Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Hybrid Graphs for Table-and-Text based Question Answering using LLMs

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper sets out to show that a question-conditioned Hybrid Graph, built from both a table and its linked passages, lets an LLM answer multi-hop Table-Text QA without any fine-tuning.

desk verdict A practical zero-shot recipe for table-text QA with real but modest gains over full-context baselines, undermined by overclaimed abstract numbers and an unanalyzed pruning failure mode. read the letter →

arxiv 2501.17767 v1 pith:B6KSXQLP submitted 2025-01-29 cs.CL cs.AI

Ankush Agarwal , Ganesh S , Chaitanya Devaguptapu This is my paper · ORCID
classification cs.CLcs.AI
keywords table-textquestionansweringhybridgraphzero-shotQAlargelanguagemodelsquestion-drivenpruningmulti-hopreasoningHybrid-QAOTT-QA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that a question-conditioned Hybrid Graph, built from both a table and its linked passages, lets an LLM answer multi-hop Table-Text QA without any fine-tuning. On Hybrid-QA and OTT-QA, it reports the top zero-shot Exact Match scores among the baselines tested, with EM gains of up to 10 percentage points on Hybrid-QA and 5.4 on OTT-QA, while cutting the reader input to about half the full context. A sympathetic reader would take this as evidence that structured and unstructured evidence can be unified and pruned at the graph level rather than through trained aggregation modules.

What carries the argument

The central object is the Hybrid Graph, which combines a sub-table of relevant columns with a bipartite entity-document graph whose named entities are linked to table cells and passages. The load-bearing mechanism is the question-conditioned Breadth-First Search: question entities are mapped to headers (or 'Others'), matched semantically to graph nodes with a 0.8 threshold, and used as starting points for a three-hop traversal; the visited paths are stored hop-wise and handed to the LLM incrementally so it can answer from the first hop that suffices or return 'None' to trigger expansion. This graph traversal is what filters noise, preserves cross-source connections, and keeps the context small.

What would settle it

On a random sample from the two dev sets, compute whether the gold answer appears among the nodes and linked passages of the 3-hop pruned subgraph that ODYSSEY would hand to the reader; if a substantial share of correctly answered questions has the answer outside that subgraph, or if the scores change sharply when the hop limit or matching threshold is moved on a fresh sample, the pruning assumption is falsified.

Watch

Extended reading notes

Core claim

The authors' central claim is that ODYSSEY, a zero-shot pipeline, answers hybrid table-text questions by first constructing a Hybrid Graph from the retrieved sub-table and named entities in linked passages, then pruning this graph with a question-driven BFS up to three hops, and only then feeding the pruned context to an LLM reader hop by hop. The paper reports that this construction achieves the best zero-shot Exact Match on both datasets among the considered baselines, improving EM by up to 10% on Hybrid-QA and 5.4% on OTT-QA, and that it reduces reader token usage by up to 53% compared to the original table-and-text context. The argument is that most questions can be answered from a small subgraph, and the hop-wise 'None' protocol provides a fallback to full context in the roughly 8-10% of cases where pruning is insufficient.

Load-bearing premise

Everything depends on the answer being reachable within three hops from the single best semantically matched question entity; if that start entity is wrong or the answer lies farther out, the pruned context can miss it with no explicit signal.

Editorial extensions

If this is right

  • If the central claim is right, zero-shot LLM systems can match or beat fine-tuned Table-Text QA models on Hybrid-QA and OTT-QA without needing human-curated training data.
  • Because almost 90% of questions are answered within one or two hops, the context passed to the reader can be kept small, making the approach more affordable and more scalable to long passages.
  • Token reductions of up to 53% compared with the full context directly lower API cost per query, which matters for real deployments.
  • The method improves results across GPT-4, GPT-3.5, and Llama-3, so the benefit comes from graph pruning rather than from a single model's capacity.

