Pith. sign in

REVIEW 4 major objections 4 minor 43 references

This paper claims that a small, task-specialized language model fine-tuned on human-validated rule-based extractions can match hand-crafted news-scraping libraries and outperform much larger zero-shot LLMs at converting news HTML into clean

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 07:53 UTC pith:5HWQAOF7

load-bearing objection Useful recipe for distilling rule-based extraction into a small model, but the evaluation's Fundus-derived reference makes the headline superiority claims unverified. the 4 major comments →

arxiv 2607.21284 v1 pith:5HWQAOF7 submitted 2026-07-23 cs.CL

news-crawler-LM: A Small Long-Context Model For High-Quality News Crawling

Pascal Stolzenburg , Jonas Golde , Max Dallabetta , Alan Akbik This is my paper
classification cs.CL
keywords news crawlingHTML-to-text extractioncontent extractionrule-based supervisionfine-tuninglong-context language modelcontrastive learningstructured JSON extraction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that a small, task-specialized language model can replace both hand-written per-site rules and expensive large zero-shot models for extracting clean article text and structured metadata from news HTML. It trains a compact model on roughly 100k rule-generated, human-validated input–output pairs from a news-crawling library and adds a contrastive objective to keep long outputs stable. On held-out publishers the model reports the best overlap-based scores on HTML-to-plaintext and HTML-to-JSON tasks, beating rule-based libraries and large zero-shot LLMs, though it concedes that on plaintext extraction the margin over rule-based libraries is small and that the reference outputs come from the same rules used for training.

Core claim

The central claim is that high-quality rule-based extraction behavior can be distilled into a small long-context model. news-crawler-LM converts raw HTML into plaintext and structured JSON with fields such as title, authors, date, body, and topics; on test publishers unseen during training it achieves the best BLEU, METEOR, ROUGE-L, Levenshtein, and Jaro-Winkler scores among all compared systems in HTML-to-plaintext, and the best overlap-based scores in HTML-to-JSON. The paper reports +4.8 BLEU and +6.1 METEOR over baselines on the HTML-to-Markdown task, and +2.2 BLEU and +4.1 METEOR on HTML-to-JSON. It also shows that zero-shot application of 32B-parameter LLMs performs poorly, that standar

What carries the argument

The mechanism is a training recipe rather than a single theorem: (1) roughly 100k input–output pairs created by Fundus extraction rules across 93 publishers and 25 languages, filtered to 24k input tokens; (2) ReaderLM-v2, a long-context transformer backbone pre-trained on HTML-to-markdown, fine-tuned with LoRA and a standard language-modeling loss; (3) a SimCTG contrastive term added to the loss, which penalizes redundant token representations and is credited for lower 5-gram repetition and more stable long outputs. The pairs carry the argument: they are what lets a small model imitate rule-based extraction decisions across unseen publishers.

Load-bearing premise

The load-bearing premise is that the Fundus rule outputs used as training references are correct, complete, and representative; if those rules are idiosyncratic, the model's edge simply reflects agreement with that one extractor.

What would settle it

A human-annotated gold set of, say, 200 news pages from unseen publishers, scored by independent extractors; news-crawler-LM's claim would be falsified if a rule-based library or a similarly small fine-tuned model matches or beats it on that gold, or if human raters find that its JSON/plaintext adds plausible details absent from the HTML more often than baselines do.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Hand-crafted per-publisher scraping rules can be compiled into a small model, so scaling to new publishers no longer requires writing and maintaining rules per site.
  • A model at the 1.5B-parameter scale can outperform 32B zero-shot LLMs on this task, making generative crawling practical in terms of compute.
  • Zero-shot general-purpose LLMs should not be assumed adequate for HTML structuring; task-specific fine-tuning is the main driver of quality.
  • For plaintext-only extraction, simple rule-based libraries remain a competitive baseline, so machine learning is not clearly necessary in the simplest setting.
  • Contrastive training reduces repetition and degeneration in structured HTML-to-text generation, a known failure mode of autoregressive models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported gains are, by the paper's own admission, measured against references produced by the same rule system used to generate training data; the headline numbers therefore measure fidelity to that system more than absolute extraction quality. A human-judged gold standard would be needed to know whether the model produces better articles than a generic boilerplate remover.
  • If the findings transfer, the same distillation recipe could be applied to other rule-heavy document types (product listings, legal filings, academic papers) wherever a reliable rule extractor exists to serve as teacher.
  • A direct experimental test of the generalization story would be to train the same model on half the supported publishers and evaluate on the other half with human-annotated gold; a model that still wins would separate the value of the supervision signal from the bias of the reference.
  • The authors note inference can take minutes for one long HTML document; an obvious extension is to apply the same supervision strategy to linear-attention or mixture-of-experts backbones to preserve quality at production-level latency.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces news-crawler-LM, a compact long-context language model built on the ReaderLM-v2 backbone with LoRA fine-tuning and a SimCTG contrastive objective, trained on HTML-to-plaintext and HTML-to-JSON pairs derived from the Fundus rule-based extraction library. The model is evaluated on held-out publishers against rule-based parsers (news-please, Boilerpipe, and, nominally, Trafilatura) and several LLM baselines, using string-overlap metrics (BLEU, METEOR, ROUGE-L, Levenshtein, Jaro-Winkler) and JSON F1/validity metrics. The authors report strong performance on HTML-to-plaintext and HTML-to-JSON, plus analyses of repetition and output-quality distributions, and release model checkpoints and datasets.

