Pith. sign in

REVIEW 4 major objections 7 minor 2 cited by

Interactive Text-to-SQL via Expected Information Gain for Disambiguation

T0 review · 4 major / 7 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 Interactive Text-to-SQL can pick the best clarification question in a single pass by scoring only the branches along the most probable candidate's path.

desk verdict Good framing, but the claimed O(N) globally optimal clarification theorem does not hold and the implementation actually uses the full EIG computation. read the letter →

arxiv 2507.06467 v1 pith:WDVBHDPP submitted 2025-07-09 cs.DB

Luyu Qiu , Jianing Li , Chi Su , Lei Chen This is my paper
classification cs.DB
keywords Text-to-SQLexpectedinformationgaininteractivedisambiguationsemanticambiguityclarificationquestionsbranchingtreeuncertainty-awareparsinglargelanguagemodels
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

This paper tries to establish that a Text-to-SQL system can resolve ambiguous natural-language queries by keeping a probability distribution over candidate SQL queries and, at each step, asking the user the single most informative clarification. The central claim is that the expected information gain (EIG) of every candidate clarification can be computed in one pass over the candidates, using only the branches that lead to the current most probable query, reducing the cost from $O(NM)$ to $O(N)$. If correct, this makes principled, uncertainty-driven disambiguation fast enough for interactive use while improving accuracy on ambiguous benchmarks. The paper reports that on an ambiguous subset of Spider and on AmbiQT, EIG-based question selection outperforms random, frequency-based, and raw-information-gain baselines, and that the same interaction module improves several existing Text-to-SQL systems on Spider 2.0.

What carries the argument

The branching tree (Definition 3.1), where internal nodes are decision variables such as selected columns, comparison operators in WHERE conditions, or join paths; edges are alternative interpretations; and leaves are full SQL candidates. The argument rides on the identity $I(X_i;Y) = P(Q^* \to X_i)\,H(X_i)$, which turns a sum over all candidates and all decision variables into a product on a single root-to-leaf path, and on the assumption that decision variables can be recovered from token-level differences across the generated candidates.

What would settle it

Build a small synthetic candidate distribution where a decision variable not on the path to the most probable candidate has larger exact mutual information $I(X_i;Y)$ than every variable on that path; if such a distribution exists, Theorem 3.2's one-pass formula is not an exact equivalence and the global-optimality claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 3.2: in a branching tree whose leaves are candidate SQL queries with probabilities, the optimal decision variable to clarify is $X^* = \arg\max_{X_i \in P(Q^*)} P(Q^* \to X_i) \cdot H(X_i)$, where $Q^*$ is the highest-probability candidate, $P(Q^* \to X_i)$ is the cumulative probability of candidates passing through node $X_i$, and $H(X_i)$ is the entropy over that node's interpretations. The paper argues that this one-path expression equals the full mutual information $I(X_i;Y)$ for variables on the path to $Q^*$, so a single scan of the candidates suffices for globally optimal disambiguation. The implemented loop is: generate candidate SQLs and probabilities with an LLM, identify decision variables by token-level comparison, ask the user the highest-EIG question, filter and renormalize the distribution over consistent candidates, and repeat until one candidate crosses a confidence threshold.

Load-bearing premise

The intended SQL query must be among the candidate set generated at the start; if it is not, no clarification question can recover it because filtering only renormalizes probabilities over the existing candidates.

Editorial extensions

If this is right

  • EIG-based clarification improves execution accuracy and exact-match accuracy over random, frequency-based, and raw-information-gain interaction strategies on an ambiguous subset of Spider.
  • The $O(N)$ scoring rule makes uncertainty-driven clarification cheap enough for interactive, real-time database question answering.
  • The interaction module can be layered onto existing Text-to-SQL systems without retraining, and it improves their accuracy on Spider 2.0 across all tested baselines.
  • Multi-turn interaction that remembers earlier clarifications beats single-turn clarification, especially for dependent or layered ambiguities.