Reading between the lines

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

  • A direct test would be measuring subgraph recall: for a random sample, does the gold answer actually lie inside the 3-hop pruned subgraph, independent of the LLM's extraction? If not, the pruning step is the bottleneck, not the reader.
  • The same question-conditioned graph pruning idea could be carried to other heterogeneous evidence, such as knowledge graphs plus text, or images with captions, wherever a small subgraph can be extracted before a multimodal LLM call.
  • The reported 3-hop limit and 0.8 matching threshold were tuned on 50 dev samples, so an extension would be to test sensitivity on fresh samples and across different semantic matchers to see whether the gains are robust.
  • A low-cost improvement could be trying multiple start entities or all entities above the threshold instead of only the best one, which might reduce the silent-failure cases where the true answer lies outside the pruned graph.
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 / 4 minor

Summary. The paper proposes ODYSSEY, a zero-shot, fine-tuning-free approach to table-and-text question answering. The method extracts entities from the question, constructs a hybrid graph that joins a retrieved sub-table with an entity-document graph built from linked passages, and then prunes the graph via semantic matching (instructor-xl, threshold 0.8) followed by a 3-hop BFS traversal. The pruned context is fed to a reader LLM hop-wise, with a fallback to the full table-text if the reader returns 'None'. The authors evaluate on 500-sample subsets of Hybrid-QA and OTT-QA using GPT-4, GPT-3.5, and Llama3-8B, comparing against full-context and summarization baselines, and also compare against fine-tuned and program-based methods. They report EM improvements of up to 10% on Hybrid-QA and 5.4% on OTT-QA, plus token reductions of up to 53%.

Significance. If the central empirical claim holds, the paper is a useful practical contribution: it shows that a relatively simple, training-free graph-pruning pipeline can match or exceed full-context prompting on hybrid table-text QA at lower token cost. The method is clearly described, the ablation of hop-wise versus one-shot retrieval is informative, and the authors are transparent about the fallback mechanism and the 50-sample hyperparameter tuning. The main weaknesses are statistical: the headline EM differences in Table 3 are small relative to the sample size, no significance tests are provided, and the comparison against fine-tuned systems on OTT-QA is not matched. The paper does not directly measure whether the pruning step retains the gold evidence, leaving the mechanism underlying the gains under-validated. Overall, the contribution is potentially valuable but the evidence as presented is not yet strong enough for a definitive acceptance.