Significance. If the evaluation were unbiased, the contribution would be practically valuable: a small specialized model for HTML-to-text extraction, public artifacts, a clear training recipe, and evidence that a contrastive objective reduces repetition. The paper is transparent about many limitations and releases reproducible resources. However, the central empirical claims rest on overlap with Fundus-derived references, and because the training targets are also Fundus outputs, the evaluation largely measures agreement with the authors' own rule system rather than extraction quality. The manuscript acknowledges this in Section 3.4, but the abstract and Section 4 do not carry the caveat, and several specific reporting inconsistencies (missing Trafilatura baseline, unsupported HTML-to-Markdown numbers, and Table 5 contradicting a 'every reported measure' claim) further undermine the headline conclusions. The core idea is promising, but the evaluation needs substantive correction before the claims can be accepted.

major comments (4)
  1. [§2.1, §3.1, §3.4] The evaluation is circular with respect to the reference standard. Training pairs D in Section 2.1 are produced by Fundus rules, and Section 3.1 states that test references are also 'derived from the Fundus extraction rules.' Section 3.4 admits a 'potential bias in favor of news-crawler-LM, as it is trained to reproduce these very decisions.' This means BLEU, METEOR, ROUGE-L, Levenshtein, and Jaro-Winkler scores measure how closely the model imitates Fundus, not extraction correctness. The claim in Section 4.1 that the model 'achieves the strongest performance across all metrics, outperforming both rule-based parsing libraries and other model-based baselines' is therefore only defensible as 'agrees most with Fundus.' A human-annotated sample or an external reference (e.g., manually cleaned article text) is needed to support the quality claim.
  2. [§3.4 vs Table 3] Section 3.4 lists Trafilatura as one of the three rule-based baselines, but Table 3 contains no Trafilatura row—only news-please and Boilerpipe appear. Because Trafilatura is a widely used and strong rule-based extractor, its absence leaves the claimed advantage over rule-based parsing libraries unverified. The paper should either add the Trafilatura results to Table 3 or explain why it is omitted.
  3. [Abstract vs §4] The abstract reports +4.8 BLEU and +6.1 METEOR for an HTML-to-Markdown task and +2.2 BLEU / +4.1 METEOR for HTML-to-JSON. However, no HTML-to-Markdown experiment appears anywhere in the paper; Tables 3 and 4 report HTML-to-plaintext and HTML-to-JSON only. The abstract's markdown-specific numbers are unsupported by the experimental section. Either add a markdown evaluation or remove these claims from the abstract.
  4. [§4.2 and Table 5] Section 4.2 states that news-crawler-LM achieves 'the best overall performance on overlap-based metrics' and 'outperforming all baselines on every reported measure.' Table 5 contradicts the latter: Qwen2.5-1.5B has higher F1 over all outputs (0.595 vs 0.523) and a higher valid-JSON rate (64.4% vs 53.3%), while F1 on valid JSON is essentially tied (0.786 vs 0.782). The sentence should be restricted to overlap-based metrics, and the trade-off between overlap and structural validity should be explicitly discussed rather than glossed over.
minor comments (4)
  1. [§5.1] The definition of 'low-repetition rate' is confusing: 'A higher low-repetition rate' is awkward, and it is unclear whether Table 6 covers plaintext, JSON, or both, while Figure 3 displays both tasks. Please define the metric precisely and state the task coverage.
  2. [§2.1 / Figure 1] The JSON schema in Figure 1 includes fields such as 'summary,' 'subheadlines,' and 'topics,' but Section 2.1 only mentions title, author, publication date, and body. Specify the exact field set and whether evaluation uses exact JSON matching or per-field fuzzy overlap.
  3. [§7 Conclusion] The sentence 'only less than 20% having a ROUGE-L scores<0.6' is imprecise and appears inconsistent with the distribution shown in Figure 4. Please state the exact threshold, task, and whether this refers to plaintext or JSON.
  4. [Limitations] Typos and grammar issues in the Limitations section ('may be bottleneck using classical models', 'mixture-of-experts ... of linear attention', 'ROUGE-L scores<0.6') should be corrected.