Reading between the lines

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

  • Not pursued in the paper: the same one-pass scoring could be transferred to any conditional generation task that carries a distribution over candidates, such as text-to-code or text-to-chart, where ambiguity also lives in structural choices.
  • The paper's framework leaves open what to do when the candidate set is incomplete; a natural extension is a threshold check that triggers candidate regeneration rather than continuing to renormalize over the same set.
  • The large exact-match gains on AmbiQT suggest the EIG criterion resolves structural choices, not just value choices; a testable extension is to compare EIG-selected questions against the gold clarification annotations to measure alignment.
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 / 7 minor

Summary. This paper proposes an interactive Text-to-SQL framework that maintains a probability distribution over candidate SQL queries and, at each turn, selects a clarification question by maximizing the expected information gain (EIG), defined as the mutual information I(X_i; Y) between a decision variable X_i and the candidate-query distribution Y (Definition 2.4, Eq. (1)). The paper's central theoretical claim is Theorem 3.2 (Section 3.3), which asserts that the optimal decision variable is found by maximizing P(Q*→X_i)·H(X_i) along the path to the most probable candidate, reducing the cost from O(NM) to O(N). The implementation in Algorithm 1, however, computes the full EIG for all M decision variables (lines 7–11). Experiments are reported on an ambiguous subset of Spider (Table 1), a single- vs multi-turn comparison (Table 2), and AmbiQT and Spider 2.0 comparisons (Tables 3–4), showing consistent improvements of the EIG strategy over the described selection heuristics and baseline models.

Significance. Timeliness is real: semantic ambiguity is a known obstacle for deployed Text-to-SQL systems, and using mutual information to select clarification questions is a sensible design that the paper correctly implements in Algorithm 1 (Eq. (11)). The ablation against Random, Min/Max Probability, and Info Gain heuristics is a reasonable experimental design; the reported gains (38.3% vs 33.8% exact match over Info Gain in Table 1; improvements across all AmbiQT categories in Table 3) are sizable; and the framework is model-agnostic, requiring no retraining of the base generator. These strengths do not offset the problems with the headline claims: Theorem 3.2 is false, the implemented algorithm does not use the claimed O(N) criterion, and neither 'globally optimal' nor 'fewer interactions' is supported by any measurement. Once the unsupported claims are removed, the residual contribution is an incremental application of standard EIG, with the empirical reliability additionally weakened by unexplained baseline numbers in Table 3 and unexamined candidate-set coverage.

