Pith. sign in

REVIEW 4 major objections 6 minor 28 references

DailyQA: A Benchmark to Evaluate Web Retrieval Augmented LLMs Based on Capturing Real-World Changes

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Reranking web results is the key to AI on date-specific queries

desk verdict DailyQA's automated pipeline is a real step forward, but the empirical claims rest on a single week and unverified Wikipedia infobox ground truth. read the letter →

arxiv 2505.17162 v1 pith:2AWSG5SZ submitted 2025-05-22 cs.IR cs.AI

Jiehan Cheng , Zhicheng Dou This is my paper · ORCID
classification cs.IRcs.AI
keywords time-sensitivequestionansweringretrieval-augmentedgenerationdynamicbenchmarkWikipediarevisionlogswebsearchrerankingLLMevaluationtemporalreasoning
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

DailyQA is a benchmark that measures whether large language models can keep up with facts that change over time. It is built automatically from Wikipedia revision logs: whenever an infobox value changes, that change becomes the seed of a question whose correct answer is the new value. Each week the benchmark adds a fresh set of questions, and each question has an answer for every date, taken from the last revision of the infobox before that date. Testing several open- and closed-source LLMs with web-search-augmented RAG pipelines, the paper finds that reranking the retrieved web documents outperforms using search snippets or raw document order, while adding date strings to queries or to reranking does not help. Overall accuracy remains far from perfect, especially for frequently updated facts, indicating that time-sensitive web QA remains a challenge.

What carries the argument

The central mechanism is the Wikipedia infobox revision log. Each revision that changes a key-value pair in an infobox is treated as a fact change; the changed value becomes the gold answer, and the page's title, first paragraph, and full infobox serve as background for query generation. This converts real-world updates into dated question-answer pairs at scale and gives every query an answer for any date by taking the last revision of the infobox before that date.

What would settle it

Manually audit a random sample of 200 DailyQA queries by looking up the answer for the specified date in an independent authoritative source; if more than a handful of gold answers are wrong or ambiguous, the benchmark's central claim of correctness collapses.

Watch

Extended reading notes

Core claim

DailyQA is an automatically updated, dynamic question-answering benchmark. It mines Wikipedia revision logs daily, focusing on changes to structured infobox fields, and generates one-hop questions whose answers are those field values. The pipeline filters for pages with recent structured edits, uses an LLM to generate descriptive questions, checks that each question can be answered from the referenced page and found via web search, and labels queries by update frequency and domain. For every query, the answer on any given date is the infobox value from the last revision before that date. On a sample week, the authors evaluate Qwen, DeepSeek-R1-Distill, and GPT-4o-mini under several RAG pipelines and report that reranking web-retrieved documents outperforms using search snippets or the raw web order, while time-based reranking does not improve results; all models perform below ceiling, especially on frequently updated queries.

Load-bearing premise

The benchmark assumes that each gold answer, taken from the value of a Wikipedia infobox field on the specified date, is factually correct and unambiguous; if that infobox value is wrong, stale, or contested, the benchmark's ground truth is wrong.

Editorial extensions

If this is right

  • Because the pipeline is fully automated, DailyQA can be refreshed weekly at low cost, avoiding the staleness of static time-sensitive QA datasets.
  • Reranking web-retrieved documents, rather than using search snippets or raw document order, is the strongest lever for improving date-specific QA performance among the pipelines tested.
  • Frequently updated queries are substantially harder than infrequently updated ones for every model tested, so progress on time-sensitive QA should be reported separately for these two difficulty buckets.
  • Adding explicit date strings to search queries or to the reranker does not help and can hurt, so better ways of injecting time awareness are needed.
  • Different model families lead in different domains, which suggests that multi-model collaboration could improve cross-domain time-sensitive QA.