Circularity Check

1 steps flagged

Main 'prediction' is a measure of agreement with the training labels: evaluation references are Fundus outputs, the same rule system used to generate training data.

specific steps
  1. self definitional [Section 2.1 (Dataset); Section 3.4 (Baselines); Section 4.1 (HTML-to-Plaintext)]
    "In contrast, we sample approximately 100k HTML-to-plaintext and JSON pairs extracted using human-defined rules from the Fundus library ... A limitation of our evaluation is that the reference outputs are derived from the Fundus extraction rules, which introduces a potential bias in favor of news-crawler-LM, as it is trained to reproduce these very decisions."

    The training targets y_i are Fundus extraction outputs (Section 2.1), and the test references are also derived from the same Fundus extraction rules (Section 3.4). Therefore BLEU/ROUGE/METEOR scores measure how closely the model reproduces the exact rule system it was optimized on. Any model fine-tuned on Fundus outputs would rank higher on this benchmark than libraries that follow different extraction objectives. Section 4.1's claim that news-crawler-LM 'outperforms both rule-based parsing libraries' is thus a statement about distance to Fundus, not extraction quality. The paper's own admission that the model is 'trained to reproduce these very decisions' confirms the reduction; the headline omits this caveat. The fact that Fundus is the authors' own library (Dallabetta et al., 2024, with

full rationale

The central circularity is that the evaluation reference is produced by the same Fundus rule system used to generate the training data, and the paper itself concedes this in Section 3.4. This makes the headline 'outperforms rule-based parsing libraries' a measure of imitation of Fundus, not of extraction quality. A distillation benchmark could legitimately measure fidelity to Fundus, but the paper uses these scores to claim superiority in extraction quality, which is the circular step. There is some independent content: the held-out-publisher split and comparisons among fine-tuned models show generalization of the learned Fundus behavior across publishers, so the circularity is partial rather than total. Other issues noted by a skeptical reader—Trafilatura listed as a baseline but absent from Table 3, abstract HTML-to-Markdown gains (+4.8 BLEU/+6.1 METEOR) not matching any reported experiment, and Table 5 contradicting the 'outperforming all baselines on every reported measure' statement—are evidence or reporting problems, not circularity per se, and do not raise the score further. Overall score 6: one or more 'predictions' reduce by construction, while the model still has a competent distillation design.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 1 invented entities

The central claim rests on the axioms that Fundus outputs are the correct ground truth and that overlap metrics against those outputs measure quality. These are domain assumptions that are not validated against human judgment. The free parameters are standard training hyperparameters, not fitted physical constants, but several (rho, split sizes, t) are chosen without ablation.

free parameters (6)
  • LoRA rank = 16
    Chosen hyperparameter for parameter-efficient fine-tuning (Section 3.2). No ablation or tuning curve is provided.
  • LoRA scaling alpha = 16
    Chosen hyperparameter (Section 3.2).
  • Learning rate = 4e-5
    Chosen hyperparameter (Section 3.2).
  • Contrastive margin rho = 0.5
    Set following Su et al. (2022), not tuned here; affects the contrastive loss in Section 2.3.
  • Repetition threshold t = 0.1
    Arbitrary threshold for the repetition analysis in Section 5.1; not justified.
  • Publisher split sizes = 7 validation / 10 test
    Number of validation and test publishers chosen by hand; no rationale given for why 10 held-out publishers is sufficient.
axioms (5)
  • domain assumption Fundus extraction rules produce correct plaintext and JSON outputs for news articles
    Used as both training supervision (Section 2.1) and evaluation reference (Section 3.1). This is the load-bearing premise of the entire evaluation.
  • domain assumption Overlap metrics (BLEU, METEOR, ROUGE-L, Levenshtein, Jaro-Winkler) measured against Fundus references reflect extraction quality
    Assumed throughout Section 4 without validation against human judgments. The paper itself notes other parsers may differ in design philosophy (Section 3.4).
  • domain assumption The 10 held-out publishers represent the difficulty of generalizing to unseen publishers
    All 10 are within the 106-publisher Fundus universe; no diversity or difficulty analysis is provided.
  • domain assumption ReaderLM-v2 / Qwen2.5 pretrained backbones are suitable bases for HTML-to-plaintext and JSON generation
    Invoked in Section 2.2; the paper does not compare against other backbones.
  • domain assumption The training data does not contain HTML-to-JSON outputs that exceed the 8k output budget
    Stated in Section 2.1 without empirical verification besides a distribution figure.
invented entities (1)
  • news-crawler-LM no independent evidence
    purpose: A small fine-tuned model for HTML-to-plaintext and HTML-to-JSON extraction
    It is the paper's system, not a new physical or conceptual entity. No independent evidence beyond the paper's own evaluation.