major comments (4)
  1. [Section 5, Table 3] The central empirical claim rests on small EM differences that may be within sampling noise. For example, on Hybrid-QA with GPT-4 the full-context baseline scores 55.40 EM and ODYSSEY w/ hopwise scores 58.40, a difference of 3.0 points; on OTT-QA with GPT-4 the corresponding numbers are 58.86 and 62.02, a difference of 3.16 points. With 500 samples, the standard error of each of these differences is roughly 3 percentage points, so the observed gains are not clearly distinguishable from noise. The paper reports no confidence intervals, significance tests, or bootstrap estimates for Table 3. Please provide these, or explicitly present the results as exploratory; without them the abstract's claim of 'best zero-shot performance' is not statistically supported.
  2. [Sections 3.2.3 and 5.4] The paper never measures whether the gold evidence actually survives the pruning step. The single-start, 0.8-threshold BFS can exclude the gold answer while the reader LLM still outputs a confident non-None answer, and the reported ~10% (Hybrid-QA) and ~8% (OTT-QA) fallback rates do not account for this silent-failure channel. The error analysis in Section 5.4 classifies 41 GPT-4 errors into categories such as Expression Mismatch and Entity-Matching, but it does not distinguish 'answer absent from the pruned context' from 'answer present but misread'. Please report the recall of gold evidence in the pruned subgraph (e.g., the fraction of questions where the gold table cell or passage is within the 3-hop context) and add an explicit error category for pruning-induced absence. This is load-bearing because the method's claimed advantage is precisely that pruning preserves the relevant evidence, and without this measurement the EM gains could be driven by the full-context fallback or by the reader's parametric knowledge.
  3. [Tables 4 and 5] The comparison against fine-tuned state-of-the-art systems is not matched. Table 4 evaluates ODYSSEY and HPROPRO on the same 200 sampled Hybrid-QA dev cases using GPT-4-0613, which is reasonable. However, Table 5 compares ODYSSEY on 500 sampled OTT-QA dev cases using GPT-4-1106-preview against published full-dev-set numbers for fine-tuned systems such as COS and CORE. This introduces both subset-selection bias and a model-version mismatch. The claim of outperforming fine-tuned SoTA on OTT-QA is therefore not supported by the present evidence. Please either evaluate on the full dev set (or a common random subset) with the same model versions, or clearly label the comparison as approximate and non-matched.
  4. [Section 4.4 and Figure 4] The two free parameters of the method—the 0.8 semantic-matching threshold and the 3-hop BFS limit—were tuned on 50 dev samples and then evaluated on a different 500-sample dev subset. The paper states that the 50 samples are distinct from the experimental set, which mitigates direct overfitting, but no sensitivity analysis is provided. Since the method's behavior can change sharply with these parameters (e.g., a lower threshold would start traversal from more entities, a higher hop limit would include more context), please report results across a small grid of threshold values and hop limits, or at least show that the main conclusions in Table 3 are stable to reasonable variations. Without such an analysis, the reader cannot assess how robust the reported gains are to the chosen hyperparameters.
minor comments (4)
  1. [Abstract and Conclusion] The abstract's 'up to 10%' and '5.4%' EM improvements are relative improvements over the full-context baseline (e.g., Hybrid-QA GPT-3.5: (44.2-40.22)/40.22 ≈ 9.9%; OTT-QA GPT-4: (62.02-58.86)/58.86 ≈ 5.4%), but the text does not say 'relative'. The absolute EM differences in Table 3 are much smaller, and readers could misinterpret the percentages as absolute gains. Please clarify this in both the abstract and the results section.
  2. [Section 5.4] The error analysis is numbered '5.4' in the main text but is referenced elsewhere as 'Appendix 5.4'; please unify the cross-references. Also, the error categories are useful, but the sample size (100 cases, 41 errors) is small; consider reporting the error analysis on the full 500-sample evaluation set if feasible.
  3. [Global] There are several typographical and formatting issues: 'we presents' in Section 5, 'Positon 4' in Figure 3, and inconsistent use of 'ODYSSEY' versus 'Odyssey' in the conclusion. These do not affect the technical content but should be cleaned up.
  4. [Section 4.4 and Table 6] The token-cost comparison in Table 6 uses OpenAI pricing as of June 2024, which is reasonable, but the cost figures are reported without mentioning that they are based on the GPT-4 Turbo price per 1K tokens. Please state the assumed price explicitly in the table caption or text, and note that the 'Summarized' baseline's token counts may depend on the specific summarization prompt used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is empirical, evaluated against external benchmarks, and the tuned hyperparameters are disclosed and not predictions.

full rationale

The paper makes no derivational claim that reduces to its own inputs. ODYSSEY is an empirical pipeline (question analysis, hybrid graph construction, traversal, reader LLM) evaluated on the Hybrid-QA and OTT-QA benchmarks. The only tuned quantities are the 0.8 semantic-matching threshold and the 3-hop traversal limit, selected on 50 dev samples as stated in Section 4.4: "We select the highest-ranked entity surpassing a 0.8 threshold... These values were determined by experimenting on 50 samples from the dev set." These are hyperparameters, not fitted predictions of the reported EM scores, and the evaluation is performed on a separately sampled 500-instance dev set. The reported EM/F1 comparisons in Tables 3-5 are against external baselines and fine-tuned models, and the token-efficiency claims in Table 6 are directly measured input sizes. There is no self-citation chain, no imported uniqueness theorem, no ansatz smuggled via citation, and no equation by which a result equals its input by construction. The abstract's 'up to 10%' improvement is not directly recoverable from Table 3 (the largest GPT-3.5 Hybrid-QA EM gap versus the full-context baseline in that table is 44.2 - 40.22 = 3.98), but this is an internal reporting inconsistency rather than circularity. Likewise, the skeptic's concern that a wrong start entity or a 3-hop limit could silently exclude gold evidence is a legitimate correctness or evaluation risk, not a circularity of the derivation chain.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the correctness of the graph construction and traversal pipeline, which itself relies on external LLM, NER, and embedding components without component-level validation. The only fitted numbers are the semantic matching threshold and the hop limit, both tuned on a small dev sample.

