Tabular Data Understanding with LLMs: A Survey of Recent Advances and Challenges

Alan Ritter; Wei Xu; Xiaofeng Wu

REVIEW 4 major objections 7 minor 2 cited by

Tabular Data Understanding with LLMs: A Survey of Recent Advances and Challenges

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 This survey argues that today's table-understanding benchmarks reward retrieval and arithmetic, under-measuring the reasoning and structural complexity real tables demand.

desk verdict A useful survey compilation whose three headline gap findings are asserted more than demonstrated, with an anonymous citation undermining one of them. read the letter →

arxiv 2508.00217 v1 pith:4G7VHNFS submitted 2025-07-31 cs.CL cs.DBcs.LG

Xiaofeng Wu , Alan Ritter , Wei Xu This is my paper

classification cs.CLcs.DBcs.LG

keywords largelanguagemodelstableunderstandingquestionansweringtext-to-SQLbenchmarksurveytabularinputrepresentationsreasoningmultimodal

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 surveys how large language models and multimodal models read tables, organizing input formats into serialized text, database schemas, images, and learned table encoders. Its central claim is that the field's benchmarks are dominated by retrieval-style questions that can be answered with SQL or arithmetic, so high leaderboard scores do not yet show real table understanding. It supports this with evidence that models fall far below humans on hierarchical tables, long passages, and multi-table questions, and that performance shifts when the same table is written in different formats. If the claim is right, the next generation of benchmarks should test diagnosis, forecasting, chart choice, and intent inference rather than lookup alone.

What carries the argument

The machinery is a taxonomy of tabular input representations: serialization (flattening a table into text like Markdown or LaTeX), database schema (showing only the SQL structure), image rendering (preserving visual layout for multimodal models), and specialized table encoders (trainable column and row embeddings), together with a structured compilation of benchmarks for Table QA, Table-to-Text, Table Fact Verification, and Text-to-SQL. The taxonomy carries the generalization argument, because each benchmark selects a representation for convenience and models are sensitive to that choice, so the same underlying table can produce materially different scores. The benchmark tables carry the reasoning-gap argument by showing which tasks are solvable with SQL and which require advanced reasoning.

What would settle it

Take the questions from a representative sample of real-world table tasks and count how many can be answered correctly by a text-to-SQL system with no additional reasoning; if a large majority already exceeds human accuracy, the survey's claim that the field is stuck at retrieval would be falsified. A second check: re-run a current model on one table set rendered in Markdown, JSON, HTML, LaTeX, and image formats; if scores no longer vary by several percentage points, the claimed representation-generalization gap has closed.

Watch

Extended reading notes

Core claim

The paper's central claim is that table understanding with LLMs has three unresolved gaps: (1) most benchmarks reward fact retrieval solvable by text-to-SQL or arithmetic, leaving higher-order reasoning such as insight identification, forecasting, and ambiguous-query understanding largely untested; (2) models are brittle when tables are hierarchically structured, large, embedded in long text, or span multiple tables, with reported accuracy below fifty percent on benchmarks like MULTIHIERTT while humans score around eighty-three percent; and (3) models generalize poorly across tabular representations, so serialization choices such as Markdown, JSON, HTML, or images can shift performance by several percentage points. The survey concludes that existing benchmarks, often built by translating SQL queries or math expressions into natural language, underestimate what real applications require.

Load-bearing premise

The load-bearing premise is that the benchmarks collected in the survey's tables fairly represent the field; the paper itself admits in its Limitations section that the selection is a snapshot, so if the chosen benchmarks skew toward clean, SQL-derived tables, the conclusion that the field is retrieval-focused could be an artifact of that selection.

Editorial extensions

If this is right

If the gap analysis is right, current leaderboard results on WTQ, TabFact, and similar benchmarks should be read as retrieval and arithmetic ability, not as evidence of general table reasoning.
Benchmark designers should add tasks that require diagnostic, predictive, and prescriptive thinking, including ambiguous user requests where the model must infer intent or ask for clarification.
Models need explicit support for hierarchical tables, long passages, and multi-table scenarios, since human-level accuracy on those inputs remains far away.
Standardizing serialization formats and training models on representation-to-representation transformations should reduce the several-percentage-point swings caused by input format.
Table foundation models that integrate specialized table encoders into larger language models are the paper's promising direction for closing the robustness gap.