pith-pipeline@v1.3.0-alltime-deepseek · 10762 in / 11390 out tokens · 115023 ms · 2026-08-01T07:53:47.652017+00:00 · methodology

0 comments
read the original abstract

Extracting structured content from news pages remains challenging due to heterogeneous HTML layouts, inconsistent markup, and substantial boilerplate such as navigation elements and advertisements. Rule-based news crawlers can achieve high extraction accuracy by encoding site-specific structure, but require manual configuration in order to generalize to new publishers. Large language models provide a more flexible alternative by reducing the need for handcrafted rules, but their high computational cost limits practical deployment. In this paper, we introduce news-crawler-LM, a small long-context language model fine-tuned on high-quality, human-validated extractions from the Fundus news-crawling library. Our model converts raw HTML into plaintext and structured JSON, including fields such as headline, author, publication date, and article body. In our experiments, news-crawler-LM outperforms strong baselines in HTML-to-Markdown and HTML-to-JSON extraction, improving performance by +4.8 BLEU and +6.1 METEOR in the HTML-to-Markdown task, and by +2.2 BLEU and +4.1 METEOR in the HTML-to-JSON task. However, we also observe that our model only slightly better compared to other rule-based parsing libraries on the HTML-to-plaintext task in evaluations on previously unseen publishers. We release all models and artifacts to the research community.

Figures

Figures reproduced from arXiv: 2607.21284 by Alan Akbik, Jonas Golde, Max Dallabetta, Pascal Stolzenburg.

Figure 1
Figure 1. Figure 1: HTML conversion task: Given the raw HTML of a crawled news article, the model is trained to extract the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distributions of subword token counts of the training data. We show HTML input counts (top), plaintext [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cumulative distribution of 5-gram repetition ratios for plaintext (left) and JSON (right) generation. Each [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Complementary cumulative distribution of ROUGE-L scores for plaintext (left) and JSON (right) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 4 canonical work pages

  1. [1]

    2025 , eprint=

    ReaderLM-v2: Small Language Model for HTML to Markdown and JSON , author=. 2025 , eprint=

    2025

  2. [2]

    2023 , eprint=

    Understanding HTML with Large Language Models , author=. 2023 , eprint=

    2023

  3. [3]

    HtmlRAG: HTML is Better Than Plain Text for Modeling Retrieved Knowledge in RAG Systems , url=

    Tan, Jiejun and Dou, Zhicheng and Wang, Wen and Wang, Mang and Chen, Weipeng and Wen, Ji-Rong , year=. HtmlRAG: HTML is Better Than Plain Text for Modeling Retrieved Knowledge in RAG Systems , url=. doi:10.1145/3696410.3714546 , booktitle=

    arXiv

  4. [4]

    Soviet physics

    Binary codes capable of correcting deletions, insertions, and reversals , author=. Soviet physics. Doklady , year=

  5. [5]

    2019 , eprint=

    Decoupled Weight Decay Regularization , author=. 2019 , eprint=

    2019

  6. [6]

    2017 , eprint=

    Adam: A Method for Stochastic Optimization , author=. 2017 , eprint=

    2017

  7. [7]

    2024 , eprint=

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author=. 2024 , eprint=

    2024

  8. [8]

    2022 , eprint=

    A Simple Contrastive Learning Objective for Alleviating Neural Text Degeneration , author=. 2022 , eprint=

    2022

  9. [9]

    2021 , eprint=

    LoRA: Low-Rank Adaptation of Large Language Models , author=. 2021 , eprint=

    2021

  10. [10]

    9th International Conference on Learning Representations, ICLR , year=

    8-bit Optimizers via Block-wise Quantization , author=. 9th International Conference on Learning Representations, ICLR , year=

  11. [11]

    A uto C lean: LLM s Can Prepare Their Training Corpus

    Shen, Xingyu and Hu, Shengding and Zhang, Xinrong and Han, Xu and Meng, Xiaojun and Wei, Jiansheng and Liu, Zhiyuan and Sun, Maosong. A uto C lean: LLM s Can Prepare Their Training Corpus. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (System Demonstr...