Reading between the lines

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

  • One natural extension is to validate the infobox-derived gold answers against independent authoritative sources; if such validation passes, DailyQA could become a reusable benchmark for tracking LLM progress over time.
  • The failure of simple time signals hints that agentic retrieval, where the model plans queries and filters documents based on content-derived dates, may be more promising than current heuristic rerankers.
  • Since the pipeline runs on Wikipedia revision logs, it could be ported to other languages or to structured databases like Wikidata, which would test whether the reranking conclusion generalizes across languages and change rates.
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 / 6 minor

Summary. The paper introduces DailyQA, a dynamic benchmark for evaluating web-retrieval-augmented LLMs on time-sensitive questions. The benchmark is constructed automatically from Wikipedia infobox revision logs: each week, a new query set is generated by Qwen-72B-Instruct from recently changed infobox values, with a daily-updated answer for each query obtained from the last revision before a requested date. The pipeline includes rule-based filtering, LLM-based correctness and descriptiveness checks, domain classification, and update-frequency classification. The authors evaluate open-source and closed-source LLMs with DuckDuckGo search and a bge-v2-m3 reranker across several RAG pipelines (no search, snippets, raw documents, reranked chunks, and time-aware reranking). The main findings are that web retrieval helps substantially, reranking improves over raw document order, time-aware reranking does not help, and frequent-update queries are harder than infrequent ones.

Significance. If the benchmark labels are reliable, DailyQA fills a real gap: existing time-sensitive QA benchmarks are mostly static or small, whereas DailyQA is designed to refresh weekly and to provide per-day answers at low cost. The automated pipeline is a practical contribution, and the paper makes a falsifiable prediction about reranking being important for this task. The inclusion of multiple LLMs, a commercial search service (perplexity.ai), and domain/frequency breakdowns is useful. However, the central validity of the benchmark depends on Wikipedia infobox values being factually correct and temporally precise on each date, and the experimental evidence is limited to one week and one evaluation date. The strengths are the transparency of the pipeline and the clear task definition; the weaknesses are the lack of human validation of labels and the absence of statistical rigor in the headline claims.

major comments (4)
  1. [Answer Extraction / Ethics] The gold answers are defined as the value of a Wikipedia infobox field at the last revision before the query date, with no verification against independent sources. The quality check in 'Query Generation and Quality Check' asks Qwen-72B to answer using the same infobox as reference, so it cannot detect factually wrong, stale, or vandalized infobox values. The Ethics section acknowledges that recently updated pages may contain unverified information but asserts without evidence that infoboxes mitigate this risk. Because all SM/Acc numbers in Tables 1-4 are computed against these labels, a nontrivial fraction of incorrect or ambiguous infobox values would change every headline conclusion. The authors should provide a human-validated sample (e.g., 100-200 query-date pairs with inter-annotator agreement) and report the error rate found; ideally, they should also cross-check a subset against independent sources.
  2. [Implementation Details / Results] All experimental results come from a single weekly update (W-2025-01-12) and a single target date (2025-01-19), with no repeated runs, no error bars, and no significance tests. The effect sizes underlying the main claim are small: for Qwen2.5-72B-Instruct with Search w/o Time, Rerank improves SM from 0.479 to 0.502 and Acc from 0.492 to 0.513 relative to Doc. Without variance estimates or multiple evaluation dates, these differences may not be statistically significant, and the statement that reranking is 'critical' is not supported. The authors should either report results across several weekly updates and repeated runs, or provide significance testing on the current single-week data.
  3. [Results, Main Results] The claim that 'rerank of web retrieval results is critical' (Abstract) is disproportionate to the observed effect sizes. In Table 1, the largest Rerank-vs-Doc gain is about 5 Acc points (Search w/ Time), while the Search w/o Time Rerank gain is about 2 points, and Rerank-T is worse than Rerank in all configurations. The paper's main practical recommendation should be calibrated to these effect sizes or substantiated with stronger evidence. This is load-bearing because the reranking finding is presented as a key takeaway for RAG pipeline design.
  4. [Metrics / Query Generation and Quality Check] The evaluation chain depends on LLMs at three points: Qwen-72B generates queries, Qwen-72B validates correctness against the infobox gold, and GPT-4o judges final accuracy using a 5-point similarity scale. No human agreement is reported for either the query validity checks or the final Acc judgments. Because the generating and validating model is from the same family (Qwen) that is later evaluated, there is a risk of systematic bias in query difficulty or judge leniency. The authors should include a human evaluation of a sample and report the distribution of GPT-4o scores, not just the thresholded Acc.
minor comments (6)
  1. [Throughout] There are numerous typos and grammatical errors, including 'benckmark', 'bertter', 'thar', 'fastly', 'Seach w/ Time' in Table 1, 'an weekly', and 'We statistic'. The paper would benefit from a careful proofreading pass.
  2. [Classification / Figures 3-4] The text says 'present it in Figure 3' when referring to the number of answer changes per query, but Figure 4 is the figure showing the percentage of queries with different answer change times. Please fix the cross-reference.
  3. [Classification] The phrase 'label queries that do not change from day 8' is ambiguous: it should specify whether day 8 is measured from the start date of the weekly update and whether the classification uses post-evaluation-date changes, which would make the frequent/infrequent split retrospective.
  4. [Table 2] The caption of Table 2 does not define SM1 and SM2; please define them as the subset-match scores on frequent-update and infrequent-update queries, respectively.
  5. [Prompt for RAG] The RAG prompt instructs the model to answer using the original text of the reference without modification, generalization, or summarization. This may be overly restrictive for numeric answers and could artificially lower performance; please justify this design choice or relax it.
  6. [Figure 1] Figure 1 is confusing: it shows the same numeric answer repeated with different dates, which makes the daily-update property hard to read. Please clarify the visualization or provide an example with a changing answer.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DailyQA labels are anchored to external Wikipedia infobox revisions and the evaluation is an empirical web-RAG comparison.