free parameters (2)
  • Semantic matching threshold = 0.8
    Section 4.4: 'We select the highest-ranked entity surpassing a 0.8 threshold.' Chosen via experiments on 50 dev samples.
  • BFS hop limit = 3
    Section 3.2.4: 'We limit the process to 3 hops due to minimal improvements observed in LLM performance beyond that during hyperparameter tuning.' Tuned on 50 Hybrid-QA dev samples.
assumptions (4)
  • domain assumption A 3-hop breadth-first traversal from the single best matched question entity contains the answer for the large majority of questions
    Section 3.2.3 and 3.2.4 rely on it; fallback to full context occurs in only 8-10% of cases, but silent failures beyond 3 hops are not analyzed.
  • domain assumption GPT-3.5-turbo can reliably extract question entities, relevant table headers, and entity-header mappings
    Question Analysis in Section 3.2.1 is the sole source of graph start points; error analysis attributes 14 of 41 errors to semantic module failures.
  • domain assumption SpaCy transformer NER and instructor-xl embeddings provide accurate entity and semantic matches
    Used in graph construction and traversal as described in Section 4.4; no separate component-level accuracy is reported.
  • domain assumption The reader LLM's 'None' output is a trustworthy signal of insufficient context
    Hop-wise expansion in Section 3.2.4 depends on this; false 'None' or false non-'None' outputs are not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Graphs for Table-and-Text based Question Answering using LLMs." pith.science (2026). https://pith.science/paper/B6KSXQLP

@misc{pith2026250117767,
  author       = {Pith},
  title        = {Pith review of: Hybrid Graphs for Table-and-Text based Question Answering using LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6KSXQLP}},
  note         = {Machine review of arXiv:2501.17767}
}
read the original abstract

Answering questions that require reasoning and aggregation across both structured (tables) and unstructured (raw text) data sources presents significant challenges. Current methods rely on fine-tuning and high-quality, human-curated data, which is difficult to obtain. Recent advances in Large Language Models (LLMs) have shown promising results for multi-hop question answering (QA) over single-source text data in a zero-shot setting, yet exploration into multi-source Table-Text QA remains limited. In this paper, we present a novel Hybrid Graph-based approach for Table-Text QA that leverages LLMs without fine-tuning. Our method constructs a unified Hybrid Graph from textual and tabular data, pruning information based on the input question to provide the LLM with relevant context concisely. We evaluate our approach on the challenging Hybrid-QA and OTT-QA datasets using state-of-the-art LLMs, including GPT-3.5, GPT-4, and LLaMA-3. Our method achieves the best zero-shot performance on both datasets, improving Exact Match scores by up to 10% on Hybrid-QA and 5.4% on OTT-QA. Moreover, our approach reduces token usage by up to 53% compared to the original context.

Figures

Figures reproduced from arXiv: 2501.17767 by the authors.

Figure 1
Figure 1. Multi-Dimensional Improvements: Our method (with GPT-4 as reader LLM) demonstrates su￾perior results on Hybrid-QA and OTT-QA. Metrics used: EM: Exact-Match with the gold answer, F1-Score, Query Info Efficiency: normalized metric ( 1 Input Token Size ) that quantifies the efficiency of using fewer input tokens to represent the same documents, w.r.t. reader LLM. the person who finished the Berlin marathon in 2:13.32 i… view at source ↗
Figure 2
Figure 2. Case study on Hybrid-QA: Comparison of our method (ODYSSEY) against various baselines on an example from the Hybrid-QA dataset. Baselines: (i) Question + Context: Providing the LLM only the question without any additional context (ii) Question + Summarized Context: Passing the question along with the summarized documents and table. Our method delivers accurate answer because the Hybrid Graph efficiently connects "St… view at source ↗
Figure 3
Figure 3. Overview of the ODYSSEY framework. Our method comprises of 3 steps: i) Question Analysis, ii) Hybrid Graph Construction, and iii) Hybrid Graph Traversal. First, we begin with Question Analysis (⃝1a in the figure) from where we get question entities, retrieved sub-table, and entity-header mapping. Next, we construct the Entity-Document Graph (⃝1b in the figure). Using entity-doc graph and retrieved sub-table, we cons… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Hopwise analysis: For ODYSSEY (our method w/ hopwise), we calculate the cumulative EM score (left-side in figure) and average token size (right￾side in figure) utilized by each hop for Llama3-8B, GPT￾3.5, and GPT-4 on Hybrid-QA. Bars in left-side of the figure denotes …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 22 canonical work pages