major comments (4)
  1. [Section 3.3, Theorem 3.2 and Eq. (9)] Theorem 3.2 is not a valid consequence of Definition 2.4. The paper defines EIG as I(X_i; Y) = H(Y) − H(Y|X_i) (Eq. (8)), but the proof replaces H(Y|X_i) with H(X_i), the entropy of the decision variable's values, and then multiplies by the cumulative branch mass P(Q*→X_i) to obtain Eq. (9). These are different quantities: since X_i is determined by the candidate tokens, the exact mutual information is I(X_i; Y) = H(X_i) (because H(X_i|Y) = 0), with no multiplicative factor, and P(Q*→X_i)·H(X_i) coincides with H(X_i) only when every candidate passes through X_i. The proof's statement that 'queries unaffected by the clarification do not alter the conditional distribution' is incorrect: unaffected candidates remain in H(Y|X_i = x) for every value x and suppress the information gain. Concretely, take P(Q1)=0.46, P(Q2)=0.44, P(Q3)=0.10 with Q*=Q1 and X_i the node whose two values separate Q1 and Q2 while Q3 does not pass through X_i. Under the paper's own convention of Eq. (5), in which branch masses are the candidate masses and the sum in Eq. (3) covers all candidates, Eq. (8) gives I(X_i;Y) = 1.3686 − (0.46·0.677 + 0.44·0.691) ≈ 0.75 bits, whereas Eq. (9) gives 0.90·H(0.511, 0.489) ≈ 0.90 bits (renormalizing the branch masses to 0.90 gives ≈0.69 bits). The two values differ, so Eq. (9) is not the EIG and the criterion can select a different clarification than the one maximizing Eq. (8). The restriction of the search to the path of Q* is likewise unjustified. The 'globally optimal disambiguation' claims in the Abstract and Sections 1.5 and 5.2 therefore rest on an untrue identity.
  2. [Algorithm 1 vs. Section 3.3; Section 4.3.1] Section 3.3 claims that the simplified criterion 'significantly simplifies and accelerates' computation and states that this 'optimized decision selection criterion' is empirically validated, but neither the implementation nor the experiments use Eq. (9). Algorithm 1 (lines 7–11) computes, for every X_i in the full set of M decision variables, the conditional entropy H(Y|X_i) and the complete EIG I(X_i; Y) = H(Y) − H(Y|X_i), which is exactly the O(NM) procedure that Section 3.3 proposes to avoid; the complexity of the implemented per-turn selection is O(NM), not O(N). The ablation in Section 4.3.1 likewise defines EIG as the full computation 'under the current distribution over SQL candidates.' No experiment measures runtime or the number of interaction steps, although the abstract claims that the method requires 'fewer interactions than baselines' and Section 3.3 advertises real-time responsiveness; Table 2 compares single- and multi-turn accuracy but reports no turn counts. Consequently, neither the O(N) efficiency claim nor the 'fewer interactions' claim is supported by any measurement, and the paper's own pseudocode contradicts its stated theoretical contribution.
  3. [Sections 3.4.1–3.4.2, 3.4.4] Sections 3.4.1–3.4.2 assume that the LLM-generated candidate set Y contains the user's intended SQL query and that token-level differences among candidates capture every relevant semantic ambiguity. If the gold interpretation is not among the candidates, no clarification question can recover it, and the post-clarification renormalization of Section 3.4.4 merely redistributes probability over an incomplete set. The paper reports no coverage statistics (for example, how often the gold query or an execution-equivalent query appears in the top-n candidate set) on Spider, AmbiQT, or Spider 2.0. This matters for reading Tables 1–3: a substantial part of the reported accuracy is likely determined by candidate-generation quality rather than by the EIG selection rule, and it is also relevant to the 'globally optimal disambiguation' claim, which presupposes that the candidate set is a faithful representation of the user's possible intents.
  4. [Section 4.4.1, Table 3] Section 4.4.1 and Table 3 do not explain how the baseline top-1 numbers were obtained, although the text notes that LogicalBeam's original evaluation used top-5 accuracy. In Table 3, the LogicalBeam exact-match and execution numbers coincide with those of T5-3B to within 0.07 percentage points on all four ambiguity categories (exact match 48.06/48.06, 42.06/41.99, 78.82/78.82, 68.32/68.32 for Column, Table, Join, and Precomputed, respectively), which is not plausible as an independent rerun of a substantially different method. The authors should either describe the protocol used to produce these baseline numbers (including the candidate-generation budget and decoding settings) or correct the table; as presented, the claimed improvements over LogicalBeam in Table 3 cannot be evaluated.
minor comments (7)
  1. [Section 4.3.1, Table 1] The text states that Random achieves 6.6% exact match on Hard questions, but Table 1 reports 10.6% for Random Select in that cell; the 'All' column for Max Probability (30.6%) is also not reproducible as a weighted average of the row's per-category values (approximately 32.5% with the numbers shown), so the text and table should be reconciled.
  2. [Sections 3.2, 3.4.1, 4.3.1] No sensitivity analysis is reported for the framework's free parameters: the number of generated candidates n, the stopping threshold τ, and the 0.7 confidence cutoff used to build the ambiguous Spider subset, which directly determines the composition of Table 1.
  3. [Abstract, Section 4.3.2] The abstract claims the method requires 'fewer interactions than baselines,' but no experiment reports interaction counts or user-effort measurements; Table 2 reports accuracy only.
  4. [Section 4.3.2] The single-turn variant is underspecified: because the post-clarification update of Section 3.4.4 filters candidates on the basis of the user's answer, the updated distribution already encodes prior answers, so it is unclear what 'does not retain the history' means and why the two variants should differ in selection behavior.
  5. [Formatting] The manuscript retains template placeholders ('PVLDB, 14(1), 2020' and 'URL_TO_YOUR_ARTIFACTS'), a typo in Algorithm 1's title ('Expeceted'), a malformed summation symbol ('Í') in the equation in Section 3.3, and a garbled caption for Figure 4; these should be cleaned up before any resubmission.
  6. [Section 3.3, Definition 3.1] The notation is overloaded: H(X_i) in Eq. (9) denotes the entropy of the decision variable while H(Y) denotes the entropy of the candidate distribution, and P(Q*→X_i) is introduced only inside the proof of Theorem 3.2; a numbered definition with distinct symbols would prevent confusion.
  7. [References, Section 1.3] Some citations appear mismatched with their context (for example, [16], a thesis on sequence labelling, is cited for IRNet), and the Section 1.3 claim that SQL is underrepresented in model training corpora is made without citation or measurement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: EIG selection is defined independently of benchmark labels; Theorem 3.2's flaw is a soundness issue, not a circular reduction.