Reading between the lines

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

One implication the authors leave implicit is that benchmark scores across papers are not directly comparable when each benchmark uses a different serialization; the reported swings of several percentage points could be treated as a minimum measurement error in any meta-analysis of table tasks.
A testable extension follows from their taxonomy: build one table set and render each table in Markdown, JSON, HTML, LaTeX, and image form, then measure whether model rankings change; if they do, input representation is a confound in every comparison.
The human-versus-model gaps they cite suggest that the next bottleneck is not model scale but structure-aware parsing and multi-source reasoning, an inference the paper supports but does not state as a prediction.
Another unstated consequence is that text-to-table and leaderboard-construction tasks, which force models to reorganize information, may be better probes of table understanding than the retrieval-heavy QA benchmarks currently used.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper is a reasonably organized survey of tabular data understanding with LLMs. The taxonomy of input representations (serialization, schema, image, encoder) is clearly presented, and the benchmark tables with their "Directions" columns are a genuinely handy orientation aid, especially for newcomers. I also appreciate the coverage of less-standard tasks like leaderboard construction and query-focused summarization. If you want a compact map of the current benchmark landscape, this does the job.

The soft spots are concentrated in the three "critical gaps" that the paper bills as its main contribution. Finding 1 says most benchmarks rely on retrieval-oriented queries solvable by SQL, but this is never quantified. No selection criteria are given beyond "major NLP conferences," and Tables 1-4 look tilted toward well-structured, SQL-derived datasets. That could be a fair observation about the field, but here it is an impression backed by selected examples, not a demonstrated trend. The paper's own Limitations section admits the selection is a snapshot, which helps, but the representativeness question is load-bearing for all three findings. The stress-test note about Awesome-Table-QA and TabMWP-style reasoning tasks is relevant; the survey does include TabMWP and Text2Analysis, but the claim is about predominance, and the paper never shows the denominator.

There are also two concrete fixable issues. First, the MMQA results used to support the model-human gap on multi-table QA are from an anonymous under-review submission. That is unverifiable and should be replaced with a published source or removed; the survey should not lean on that for a central claim. Second, the Sui et al. sensitivity figures appear in two places with very different magnitudes (up to 5% vs up to 20%/50%). These may describe different manipulations, but as written they look inconsistent and need reconciliation.

Overall, the survey is a fair synthesis and its gap analysis is directionally plausible, but the evidence for the headline findings is thinner than the rhetoric. That is a revision problem, not a rejection problem. A serious referee can push for the missing quantification and source hygiene. I would send it to review, but I would not cite it in my own work until the anonymous reference and the internal inconsistency are cleaned up.

Referee Report

4 major / 7 minor

Summary. This survey reviews recent work on tabular data understanding with LLMs and MLLMs. It organizes the field through a taxonomy of tabular input representations (serialization, data schema, image, and table encoders) and surveys major table-related tasks (TQA, table-to-text, table fact verification, leaderboard construction, and text-to-SQL), with benchmark summaries in Tables 1-4. The paper's central contribution is a set of three findings: (1) current benchmarks are dominated by retrieval-focused tasks that require minimal reasoning beyond mathematical and logical operations; (2) models are not robust to complex table structures, large tables, long contexts, or multi-table scenarios; and (3) models generalize poorly across different tabular representations. It also proposes future directions, including higher-level reasoning benchmarks, serialization-to-serialization tasks, and scientific-document table understanding.

Significance. If the central findings hold, the survey provides a valuable organizing framework and a forward-looking agenda for table understanding research. The taxonomy of input representations (Figure 4) and the compiled benchmark tables are useful resources for researchers entering the field. The paper also credits recent concrete advances (e.g., TableGPT2, decomposition pipelines, Spider 2, Text2Analysis) and identifies relatively underexplored tasks, such as leaderboard construction and serialization-to-serialization. However, the strength of the findings depends on the representativeness of the benchmark selection and on the reliability of a few key cited results, particularly the anonymous MMQA submission. These dependencies are load-bearing for the paper's inductive claims and require attention.

major comments (4)