  1. [1]

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

    arXiv 2023

  2. [2]

    Wenhu Chen, Ming-Wei Chang, Eva Schlinger, William Yang Wang, and William W Cohen. 2020 a . Open question answering over tables and text. In International Conference on Learning Representations

    work page 2020

  3. [3]

    Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. 2020 b . Hybridqa: A dataset of multi-hop question answering over tabular and textual data. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1026--1036

    work page 2020

  4. [4]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    arXiv 2018

  5. [5]

    Julian Martin Eisenschlos, Maharshi Gor, Thomas M \"u ller, and William W Cohen. 2021. Mate: multi-view attention for table transformer efficiency. arXiv preprint arXiv:2109.04312

    work page Pith review arXiv 2021

  6. [6]

    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

    arXiv 2023

  7. [7]

    Jonathan Herzig, Pawe Krzysztof Nowak, Thomas M \"u ller, Francesco Piccinno, and Julian Martin Eisenschlos. 2020. Tapas: Weakly supervised table parsing via pre-training. arXiv preprint arXiv:2004.02349

    arXiv 2020

  8. [8]

    Minghao Hu, Yuxing Peng, Zhen Huang, Xipeng Qiu, Furu Wei, and Ming Zhou. 2018. Reinforced mnemonic reader for machine reading comprehension. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4099--4106

    work page 2018

Show all 46 references
  1. [9]

    Binyuan Hui, Ruiying Geng, Lihan Wang, Bowen Qin, Bowen Li, Jian Sun, and Yongbin Li. 2022. s ^ 2 SQL : Injecting syntax to question-schema interaction graph encoder for text-to- SQL parsers. arXiv preprint arXiv:2203.06958

    2022 arXiv

  2. [10]

    Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Structgpt: A general framework for large language model to reason over structured data. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9237--9251

    2023

  3. [11]

    Hanlei Jin, Yang Zhang, Dan Meng, Jun Wang, and Jinghua Tan. 2024. A comprehensive survey on process-oriented automatic text summarization with exploration of llm-based methods. arXiv preprint arXiv:2403.02901

    2024

  4. [12]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)...

    2020

  5. [13]

    Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 39--48

    2020

  6. [14]

    Vishwajeet Kumar, Yash Gupta, Saneem Chemmengath, Jaydeep Sen, Soumen Chakrabarti, Samarth Bharadwaj, and Feifei Pan. 2023. Multi-row, multi-span distant supervision for table+ text question answering. In Proceedings of the 61st Annual Meeting of the Association for Computatio...

    2023

  7. [15]

    Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ili \'c , Daniel Hesslow, Roman Castagn \'e , Alexandra Sasha Luccioni, Fran c ois Yvon, Matthias Gall \'e , et al. 2023. Bloom: A 176b-parameter open-access multilingual language model

    2023

  8. [16]

    Sung-Min Lee, Eunhwan Park, Daeryong Seo, Donghyeon Jeon, Inho Kang, and Seung-Hoon Na. 2023. Mafid: Moving average equipped fusion-in-decoder for question answering over tabular and textual data. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2...

    2023

  9. [17]

    Fangyu Lei, Xiang Li, Yifan Wei, Shizhu He, Yiming Huang, Jun Zhao, and Kang Liu. 2023. S3hqa: A three-stage approach for multi-hop text-table hybrid question answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short...

    2023

  10. [18]

    Haoyang Li, Jing Zhang, Cuiping Li, and Hong Chen. 2023 a . Resdsql: Decoupling schema linking and skeleton parsing for text-to-sql. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 13067--13075

    2023

  11. [19]

    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

  12. [20]

    Ruosen Li and Xinya Du. 2023. Leveraging structured information for explainable multi-hop question answering and reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6779--6789

    2023

  13. [21]

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

    2019 arXiv

  14. [22]

    Kaixin Ma, Hao Cheng, Xiaodong Liu, Eric Nyberg, and Jianfeng Gao. 2022. Open-domain question answering via chain of reasoning over heterogeneous knowledge. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5360--5374

    2022

  15. [23]

    Kaixin Ma, Hao Cheng, Yu Zhang, Xiaodong Liu, Eric Nyberg, and Jianfeng Gao. 2023. Chain-of-skills: A configurable model for open-domain question answering. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), page...