full rationale

The paper's central construction is dataset building, not a fitted derivation. Queries are generated from Wikipedia infobox values, and the gold answer for any date is defined as the last infobox value before that date. This anchors the benchmark to an external, independently editable source (Wikipedia revision logs) rather than to the model outputs being evaluated. The quality-check step uses Qwen-72B to confirm that the generated query is answerable from the golden infobox, but this is a filtering step, not a prediction target: the final evaluation never provides the golden infobox to the tested models and instead relies on DuckDuckGo web retrieval. The headline claims (rerank matters, LLMs struggle with time-sensitive information) are derived from empirical comparisons across pipelines and models on this externally anchored benchmark, with no fitted parameters or self-citation chain carrying the argument. The paper contains no self-citations from the present authors, and no uniqueness theorem or ansatz is imported from prior work. The unverified factual accuracy of some infobox values is a data-quality concern that could affect the benchmark's validity, but it is not circularity because the labels are not constructed from the models' predictions or from the paper's own conclusions.

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

The benchmark rests on domain assumptions about Wikipedia's reliability, LLM-based quality control, and the representativeness of the chosen search engine and judge. There are no fitted parameters in a derivation, but several hand-set design choices (domain quota, top-k, evaluation date, accuracy threshold, frequent/infrequent cutoff) shape the results.

free parameters (5)
  • domain quota = 750 per week
    To balance domain distribution, the authors cap each domain at 750 queries per weekly update, which changes the composition of the benchmark and the difficulty mix.
  • top_k retrieved references = 12
    All RAG pipelines use the top 12 snippets, documents, or chunks as references; results depend on this choice.
  • evaluation week and date = Week starting 2025-01-12, query date 2025-01-19
    All experimental results come from a single evaluation week and a single target date, limiting generalizability.
  • accuracy threshold = GPT-4o scores 4 and 5 counted as correct
    The model-based Acc metric treats similarity scores 4 or 5 as correct; this threshold affects all Acc numbers.
  • frequent/infrequent cutoff = no change after day 8 is infrequent
    Queries are labeled frequent if they change on or after day 8 relative to the start date; this arbitrary cutoff defines the difficulty analysis.
assumptions (5)
  • domain assumption Wikipedia infobox values are timely, complete, and accurate enough to serve as ground truth for the answers.
    The pipeline treats infobox changes as real-world factual changes and uses the latest pre-date infobox value as the correct answer, without cross-checking other sources. Section: Wiki Data Collection and Processing, Answer Extraction.
  • domain assumption LLM-generated queries are unambiguous and answerable after the automatic quality check, despite no human verification.
    The correctness check asks Qwen-72B to answer from the golden documents and requires subset match; this assumes the LLM judge is reliable. Section: Query Generation and Quality Check.
  • domain assumption DuckDuckGo search results are a representative web retrieval source for evaluating RAG systems in this benchmark.
    All evaluation uses DuckDuckGo's API; results may differ on other search engines. Section: Implementation Details.
  • domain assumption GPT-4o's similarity judgments provide a valid accuracy measure for answers.
    The model-based Acc metric relies on GPT-4o comparing predictions to standard answers; no human evaluation is reported. Section: Metrics.
  • domain assumption One-hop questions are sufficient to measure time-sensitivity without confounding by multi-step reasoning.
    The authors state this as a design goal; it limits the scope of what DailyQA measures. Section: Limitations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DailyQA: A Benchmark to Evaluate Web Retrieval Augmented LLMs Based on Capturing Real-World Changes." pith.science (2026). https://pith.science/paper/2AWSG5SZ

