by ablations (Table 5, Table 7). The low-TED analysis in Table 8 is partly anticipated by the AST-similar candidate filtering used for training labels, but inference-time retrieval does not use AST distance, so this is a supportive analysis rather than a forced prediction. Overall, the central claim has independent empirical content and is externally benchmarked.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or ontological entities. Its load-bearing assumptions are the usefulness of attention scores for linking, the validity of Llama 3.1-8B likelihoods as training labels, the completeness of gold-SQL-derived pruning targets, and the manually set attention thresholds. These are empirical modeling choices rather than derived constants.

free parameters (3)
  • attention thresholds = tau_tab=0.66, tau_col=0.43
    Thresholds for deciding when an attention score creates a link between a question token and a schema item. Stated in Section 4.1 without a validation procedure, so they may have been tuned on the Spider dev set.
  • number of demonstrations = 5
    Five demonstrations are used for all experiments. This is a standard few-shot choice but is not justified or swept.
  • candidate set size = 100
    For each training anchor, 100 structurally similar SQL candidates are selected by tree edit distance before LLM scoring; this limit is a design choice that constrains which demonstrations can be labeled as positive.
assumptions (4)
  • domain assumption Attention scores from the schema-pruning cross-encoder reflect semantic relevance between question tokens and schema items.
    Section 3.1 uses attention scores above a threshold to construct graph edges without a schema-linking dataset. This assumes the classifier's attention weights encode the linking relation.
  • ad hoc to paper Likelihood of the gold SQL under Llama 3.1-8B is a valid measure of demonstration usefulness.
    Eq. 1 in Section 3.2 defines positive and negative demonstrations by this likelihood. The proxy is model-specific and may not transfer to other LLMs.
  • domain assumption Tree-edit-distance-based candidate selection over abstract syntax trees yields the most useful demonstrations.
    Section 3.2 restricts the candidate pool to 100 structurally similar SQLs before LLM scoring, assuming AST similarity correlates with demonstration usefulness.
  • domain assumption Spider training-set gold SQL queries provide complete ground truth for which schema items are relevant.
    The schema pruning classifier is trained with binary labels derived from the gold SQL. This assumes every relevant schema item appears in the gold SQL and that items not in the SQL are irrelevant.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DCG-SQL: Enhancing In-Context Learning for Text-to-SQL with Deep Contextual Schema Link Graph." pith.science (2026). https://pith.science/paper/MTAO4ZFJ

@misc{pith2026250519956,
  author       = {Pith},
  title        = {Pith review of: DCG-SQL: Enhancing In-Context Learning for Text-to-SQL with Deep Contextual Schema Link Graph},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MTAO4ZFJ}},
  note         = {Machine review of arXiv:2505.19956}
}
read the original abstract

Text-to-SQL, which translates a natural language question into an SQL query, has advanced with in-context learning of Large Language Models (LLMs). However, existing methods show little improvement in performance compared to randomly chosen demonstrations, and significant performance drops when smaller LLMs (e.g., Llama 3.1-8B) are used. This indicates that these methods heavily rely on the intrinsic capabilities of hyper-scaled LLMs, rather than effectively retrieving useful demonstrations. In this paper, we propose a novel approach for effectively retrieving demonstrations and generating SQL queries. We construct a Deep Contextual Schema Link Graph, which contains key information and semantic relationship between a question and its database schema items. This graph-based structure enables effective representation of Text-to-SQL samples and retrieval of useful demonstrations for in-context learning. Experimental results on the Spider benchmark demonstrate the effectiveness of our approach, showing consistent improvements in SQL generation performance and efficiency across both hyper-scaled LLMs and small LLMs. The code is available at https://github.com/jjklle/DCG-SQL}{https://github.com/jjklle/DCG-SQL.

Figures

Figures reproduced from arXiv: 2505.19956 by the authors.

Figure 1
Figure 1. Overview of our proposed method for Text-to-SQL. The method consists of three main steps: (1) Deep Contextual Schema Link Graph Construction, which prunes irrelevant schema items and builds a schema link graph that captures the contextual relationship between the question tokens and the schema items; (2) Graph-based Demonstration Retrieval, where graph embeddings are used to retrieve demonstrations relevant to the g… view at source ↗
Figure 2
Figure 2. Deep Contextual Schema Link Graph Construction. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results. These results demonstrate a comparison between our generated queries and those from [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results. These results demonstrate a comparison between our generated queries and those from [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Sign in 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

  1. EcoTable: Cost-effective Table Integration in Data Lakes for Natural Language Queries

    cs.DB 2026-06 unverdicted novelty 6.0 of 10

    Query-driven table integration that uses Steiner-tree search to choose which joins LLMs must verify, reporting 30%+ accuracy gains at 5x lower LLM cost.