full rationale

The derivation chain is self-contained and does not reduce to its own inputs. The EIG criterion in Definition 2.4 and Eq. (8) is defined from first principles of information theory, and the probabilities P(Q_i) come from the LLM's generation, not from fitting to the benchmark labels or to the gold disambiguation annotations. The empirical protocol simulates user responses using gold annotations only after the clarification question has been selected by the EIG rule, so the gold answer is not used to choose the question. There are no load-bearing self-citations: the cited baselines, datasets, and prior systems are external, and no uniqueness theorem or prior result by the same authors is invoked to force the framework. The main mathematical weakness is Theorem 3.2, whose claimed identity Eq. (9) = Eq. (8) is not a valid derivation and is internally inconsistent with Algorithm 1, which computes the full EIG via Eq. (11). However, an invalid or unsupported theorem is a correctness or soundness problem, not a circularity, because Eq. (9) is not equivalent to Eq. (8) by construction and the conclusion is not assumed in the definition of the criterion. Therefore no circular step is exhibited, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The framework depends on LLM candidate generation, probability calibration, token-level branch extraction, and the correctness of Theorem 3.2. No new physical entities are introduced; the free parameters are the number of candidates, the stopping threshold, and the ambiguity cutoff used to build the Spider subset.

free parameters (3)
  • Number of SQL candidates n = not stated (examples use 4 to 6)
    Chosen by hand; the entire filtering pipeline depends on it, and no sensitivity analysis is given.
  • Stopping confidence threshold tau = not stated
    Predefined stopping criterion in Section 3.2; its value is never reported or swept.
  • Ambiguity cutoff 0.7 = 0.7
    Used in Section 4.3.1 to select the ambiguous Spider subset; chosen by hand, no justification.
assumptions (5)
  • domain assumption The generated candidate set Y always contains the user's intended SQL query.
    Section 3.4.1: candidates come from an LLM; if the correct SQL is missing, filtering and renormalizing cannot recover it. No coverage statistics are reported.
  • domain assumption Normalized LLM generation probabilities reflect the true posterior P(Q | q, S).
    Definition 2.3 and Section 3.4.1 treat model scores as probabilities over interpretations; LLM probabilities are not calibrated confidence.
  • domain assumption Decision variables extracted by token-level comparison cover all semantic ambiguities.
    Section 3.4.2: variables are only created where token strings differ across candidates; semantic ambiguities that coincide in tokenization remain invisible.
  • ad hoc to paper Theorem 3.2's step that unaffected candidates do not alter conditional entropy.
    Section 3.3 proof: the drop from H(Y) to P(Q* -> X_i) * H(X_i) discards entropy mass from candidates not passing through X_i; this step is mathematically invalid.
  • domain assumption Users always give the answer consistent with the gold annotation.
    Section 4.1: 'We simulate user responses using the gold disambiguation annotations'; this removes user error and makes interaction look more effective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Interactive Text-to-SQL via Expected Information Gain for Disambiguation." pith.science (2026). https://pith.science/paper/WDVBHDPP