    2023

  16. [24]

    Will Orr and Kate Crawford. 2023. The social construction of datasets: On the practices, processes and challenges of dataset creation for machine learning

    2023

  17. [25]

    Liangming Pan, Wenhu Chen, Wenhan Xiong, Min-Yen Kan, and William Yang Wang. 2021. Unsupervised multi-hop question answering by question generation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lang...

    2021

  18. [26]

    Ethan Perez, Patrick Lewis, Wen-tau Yih, Kyunghyun Cho, and Douwe Kiela. 2020. Unsupervised question decomposition for question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8864--8880

    2020

  19. [27]

    Bryan Perozzi, Bahare Fatemi, Dustin Zelle, Anton Tsitsulin, Mehran Kazemi, Rami Al-Rfou, and Jonathan Halcrow. 2024. Let your graph do the talking: Encoding structured data for llms. arXiv preprint arXiv:2402.05862

    2024 arXiv

  20. [28]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1--67

    2020

  21. [29]

    Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

    2009

  22. [30]

    Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. 2016. Bidirectional attention flow for machine comprehension. arXiv preprint arXiv:1611.01603

    2016 arXiv

  23. [31]

    Yeon Seonwoo, Ji-Hoon Kim, Jung-Woo Ha, and Alice Oh. 2020. Context-aware answer extraction in question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2418--2428

    2020

  24. [32]

    Qi Shi, Han Cui, Haofeng Wang, Qingfu Zhu, Wanxiang Che, and Ting Liu. 2024. Exploring hybrid question answering via program-based prompting. arXiv preprint arXiv:2402.10812

    2024 arXiv

  25. [33]

    Haitian Sun, William W Cohen, and Ruslan Salakhutdinov. 2021. Iterative hierarchical attention for answering complex questions over long documents. arXiv preprint arXiv:2106.00200

    2021 arXiv

  26. [34]

    Xiaoyu Tan, Haoyu Wang, Xihe Qiu, Yuan Cheng, Yinghui Xu, Wei Chu, and Yuan Qi. 2024. Struct-x: Enhancing large language models reasoning with structured data. arXiv preprint arXiv:2407.12522

    2024 arXiv

  27. [35]

    Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805

    2023 arXiv

  28. [36]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vo...

    2023

  29. [37]

    Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. Rat-sql: Relation-aware schema encoding and linking for text-to-sql parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7567--7578

    2020

  30. [38]

    Yingyao Wang, Junwei Bao, Chaoqun Duan, Youzheng Wu, Xiaodong He, and Tiejun Zhao. 2022. Muger2: Multi-granularity evidence retrieval and reasoning for hybrid question answering. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6687--6697

    2022

  31. [39]

    Zhiruo Wang, Haoyu Dong, Ran Jia, Jia Li, Zhiyi Fu, Shi Han, and Dongmei Zhang. 2021. Tuta: Tree-based transformers for generally structured table pre-training. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1780--1790

    2021

  32. [40]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  33. [41]

    Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I Wang, et al. 2022. Unifiedskg: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. In Proceedings ...

    2022

  34. [42]

    Tianshu Zhang, Xiang Yue, Yifei Li, and Huan Sun. 2023. Tablellama: Towards open large generalist models for tables. arXiv preprint arXiv:2311.09206

    2023 arXiv

  35. [43]

    Bertscore: Evaluating text generation with bert

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations

  36. [44]

    Wanjun Zhong, Junjie Huang, Qian Liu, Ming Zhou, Jiahai Wang, Jian Yin, and Nan Duan. 2022. Reasoning over hybrid chain for table-and-text open domain qa. arXiv preprint arXiv:2201.05880

    2022 arXiv

  37. [45]

    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...

  38. [46]

    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 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.