Reference graph

Works this paper leans on

42 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Steven Bird and Edward Loper. 2004. https://aclanthology.org/P04-3031 NLTK : The natural language toolkit . In Proceedings of the ACL Interactive Poster and Demonstration Sessions , pages 214--217, Barcelona, Spain. Association for Computational Linguistics

    work page 2004

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  5. [5]

    Shuaichen Chang and Eric Fosler-Lussier. 2023. How to prompt llms for text-to-sql: A study in zero-shot, single-domain, and cross-domain settings. arXiv preprint arXiv:2305.11853

    arXiv 2023

  6. [6]

    DeepSeek-AI, :, Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, Huazuo Gao, et al. 2024. https://arxiv.org/abs/2401.02954 Deepseek llm: Scaling open-source language models with longtermism . Preprint, arXiv:2401.02954

    arXiv 2024

  7. [7]

    Xiang Deng, Ahmed Hassan Awadallah, Christopher Meek, Oleksandr Polozov, Huan Sun, and Matthew Richardson. 2020. Structure-grounded pretraining for text-to-sql. arXiv preprint arXiv:2010.12773

    arXiv 2020

  8. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long a...

    doi:10.18653/v1/n19-1423 2019

Show all 42 references
  1. [9]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  2. [10]

    Andrea Galassi, Marco Lippi, and Paolo Torroni. 2020. Attention in natural language processing. IEEE transactions on neural networks and learning systems, 32(10):4291--4308

    2020

  3. [11]

    Yujian Gan, Xinyun Chen, Qiuping Huang, Matthew Purver, John R Woodward, Jinxia Xie, and Pengsheng Huang. 2021 a . Towards robustness of text-to-sql models against synonym substitution. arXiv preprint arXiv:2106.01065

    2021 arXiv

  4. [12]

    Yujian Gan, Xinyun Chen, and Matthew Purver. 2021 b . Exploring underexplored limitations of cross-domain text-to-sql generalization. arXiv preprint arXiv:2109.05157

    2021 arXiv

  5. [13]

    Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. 2024. https://doi.org/10.14778/3641204.3641221 Text-to-sql empowered by large language models: A benchmark evaluation . Proc. VLDB Endow., 17(5):1132–1145

    2024

  6. [14]

    Chunxi Guo, Zhiliang Tian, Jintao Tang, Pancheng Wang, Zhihua Wen, Kang Yang, and Ting Wang. 2024. Prompting gpt-3.5 for text-to-sql with de-semanticization and skeleton retrieval. In PRICAI 2023: Trends in Artificial Intelligence, pages 262--274, Singapore. Springer Nature Singapore

    2024

  7. [15]

    Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, Jian-Guang Lou, Ting Liu, and Dongmei Zhang. 2019. https://doi.org/10.18653/v1/P19-1444 Towards complex text-to- SQL in cross-domain database with intermediate representation . In Proceedings of the 57th Annual Meeting of the Associa...

    2019 doi

  8. [16]

    Matthew Honnibal and Ines Montani. 2017. spaCy 2 : Natural language understanding with B loom embeddings, convolutional neural networks and incremental parsing. To appear

    2017

  9. [17]

    Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen tau Yih. 2020. https://arxiv.org/abs/2004.04906 Dense passage retrieval for open-domain question answering . Preprint, arXiv:2004.04906

    2020 arXiv

  10. [18]

    Dongjun Lee, Choongwon Park, Jaehyuk Kim, and Heesoo Park. 2025. https://aclanthology.org/2025.coling-main.24/ MCS - SQL : Leveraging multiple prompts and multiple-choice selection for text-to- SQL generation . In Proceedings of the 31st International Conference on Computation...

    2025

  11. [19]

    Wenqiang Lei, Weixin Wang, Zhixin Ma, Tian Gan, Wei Lu, Min-Yen Kan, and Tat-Seng Chua. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.564 Re-examining the role of schema linking in text-to- SQL . In Proceedings of the 2020 Conference on Empirical Methods in Natural Languag...

    2020 doi

  12. [20]

    Haoyang Li, Jing Zhang, Cuiping Li, and Hong Chen. 2023 a . https://doi.org/10.1609/aaai.v37i11.26535 Resdsql: decoupling schema linking and skeleton parsing for text-to-sql . In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conf...

    2023 doi

  13. [21]

    Jinyang Li, Binyuan Hui, Reynold Cheng, Bowen Qin, Chenhao Ma, Nan Huo, Fei Huang, Wenyu Du, Luo Si, and Yongbin Li. 2023 b . Graphix-t5: Mixing pre-trained transformers with graph-aware layers for text-to-sql parsing. In Proceedings of the AAAI Conference on Artificial Intell...

    2023

  14. [22]

    Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023 c . https://doi.org/10.18653/v1/2023.acl-long.256 Unified demonstration retriever for in-context learning . In Proceedings of the 61st Annual Meeting of the Associati...