@misc{pith2026250706467,
  author       = {Pith},
  title        = {Pith review of: Interactive Text-to-SQL via Expected Information Gain for Disambiguation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WDVBHDPP}},
  note         = {Machine review of arXiv:2507.06467}
}
read the original abstract

Relational databases are foundational to numerous domains, including business intelligence, scientific research, and enterprise systems. However, accessing and analyzing structured data often requires proficiency in SQL, which is a skill that many end users lack. With the development of Natural Language Processing (NLP) technology, the Text-to-SQL systems attempt to bridge this gap by translating natural language questions into executable SQL queries via an automated algorithm. Yet, when operating on complex real-world databases, the Text-to-SQL systems often suffer from ambiguity due to natural ambiguity in natural language queries. These ambiguities pose a significant challenge for existing Text-to-SQL translation systems, which tend to commit early to a potentially incorrect interpretation. To address this, we propose an interactive Text-to-SQL framework that models SQL generation as a probabilistic reasoning process over multiple candidate queries. Rather than producing a single deterministic output, our system maintains a distribution over possible SQL outputs and seeks to resolve uncertainty through user interaction. At each interaction step, the system selects a branching decision and formulates a clarification question aimed at disambiguating that aspect of the query. Crucially, we adopt a principled decision criterion based on Expected Information Gain to identify the clarification that will, in expectation, most reduce the uncertainty in the SQL distribution.

Figures

Figures reproduced from arXiv: 2507.06467 by the authors.

Figure 2
Figure 2. Illustration of candidate branches. solely on textual cues are inherently ill-equipped to disambiguate such uncertainties, frequently producing results misaligned with users’ true analytical intentions. 1.4 Semantic Ambiguity and User Intent The ambiguity discussed here is particularly prevalent and prob￾lematic in real-world analytical contexts. Unlike schema-related ambiguity arising from database structures, we e… view at source ↗
Figure 3
Figure 3. The proposed interactive Text-to-SQL system. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Illustration of the (a) input user query and several candidate SQLs with possibility; and (b) the decision branches with [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Illustration of the graph representation of branch decisions process. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. ABISS: Evaluating Text-to-SQL Systems Through Agent Interaction

    cs.DB 2026-07 conditional novelty 6.0 of 10

    A new benchmark shows that text-to-SQL models detect problematic questions but fail to pinpoint the exact problem type and to resolve the question after a useful clarification.

  2. EGREFINE: An Execution-Grounded Optimization Framework for Text-to-SQL Schema Refinement

    cs.DB 2026-05 unverdicted novelty 6.0 of 10

    EGRefine optimizes column renamings via execution-grounded verification and view materialization to recover Text-to-SQL accuracy lost to schema naming issues while guaranteeing query equivalence.

Reference graph

Works this paper leans on

39 extracted references · 16 canonical work pages · cited by 2 Pith papers

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    arXiv 2023

  2. [2]

    Adithya Bhaskar, Tushar Tomar, Ashutosh Sathe, and Sunita Sarawagi. 2023. Benchmarking and improving text-to-sql generation under ambiguity. arXiv preprint arXiv:2310.13659 (2023)

    arXiv 2023

  3. [3]

    Ruisheng Cao, Lu Chen, Zhi Chen, Yanbin Zhao, Su Zhu, and Kai Yu. 2021. LGESQL: line graph enhanced text-to-SQL model with mixed local and non-local relations. arXiv preprint arXiv:2106.01093 (2021)

    arXiv 2021

  4. [4]

    Wenhu Chen. 2022. Large language models are few (1)-shot table reasoners. arXiv preprint arXiv:2210.06710 (2022)

    arXiv 2022

  5. [5]

    DongHyun Choi, Myeong Cheol Shin, EungGyun Kim, and Dong Ryeol Shin

    work page

  6. [6]

    Mingwen Dong, Nischal Ashok Kumar, Yiqun Hu, Anuj Chauhan, Chung-Wei Hang, Shuaichen Chang, Lin Pan, Wuwei Lan, Henghui Zhu, Jiarong Jiang, et al