[§2.1, Tables 1-4] The claim that 'most benchmarks rely on detailed queries that prompt models to extract specific information' and the finding of a 'predominance of retrieval-focused tasks' is an inductive generalization about the field, but the survey does not define the population of benchmarks from which Tables 1-4 are drawn. The Limitations section states only that the selection draws from 'major NLP conferences, including ACL, EMNLP, NAACL, and ICLR, along with relevant studies from other domains and preprints,' with no inclusion/exclusion criteria, search protocol, or date range. Without a defined scope, the 'predominance' claim may reflect the authors' curation rather than the field's actual distribution. Please either specify the selection methodology and, if feasible, quantify the share of retrieval-focused versus reasoning-focused benchmarks within that defined corpus, or rephrase the finding as applying to the surveyed benchmarks rather than to the field as a whole. This is load-bearing for Finding 1 and, to a lesser extent, Finding 3.
[§2.2, Table 1 (MMQA)] Key evidence for the model-human gap on multi-table QA relies on MMQA, which is cited as 'Anonymous, 2024' and 'Under review' in the references. The reported numbers (o1-preview exact-match slightly above 50%, human performance approximately 89%) cannot be verified, and the benchmark is not available in a citable, public form. Because this is one of the primary supports for Finding 2, please either replace it with published, verifiable benchmarks (e.g., MultiTabQA or other peer-reviewed multi-table QA datasets) or explicitly mark the finding as preliminary and dependent on an unreviewed submission. Additionally, the name 'MMQA' collides with the published MMQA (Gupta et al., 2018) that is listed as the source for MMCoQA in Table 1; please disambiguate, e.g., 'MMQA (2024, anonymous submission)'.
[§2.3 vs. §3.1] Both sections attribute representation-sensitivity results to Sui et al. (2024), but the reported magnitudes appear inconsistent. Section 2.3 states that performance varies 'by up to 5% depending on how closely input formats align with the data encountered during pretraining,' while Section 3.1 reports that omitting marked partitions or altering input order causes performance drops of 'up to 20%' and removing example shots leads to deteriorations 'as much as 50%.' These could refer to different experimental conditions (format alignment versus input-design perturbations), but the paper does not say so. Please clarify the relationship between these numbers, define the quantity being measured in each case, and state this explicitly to avoid the appearance of an internal contradiction.
[§2.1, Figure 2] The distinction between 'Solvable by SQL' and 'Unsolvable by SQL' is central to Finding 1, but the figure's layout is difficult to parse. The left column is labeled 'A) Solvable by SQL,' yet the bottom portion of that column appears to include a Spider 2 ambiguous high-level question, which the text describes as going 'beyond the capabilities of SQL-based approaches.' Please redraw the figure so that each example is unambiguously assigned to one of the two categories, and ensure the caption matches the visual grouping.

minor comments (7)

[Table 1] The 'Directions' column categories (e.g., 'Input Complexity', 'Reasoning Difficulty') are not defined; consider adding a footnote with one-sentence definitions or examples. Also, the 'Output' column entries for WTQ and SQA appear truncated by the table formatting; please check the rendering.
[§3.1] The sentence 'Furthermore, they show a balanced combination of augmentation data (e.g., table sizes and keyword explanations) and sampled table text has proven effective' has a subject-verb agreement problem; the subject is 'a balanced combination,' so the verb should be 'has proven' or the sentence should be rephrased.
[§4.2] The claim that table-to-text 'has become less prominent, as the number of related publications has steadily decreased since 2021' is unsupported by a citation or a count; please add evidence or soften the statement to a qualitative observation.
[References / §2.2] The 'Anonymous, 2024' entry should be flagged in the text as an anonymous under-review submission at every point of citation, not only in the reference list, so readers can weigh the evidence accordingly.
[Limitations] The Limitations section is candid that the survey is a snapshot, but it does not mention the reliance on anonymous or unreviewed sources (MMQA) or the absence of a systematic search protocol; adding these caveats would improve transparency.
[Figure 5] The comparison of LaTeX, X-separated, and JSON serializations is useful, but the figure text is small and the source 'semi-structured table' is not clearly shown; consider enlarging the figure or adding annotations.
[Table 2] QTSUMM is listed under Table-to-Text/Summarization, but it is also discussed in §4.1 as a query-focused task bridging TQA and summarization. Please clarify the task categorization in the table or in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: this is a literature review whose findings are inductive generalizations from external, verifiable benchmarks, and the paper explicitly scopes its selection as a snapshot.