@misc{pith2026250517162,
  author       = {Pith},
  title        = {Pith review of: DailyQA: A Benchmark to Evaluate Web Retrieval Augmented LLMs Based on Capturing Real-World Changes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AWSG5SZ}},
  note         = {Machine review of arXiv:2505.17162}
}
read the original abstract

We propose DailyQA, an automatically updated dynamic dataset that updates questions weekly and contains answers to questions on any given date. DailyQA utilizes daily updates from Wikipedia revision logs to implement a fully automated pipeline of data filtering, query generation synthesis, quality checking, answer extraction, and query classification. The benchmark requires large language models (LLMs) to process and answer questions involving fast-changing factual data and covering multiple domains. We evaluate several open-source and closed-source LLMs using different RAG pipelines with web search augmentation. We compare the ability of different models to process time-sensitive web information and find that rerank of web retrieval results is critical. Our results indicate that LLMs still face significant challenges in handling frequently updated information, suggesting that DailyQA benchmarking provides valuable insights into the direction of progress for LLMs and RAG systems.

Figures

Figures reproduced from arXiv: 2505.17162 by the authors.

Figure 1
Figure 1. A example for DailyQA. The answer to “LeBron [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our DailyQA dataset construction pipeline, which includes filtration and process of the raw data (Wiki [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The number of answer changes relative to the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Percentage of queries with different answer change [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: An example of an infobox from a wikipedia page (left), and its processed data struc￾ture (right). The infobox is from the wikipedi https://en.wikipedia.org/wiki/Battle of Chakdara. As shown in [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 9
Figure 9. Figure 9: Examples for generated queries in different do [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 7
Figure 7. Figure 7: Prompt for Query Generation Prompt for RAG This is the retrieval augmented generation scenarios where you need to answer questions based on the references. Note that the information in the references is time￾sensitive, and your answer should match the time requirements…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 13 canonical work pages

  1. [1]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al

    Brown, T.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J. D.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877--1901

    2020

  3. [3]

    Chatterjee, S.; Mackie, I.; and Dalton, J. 2024. DREQ: Document Re-ranking Using Entity-Based Query Understanding. In European Conference on Information Retrieval, 210--229. Springer

    work page 2024

  4. [4]

    Chen, W.; Wang, X.; and Wang, W. Y. 2021. A dataset for answering time-sensitive questions. arXiv preprint arXiv:2108.06314

    arXiv 2021

  5. [5]

    Devlin, J. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    arXiv 2018

  6. [6]

    R.; Eisenschlos, J

    Dhingra, B.; Cole, J. R.; Eisenschlos, J. M.; Gillick, D.; Eisenstein, J.; and Cohen, W. W. 2021. Time-Aware Language Models as Temporal Knowledge Bases. Transactions of the Association for Computational Linguistics, 10: 257--273

    work page 2021

  7. [7]

    Gade, A.; and Jetcheva, J. 2024. It's About Time: Incorporating Temporality in Retrieval Augmented Language Models. arXiv preprint arXiv:2401.13222

    arXiv 2024

  8. [8]

    Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    arXiv 2025

Show all 28 references
  1. [9]

    Guo, T.; Zhang, T.; Wu, H.; Li, H.; Qiao, R.; and Sun, X. 2024. Multimodal Label Relevance Ranking via Reinforcement Learning. In European Conference on Computer Vision, 391--408. Springer

    2024

  2. [10]

    Kanhabua, N.; and N rv a g, K. 2012. Learning to rank search results for time-sensitive queries. In Proceedings of the 21st ACM International Conference on Information and Knowledge Management, CIKM '12, 2463–2466. New York, NY, USA: Association for Computing Machinery. ISBN 9...