    2023 doi

  15. [23]

    Zhishuai Li, Xiang Wang, Jingjing Zhao, Sun Yang, Guoqing Du, Xiaoru Hu, Bin Zhang, Yuxiao Ye, Ziyue Li, Rui Zhao, and Hangyu Mao. 2024. https://arxiv.org/abs/2403.09732 Pet-sql: A prompt-enhanced two-round refinement of text-to-sql with cross-consistency . Preprint, arXiv:2403.09732

    2024 arXiv

  16. [24]

    Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2020. https://arxiv.org/abs/2012.12627 Bridging textual and tabular data for cross-domain text-to-sql semantic parsing . Preprint, arXiv:2012.12627

    2020 arXiv

  17. [25]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized bert pretraining approach . Preprint, arXiv:1907.11692

    2019 arXiv

  18. [26]

    Man Luo, Xin Xu, Yue Liu, Panupong Pasupat, and Mehran Kazemi. 2024. In-context learning with retrieved demonstrations for language models: A survey. arXiv preprint arXiv:2401.11624

    2024 arXiv

  19. [27]

    Linyong Nan, Yilun Zhao, Weijin Zou, Narutatsu Ri, Jaesung Tae, Ellen Zhang, Arman Cohan, and Dragomir Radev. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.996 Enhancing text-to- SQL capabilities of large language models: A study on prompt design strategies . In Findin...

    2023 doi

  20. [28]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff ...

    2024 arXiv

  21. [29]

    Mohammadreza Pourreza and Davood Rafiei. 2023. https://proceedings.neurips.cc/paper_files/paper/2023/file/72223cc66f63ca1aa59edaec1b3670e6-Paper-Conference.pdf Din-sql: Decomposed in-context learning of text-to-sql with self-correction . In Advances in Neural Information Proce...

    2023

  22. [30]

    Jiexing Qi, Jingyao Tang, Ziwei He, Xiangpeng Wan, Yu Cheng, Chenghu Zhou, Xinbing Wang, Quanshi Zhang, and Zhouhan Lin. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.211 RASAT : Integrating relational structures into pretrained S eq2 S eq model for text-to- SQL . In Proce...

    2022 doi

  23. [31]

    Nitarshan Rajkumar, Raymond Li, and Dzmitry Bahdanau. 2022. https://arxiv.org/abs/2204.00498 Evaluating the text-to-sql capabilities of large language models . Preprint, arXiv:2204.00498

    2022 arXiv

  24. [32]

    Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2022. https://arxiv.org/abs/2112.08633 Learning to retrieve prompts for in-context learning . Preprint, arXiv:2112.08633

    2022 arXiv

  25. [33]

    Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. 2018. https://doi.org/10.18653/v1/N18-2074 Self-attention with relative position representations . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Lang...

    2018 doi

  26. [34]

    Zhili Shen, Pavlos Vougiouklis, Chenxin Diao, Kaustubh Vyas, Yuanyi Ji, and Jeff Z. Pan. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.449 Improving retrieval-augmented text-to- SQL with AST -based ranking and schema pruning . In Proceedings of the 2024 Conference on Empir...

    2024 doi

  27. [35]

    Kummerfeld, and Tianyi Zhang

    Yuan Tian, Zheng Zhang, Zheng Ning, Toby Jia-Jun Li, Jonathan K. Kummerfeld, and Tianyi Zhang. 2024. https://arxiv.org/abs/2305.07372 Interactive text-to-sql generation via editable step-by-step explanations . Preprint, arXiv:2305.07372

    2024 arXiv

  28. [36]

    Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. https://doi.org/10.18653/v1/2020.acl-main.677 RAT-SQL : Relation-aware schema encoding and linking for text-to- SQL parsers . In Proceedings of the 58th Annual Meeting of the Association ...

    2020 doi

  29. [37]

    Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. https://arxiv.org/abs/2206.07682 Emergen...

    2022 arXiv

  30. [38]

    Tao Yu, Michihiro Yasunaga, Kai Yang, Rui Zhang, Dongxu Wang, Zifan Li, and Dragomir Radev. 2018 a . https://doi.org/10.18653/v1/D18-1193 S yntax SQLN et: Syntax tree networks for complex and cross-domain text-to- SQL task . In Proceedings of the 2018 Conference on Empirical M...

    2018 doi

  31. [39]

    Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and Dragomir Radev. 2018 b . https://doi.org/10.18653/v1/D18-1425 S pider: A large-scale human-labeled dataset for complex and cross-domain se...

    2018 doi

  32. [40]

    Hanchong Zhang, Ruisheng Cao, Lu Chen, Hongshen Xu, and Kai Yu. 2023 a . https://doi.org/10.18653/v1/2023.findings-emnlp.227 ACT - SQL : In-context learning for text-to- SQL with automatically-generated chain-of-thought . In Findings of the Association for Computational Lingui...

    2023 doi

  33. [41]

    Kun Zhang, Xiexiong Lin, Yuanzhuo Wang, Xin Zhang, Fei Sun, Cen Jianhe, Hexiang Tan, Xuhui Jiang, and Huawei Shen. 2023 b . https://doi.org/10.18653/v1/2023.findings-emnlp.48 R e FSQL : A retrieval-augmentation framework for text-to- SQL generation . In Findings of the Associa...

    2023 doi

  34. [42]

    Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. 2022. https://arxiv.org/abs/2210.03493 Automatic chain of thought prompting in large language models . Preprint, arXiv:2210.03493

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.