  7. [7]

    Matthieu Futeral, Cordelia Schmid, Ivan Laptev, Benoît Sagot, and Rachel Bawden

    work page

  8. [8]

    Yujian Gan, Xinyun Chen, Jinxia Xie, Matthew Purver, John R Woodward, John Drake, and Qiaofu Zhang. 2021. Natural SQL: Making SQL easier to infer from natural language specifications. arXiv preprint arXiv:2109.05153 (2021)

    work page Pith review arXiv 2021

Show all 39 references
  1. [9]

    Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. 2023. Text-to-sql empowered by large language models: A benchmark evaluation. arXiv preprint arXiv:2308.15363 (2023)

    2023 arXiv

  2. [10]

    Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, Jian-Guang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-sql in cross-domain database with intermediate representation. arXiv preprint arXiv:1905.08205 (2019)

    2019 arXiv

  3. [11]

    Zhixin Guo, Minyxuan Yan, Jiexing Qi, Jianping Zhou, Ziwei He, Zhouhan Lin, Guanjie Zheng, and Xinbing Wang. 2023. Few-shot table-to-text generation with prompt planning and knowledge memorization. arXiv preprint arXiv:2302.04415 (2023)

    2023 arXiv

  4. [12]

    Izzeddin Gür, Semih Yavuz, Yu Su, and Xifeng Yan. 2018. Dialsql: Dialogue based structured query generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . 1339–1349

    2018

  5. [13]

    Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno, and Julian Martin Eisenschlos. 2020. TaPas: Weakly supervised table parsing via pre-training. arXiv preprint arXiv:2004.02349 (2020)

    2020 arXiv

  6. [14]

    Wonseok Hwang, Jinyeong Yim, Seunghyun Park, and Minjoon Seo. 2019. A comprehensive exploration on wikisql with table-aware word contextualization. arXiv preprint arXiv:1902.01069 (2019)

    2019 arXiv

  7. [15]

    Srinivasan Iyer, Ioannis Konstas, Alvin Cheung, Jayant Krishnamurthy, and Luke Zettlemoyer. 2017. Learning a neural semantic parser from user feedback. arXiv preprint arXiv:1704.08760 (2017)

    2017 arXiv

  8. [16]

    Kazuya Kawakami. 2008. Supervised sequence labelling with recurrent neural networks. Ph.D. Dissertation. Ph. D. thesis. 12

    2008

  9. [17]

    Fei Li and Hosagrahar V Jagadish. 2014. Constructing an interactive natural language interface for relational databases. Proceedings of the VLDB Endowment 8, 1 (2014), 73–84

    2014

  10. [18]

    What Do You Mean by That?

    Yuntao Li, Bei Chen, Qian Liu, Yan Gao, Jian-Guang Lou, Yan Zhang, and Dong- mei Zhang. 2020. " What Do You Mean by That?" A Parser-Independent Inter- active Approach for Enhancing Text-to-SQL. arXiv preprint arXiv:2011.04151 (2020)

    2020 arXiv

  11. [19]

    Yihang Li, Shuichiro Shimizu, Weiqi Gu, Chenhui Chu, and Sadao Kurohashi

  12. [20]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  13. [21]

    Aiwei Liu, Xuming Hu, Lijie Wen, and Philip S Yu. 2023. A comprehen- sive evaluation of ChatGPT’s zero-shot Text-to-SQL capability. arXiv preprint arXiv:2303.13547 (2023)

    2023 arXiv

  14. [22]

    Arpit Narechania, Adam Fourney, Bongshin Lee, and Gonzalo Ramos. 2021. DIY: Assessing the correctness of natural language to sql systems. In Proceedings of the 26th International Conference on Intelligent User Interfaces . 597–607

    2021

  15. [23]

    arXiv preprint arXiv:2201.08054 (2022)

    Visa: An ambiguous subtitles dataset for visual scene-aware machine translation. arXiv preprint arXiv:2201.08054 (2022)