full rationale

The paper is a survey with no derived quantities, fitted parameters, or formal reductions, so there is no derivation chain that could collapse into its own inputs. Its three findings are inductive generalizations from external published benchmarks and prior work: the claim that most benchmarks are retrieval-focused is supported by the cited datasets in Tables 1-4 (WTQ, Spider, TabFact, HiTab, MULTIHIERTT, etc.) and by specific named prior results; the robustness claim relies on reported model and human scores from those benchmarks; and the representation-sensitivity claim relies on published studies such as Sui et al. (2024) and Gao et al. (2023). None of these numbers is produced by the survey itself, and no fitted parameter is renamed as a prediction. The selection-bias concern sometimes raised against such surveys is addressed in the Limitations section, which states that the selection 'primarily draws from major NLP conferences' and that the survey 'offers a snapshot of current progress rather than a definitive account'; this explicit scoping prevents the representative-ness assumption from being hidden. The author self-citations (Park et al., 2025; Bai et al., 2024) appear only as examples of scientific-document table extraction and are not load-bearing for any of the three findings; they are also peer-reviewed and externally checkable. The anonymous MMQA (2024) citation is unverifiable, but the survey uses it only as one piece of external evidence for the multi-table QA gap, not as a justification that reduces to the survey's own claims. No equation equates an output to an input, no uniqueness theorem from the authors' prior work is invoked to force a choice, and no known result is merely renamed. The paper therefore has no significant circularity.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

No free parameters or invented entities. The survey is a literature review; its conclusions rest on the unstated premise that the collected benchmarks represent the field, which the authors themselves qualify as a snapshot in the Limitations section.

assumptions (1)

domain assumption The benchmark tables (Tables 1-4) are a representative and accurate summary of table-understanding tasks in the field.
The three headline findings about limited reasoning, robustness, and generalization are inferred from this set of benchmarks; if the set is biased, the findings do not generalize. The authors acknowledge in the Limitations section that the selection is 'a snapshot'.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Tabular Data Understanding with LLMs: A Survey of Recent Advances and Challenges." pith.science (2026). https://pith.science/paper/4G7VHNFS

@misc{pith2026250800217,
  author       = {Pith},
  title        = {Pith review of: Tabular Data Understanding with LLMs: A Survey of Recent Advances and Challenges},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4G7VHNFS}},
  note         = {Machine review of arXiv:2508.00217}
}

read the original abstract

Tables have gained significant attention in large language models (LLMs) and multimodal large language models (MLLMs) due to their complex and flexible structure. Unlike linear text inputs, tables are two-dimensional, encompassing formats that range from well-structured database tables to complex, multi-layered spreadsheets, each with different purposes. This diversity in format and purpose has led to the development of specialized methods and tasks, instead of universal approaches, making navigation of table understanding tasks challenging. To address these challenges, this paper introduces key concepts through a taxonomy of tabular input representations and an introduction of table understanding tasks. We highlight several critical gaps in the field that indicate the need for further research: (1) the predominance of retrieval-focused tasks that require minimal reasoning beyond mathematical and logical operations; (2) significant challenges faced by models when processing complex table structures, large-scale tables, length context, or multi-table scenarios; and (3) the limited generalization of models across different tabular representations and formats.

Figures

Figures reproduced from arXiv: 2508.00217 by the authors.

**Figure 2.** The left side illustrates examples of tasks that can be addressed with SQL-based methods such as typical Text-to-SQL [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Illustration of the proposed task: Scientific Doc [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Taxonomy of table input representation methods, encompassing serialization, image, specialized table encoders, and Ours 60.1 [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Comparison of serialization methods for semi [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

**Figure 6.** Figure 6: Illustration of automatic leaderboard construction [PITH_FULL_IMAGE:figures/full_fig_p008_6.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. Full citation record

Towards Agents That Know When They Don't Know: Uncertainty as a Control Signal for Structured Reasoning
cs.AI 2025-09 conditional novelty 6.0 of 10

An LLM agent using retrieval and summary uncertainty as training rewards and inference filters produces more factual, useful multi-omics summaries and better downstream survival predictions.
LLM-Based Instance-Driven Heuristic Bias In the Context of a Biased Random Key Genetic Algorithm
cs.NE 2025-09 conditional novelty 4.0 of 10

An LLM-generated, per-instance bias vector improves a BRKGA on the NP-hard Longest Run Subsequence problem, with statistically significant gains on 15 of 35 instance groups, concentrated on complex instances.