    2012

  3. [11]

    Karpukhin, V.; O g uz, B.; Min, S.; Lewis, P.; Wu, L.; Edunov, S.; Chen, D.; and Yih, W.-t. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906

    2020 arXiv

  4. [12]

    L.; Asai, A.; Yu, X

    Kasai, J.; Sakaguchi, K.; Takahashi, Y.; Bras, R. L.; Asai, A.; Yu, X. V.; Radev, D. R.; Smith, N. A.; Choi, Y.; and Inui, K. 2022. RealTime QA: What's the Answer Right Now? ArXiv, abs/2207.13332

    2022 arXiv

  5. [13]

    Khattab, O.; and Zaharia, M. 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, 39--48

    2020

  6. [14]

    u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \

    Lewis, P.; Perez, E.; Piktus, A.; Petroni, F.; Karpukhin, V.; Goyal, N.; K \"u ttler, H.; Lewis, M.; Yih, W.-t.; Rockt \"a schel, T.; et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33: 9459--9474

    2020

  7. [15]

    Luo, R.; Sun, L.; Xia, Y.; Qin, T.; Zhang, S.; Poon, H.; and Liu, T.-Y. 2022. BioGPT: generative pre-trained transformer for biomedical text generation and mining. Briefings in bioinformatics, 23(6): bbac409

    2022

  8. [16]

    McDonald, R.; Brokos, G.; and Androutsopoulos, I. 2018. Deep Relevance Ranking Using Enhanced Document-Query Interactions. In Riloff, E.; Chiang, D.; Hockenmaier, J.; and Tsujii, J., eds., Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, ...

    2018

  9. [17]

    M.; Alghisi, S.; and Riccardi, G

    Mousavi, S. M.; Alghisi, S.; and Riccardi, G. 2024. DyKnow: Dynamically Verifying Time-Sensitive Factual Knowledge in LLMs. In Conference on Empirical Methods in Natural Language Processing

    2024

  10. [18]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PMLR

    2021

  11. [19]

    Ramesh, A.; Pavlov, M.; Goh, G.; Gray, S.; Voss, C.; Radford, A.; Chen, M.; and Sutskever, I. 2021. Zero-shot text-to-image generation. In International conference on machine learning, 8821--8831. Pmlr

    2021

  12. [20]

    E.; and Zaragoza, H

    Robertson, S. E.; and Zaragoza, H. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Found. Trends Inf. Retr., 3(4): 333--389

    2009

  13. [21]

    Salinas, D.; Flunkert, V.; Gasthaus, J.; and Januschowski, T. 2020. DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International journal of forecasting, 36(3): 1181--1191

    2020

  14. [22]

    R.; Cole-Lewis, H.; et al

    Singhal, K.; Tu, T.; Gottweis, J.; Sayres, R.; Wulczyn, E.; Amin, M.; Hou, L.; Clark, K.; Pfohl, S. R.; Cole-Lewis, H.; et al. 2025. Toward expert-level medical question answering with large language models. Nature Medicine, 1--8

    2025

  15. [23]

    T.; and Chen, Z

    Siyue, Z.; Xue, Y.; Zhang, Y.; Wu, X.; Luu, A. T.; and Chen, Z. 2024. MRAG: A Modular Retrieval Framework for Time-Sensitive Question Answering. ArXiv, abs/2412.15540

    2024 arXiv

  16. [24]

    N.; Saeidi, A.; Handa, D.; Seth, A.; Son, T

    Uddin, M. N.; Saeidi, A.; Handa, D.; Seth, A.; Son, T. C.; Blanco, E.; Corman, S.; and Baral, C. 2024. UnSeenTimeQA: Time-Sensitive Question-Answering Beyond LLMs' Memorization. ArXiv, abs/2407.03525

    2024 arXiv

  17. [25]

    Vu, T.; Iyyer, M.; Wang, X.; Constant, N.; Wei, J.; Wei, J.; Tar, C.; Sung, Y.-H.; Zhou, D.; Le, Q.; and Luong, T. 2023. FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation. In Annual Meeting of the Association for Computational Linguistics

    2023

  18. [26]

    Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; et al. 2024 a . Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  19. [27]

    Yang, W.; Li, Y.; Fang, M.; and Chen, L. 2024 b . Enhancing Temporal Sensitivity and Reasoning for Time-Sensitive Question Answering. In Al-Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Findings of the Association for Computational Linguistics: EMNLP 2024, 14495--14508. Miam...

    2024

  20. [28]

    J.; and Choi, E

    Zhang, M. J.; and Choi, E. 2021. SituatedQA: Incorporating extra-linguistic contexts into QA. arXiv preprint arXiv:2109.06157

    2021 arXiv

Pith tools

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