    2022 arXiv

  16. [24]

    Bowen Qin, Binyuan Hui, Lihan Wang, Min Yang, Jinyang Li, Binhua Li, Ruiying Geng, Rongyu Cao, Jian Sun, Luo Si, et al. 2022. A survey on text-to-sql parsing: Concepts, methods, and future directions. arXiv preprint arXiv:2208.13629 (2022)

    2022 arXiv

  17. [25]

    Nitarshan Rajkumar, Raymond Li, and Dzmitry Bahdanau. 2022. Evaluating the text-to-sql capabilities of large language models. arXiv preprint arXiv:2204.00498 (2022)

    2022 arXiv

  18. [26]

    Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems 27 (2014)

    2014

  19. [27]

    Jonathan Pilault, Xavier Garcia, Arthur Bražinskas, and Orhan Firat. 2023. Interactive-chain-prompting: Ambiguity resolution for crosslingual conditional generation with interaction. arXiv preprint arXiv:2301.10309 (2023)

    2023 arXiv

  20. [28]

    Bing Wang, Yan Gao, Zhoujun Li, and Jian-Guang Lou. 2023. Know what I don’t know: Handling ambiguous and unknown questions for text-to-SQL. In Findings of the Association for Computational Linguistics: ACL 2023 . 5701–5714

    2023

  21. [29]

    Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2019. Rat-sql: Relation-aware schema encoding and linking for text-to-sql parsers. arXiv preprint arXiv:1911.04942 (2019)

    2019 arXiv

  22. [30]

    Xiaojun Xu, Chang Liu, and Dawn Song. 2017. Sqlnet: Generating structured queries from natural language without reinforcement learning. arXiv preprint arXiv:1711.04436 (2017)

    2017 arXiv

  23. [31]

    Bing Wang, Yan Gao, Zhoujun Li, and Jian-Guang Lou. 2022. Know What I don’t Know: Handling Ambiguous and Unanswerable Questions for Text-to-SQL. arXiv preprint arXiv:2212.08902 (2022)

    2022 arXiv

  24. [32]

    Pengcheng Yin, Graham Neubig, Wen-tau Yih, and Sebastian Riedel. 2020. TaBERT: Pretraining for joint understanding of textual and tabular data. arXiv preprint arXiv:2005.08314 (2020)

    2020 arXiv

  25. [33]

    Tao Yu, Chien-Sheng Wu, Xi Victoria Lin, Bailin Wang, Yi Chern Tan, Xinyi Yang, Dragomir Radev, Richard Socher, and Caiming Xiong. 2020. Grappa: Grammar-augmented pre-training for table semantic parsing. arXiv preprint arXiv:2009.13845 (2020)

    2020 arXiv

  26. [34]

    Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, et al. 2018. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task. arXiv preprint arXiv:1809.08887 (2018)

    2018 arXiv

  27. [35]

    Ziyu Yao, Yu Su, Huan Sun, and Wen-tau Yih. 2019. Model-based interactive semantic parsing: A unified framework and a text-to-SQL case study. arXiv preprint arXiv:1910.05389 (2019)

    2019 arXiv

  28. [39]

    Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2sql: Generating structured queries from natural language using reinforcement learning. arXiv preprint arXiv:1709.00103 (2017). 13

    2017 arXiv

  29. [2021]

    Computational Linguistics 47, 2 (2021), 309–332

    Ryansql: Recursively applying sketch-based slot fillings for complex text-to- sql in cross-domain databases. Computational Linguistics 47, 2 (2021), 309–332

    2021

  30. [2022]

    arXiv preprint arXiv:2212.10140 (2022)

    Tackling ambiguity with images: Improved multimodal machine translation and contrastive evaluation. arXiv preprint arXiv:2212.10140 (2022)

    2022 arXiv

  31. [2024]

    arXiv preprint arXiv:2410.11076 (2024)

    PRACTIQ: A Practical Conversational Text-to-SQL dataset with Ambiguous and Unanswerable Queries. arXiv preprint arXiv:2410.11076 (2024)

    2024

Pith tools

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