Reference graph

Works this paper leans on

12 extracted references · 3 canonical work pages · cited by 2 Pith papers

[2]

Peter Baile Chen, Yi Zhang, and Dan Roth

Dr.spider: A diagnostic evaluation bench- mark towards text-to-sql robustness.Preprint, arXiv:2301.08881. Peter Baile Chen, Yi Zhang, and Dan Roth. 2024. Is table retrieval a solved problem? exploring join- aware multi-table retrieval. InProceedings of the 62nd Annual Meeting of the Association for Computa- tional Linguistics (ACL). ArXiv:2404.09889 [cs.I...

arXiv 2024
[5]

InProceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 4320–4333, Online

Tapas: Weakly supervised table parsing via pre-training. InProceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 4320–4333, Online. Association for Computa- tional Linguistics. Atin Sakkeer Hussain. 2025. Artemis-da: An advanced reasoning and transformation engine for multi- step insight synthesis in data analyti...

arXiv 2025
[7]

TableQAKit: A Comprehensive and Practical Toolkit for Table-based Question Answering

Tableqakit: A comprehensive and practical toolkit for table-based question answering.Preprint, arXiv:2310.15075. Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Rongyu Cao, Ruiying Geng, Nan Huo, Xuanhe Zhou, Chenhao Ma, Guoliang Li, Kevin C. C. Chang, Fei Huang, Reynold Cheng, and Yongbin Li. 2023a. Can llm alread...

work page Pith review arXiv
[10]

In Proceedings of the first Workshop on Information Extraction from Scientific Publications, pages 20–25, Online

TELIN: Table entity LINker for extracting leaderboards from machine learning publications. In Proceedings of the first Workshop on Information Extraction from Scientific Publications, pages 20–25, Online. Association for Computational Linguistics. Yunhu Ye, Binyuan Hui, Min Yang, Binhua Li, Fei Huang, and Yongbin Li. 2023. Large language mod- els are vers...

arXiv 2023
[11]

InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588–6600, Dublin, Ireland

MultiHiertt: Numerical reasoning over multi hierarchical tabular and textual data. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588–6600, Dublin, Ireland. Association for Computational Linguistics. Yilun Zhao, Zhenting Qi, Linyong Nan, Boyu Mi, Yixin Liu, Weijin Zou, Simeng Han, ...

arXiv doi:10.48550/arxiv.2406.08100 2023
[2017]

Preprint, arXiv:1709.00103

Seq2sql: Generating structured queries from natural language using reinforcement learning. Preprint, arXiv:1709.00103. Wei Zhou, Mohsen Mesgar, Heike Adel, and Annemarie Friedrich. 2024. FREB-TQA: A fine-grained robust- ness evaluation benchmark for table question answer- ing. InProceedings of the 2024 Conference of the North American Chapter of the Assoc...

arXiv 2017
[2019]

Step-by-step: Separating planning from re- alization in neural data-to-text generation.Preprint, arXiv:1904.03396. Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Kry´sci´nski, Nick Schoelkopf, Riley Kong, Xiangru Tang, Murori Mutuma, Ben Rosand, Isabel Trindade, Renusree Bandaru, Jacob Cunningham, Caiming Xiong, ...

work page Pith review arXiv 1904
[2020]

Computer Speech and Language, 59:123–156

Evaluating the state-of-the-art of end-to-end natural language generation: The e2e nlg challenge. Computer Speech and Language, 59:123–156. Xi Fang, Weijie Xu, Fiona Anting Tan, Jiani Zhang, Ziqing Hu, Yanjun Qi, Scott Nickleach, Diego Socol- insky, Srinivasan Sengamedu, and Christos Faloutsos

work page

Show all 12 references

[2021]

Towards table-to-text generation with numer- ical reasoning. InProceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1451–1465, Online....

2024 doi arXiv
[2022]

AIT-QA: Question answering dataset over complex tables in the airline industry. InProceed- ings of the 2022 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies: Industry Track, pages 305–314, Hybrid: Seattle,...

2022 arXiv
[2023]

Fan Bai, Junmo Kang, Gabriel Stanovsky, Dayne Fre- itag, Mark Dredze, and Alan Ritter

Transformers for tabular data representation: A survey of models and applications.Transactions of the Association for Computational Linguistics, 11:227–249. Fan Bai, Junmo Kang, Gabriel Stanovsky, Dayne Fre- itag, Mark Dredze, and Alan Ritter. 2024. Schema- driven information ...

2024 arXiv
[2024]

Preprint, arXiv:2402.17944

Large language models(llms) on tabular data: Prediction, generation, and understanding – a survey. Preprint, arXiv:2402.17944. Yujian Gan, Xinyun Chen, Qiuping Huang, Matthew Purver, John R. Woodward, Jinxia Xie, and Peng- sheng Huang. 2021a. Towards robustness of text-to- sql...

2023 arXiv

Pith tools

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

[1] [2]

Peter Baile Chen, Yi Zhang, and Dan Roth

Dr.spider: A diagnostic evaluation bench- mark towards text-to-sql robustness.Preprint, arXiv:2301.08881. Peter Baile Chen, Yi Zhang, and Dan Roth. 2024. Is table retrieval a solved problem? exploring join- aware multi-table retrieval. InProceedings of the 62nd Annual Meeting of the Association for Computa- tional Linguistics (ACL). ArXiv:2404.09889 [cs.I...

arXiv 2024

[2] [5]

InProceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 4320–4333, Online

Tapas: Weakly supervised table parsing via pre-training. InProceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 4320–4333, Online. Association for Computa- tional Linguistics. Atin Sakkeer Hussain. 2025. Artemis-da: An advanced reasoning and transformation engine for multi- step insight synthesis in data analyti...

arXiv 2025

[3] [7]

TableQAKit: A Comprehensive and Practical Toolkit for Table-based Question Answering

Tableqakit: A comprehensive and practical toolkit for table-based question answering.Preprint, arXiv:2310.15075. Jinyang Li, Binyuan Hui, Ge Qu, Jiaxi Yang, Binhua Li, Bowen Li, Bailin Wang, Bowen Qin, Rongyu Cao, Ruiying Geng, Nan Huo, Xuanhe Zhou, Chenhao Ma, Guoliang Li, Kevin C. C. Chang, Fei Huang, Reynold Cheng, and Yongbin Li. 2023a. Can llm alread...

work page Pith review arXiv

[4] [10]

In Proceedings of the first Workshop on Information Extraction from Scientific Publications, pages 20–25, Online

TELIN: Table entity LINker for extracting leaderboards from machine learning publications. In Proceedings of the first Workshop on Information Extraction from Scientific Publications, pages 20–25, Online. Association for Computational Linguistics. Yunhu Ye, Binyuan Hui, Min Yang, Binhua Li, Fei Huang, and Yongbin Li. 2023. Large language mod- els are vers...

arXiv 2023

[5] [11]

InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588–6600, Dublin, Ireland

MultiHiertt: Numerical reasoning over multi hierarchical tabular and textual data. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6588–6600, Dublin, Ireland. Association for Computational Linguistics. Yilun Zhao, Zhenting Qi, Linyong Nan, Boyu Mi, Yixin Liu, Weijin Zou, Simeng Han, ...

arXiv doi:10.48550/arxiv.2406.08100 2023

[6] [2017]

Preprint, arXiv:1709.00103

Seq2sql: Generating structured queries from natural language using reinforcement learning. Preprint, arXiv:1709.00103. Wei Zhou, Mohsen Mesgar, Heike Adel, and Annemarie Friedrich. 2024. FREB-TQA: A fine-grained robust- ness evaluation benchmark for table question answer- ing. InProceedings of the 2024 Conference of the North American Chapter of the Assoc...

arXiv 2017

[7] [2019]

Step-by-step: Separating planning from re- alization in neural data-to-text generation.Preprint, arXiv:1904.03396. Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Kry´sci´nski, Nick Schoelkopf, Riley Kong, Xiangru Tang, Murori Mutuma, Ben Rosand, Isabel Trindade, Renusree Bandaru, Jacob Cunningham, Caiming Xiong, ...

work page Pith review arXiv 1904

[8] [2020]

Computer Speech and Language, 59:123–156

Evaluating the state-of-the-art of end-to-end natural language generation: The e2e nlg challenge. Computer Speech and Language, 59:123–156. Xi Fang, Weijie Xu, Fiona Anting Tan, Jiani Zhang, Ziqing Hu, Yanjun Qi, Scott Nickleach, Diego Socol- insky, Srinivasan Sengamedu, and Christos Faloutsos

work page