    work page doi:10.18653/v1/2025.naacl-demo.9 2025

  12. [12]

    Understanding HTML with Large Language Models

    Gur, Izzeddin and Nachum, Ofir and Miao, Yingjie and Safdari, Mustafa and Huang, Austin and Chowdhery, Aakanksha and Narang, Sharan and Fiedel, Noah and Faust, Aleksandra. Understanding HTML with Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023. doi:10.18653/v1/2023.findings-emnlp.185

    doi:10.18653/v1/2023.findings-emnlp.185 2023

  13. [13]

    Trafilatura: A Web Scraping Library and Command-Line Tool for Text Discovery and Extraction

    Barbaresi, Adrien. Trafilatura: A Web Scraping Library and Command-Line Tool for Text Discovery and Extraction. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing: System Demonstrations. 2021. doi:10.18653/v1/2021.acl-demo.15

    doi:10.18653/v1/2021.acl-demo.15 2021

  14. [14]

    2024 , eprint=

    The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale , author=. 2024 , eprint=

    2024

  15. [15]

    Automated Fact-Checking of Claims from W ikipedia

    Sathe, Aalok and Ather, Salar and Le, Tuan Manh and Perry, Nathan and Park, Joonsuk. Automated Fact-Checking of Claims from W ikipedia. Proceedings of the Twelfth Language Resources and Evaluation Conference. 2020

    2020

  16. [16]

    Generating Fact Checking Summaries for Web Claims

    Mishra, Rahul and Gupta, Dhruv and Leippold, Markus. Generating Fact Checking Summaries for Web Claims. Proceedings of the Sixth Workshop on Noisy User-generated Text (W-NUT 2020). 2020. doi:10.18653/v1/2020.wnut-1.12

    doi:10.18653/v1/2020.wnut-1.12 2020

  17. [17]

    W eb IE : Faithful and Robust Information Extraction on the Web

    Whitehouse, Chenxi and Vania, Clara and Aji, Alham Fikri and Christodoulopoulos, Christos and Pierleoni, Andrea. W eb IE : Faithful and Robust Information Extraction on the Web. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. doi:10.18653/v1/2023.acl-long.428

    work page doi:10.18653/v1/2023.acl-long.428 2023

  18. [18]

    2020 , eprint=

    The Curious Case of Neural Text Degeneration , author=. 2020 , eprint=

    2020

  19. [19]

    T ext R unner: Open Information Extraction on the Web

    Yates, Alexander and Banko, Michele and Broadhead, Matthew and Cafarella, Michael and Etzioni, Oren and Soderland, Stephen. T ext R unner: Open Information Extraction on the Web. Proceedings of Human Language Technologies: The Annual Conference of the North A merican Chapter of the Association for Computational Linguistics ( NAACL - HLT ). 2007

    2007

  20. [20]

    Transformers: State-of-the-Art Natural Language Processing

    Wolf, Thomas and Debut, Lysandre and Sanh, Victor and Chaumond, Julien and Delangue, Clement and Moi, Anthony and Cistac, Pierric and Rault, Tim and Louf, Remi and Funtowicz, Morgan and Davison, Joe and Shleifer, Sam and von Platen, Patrick and Ma, Clara and Jernite, Yacine and Plu, Julien and Xu, Canwen and Le Scao, Teven and Gugger, Sylvain and Drame, M...

    2020

  21. [21]

    2020 , eprint=

    Hate is the New Infodemic: A Topic-aware Modeling of Hate Speech Diffusion on Twitter , author=. 2020 , eprint=

    2020

  22. [22]

    2020 , eprint=

    Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention , author=. 2020 , eprint=

    2020

  23. [23]

    A Survey on Mixture of Experts in Large Language Models , ISSN=

    Cai, Weilin and Jiang, Juyong and Wang, Fan and Tang, Jing and Kim, Sunghun and Huang, Jiayi , year=. A Survey on Mixture of Experts in Large Language Models , ISSN=. doi:10.1109/tkde.2025.3554028 , journal=

    arXiv 2025

  24. [24]

    , author=

    String comparator metrics and enhanced decision rules in the fellegi-sunter model of record linkage. , author=. 1990 , publisher=

    1990

  25. [25]

    , title =

    Damerau, Fred J. , title =. Commun. ACM , month = mar, pages =. 1964 , issue_date =. doi:10.1145/363958.363994 , abstract =

    arXiv 1964

  26. [26]

    ROUGE : A Package for Automatic Evaluation of Summaries

    Lin, Chin-Yew. ROUGE : A Package for Automatic Evaluation of Summaries. Text Summarization Branches Out. 2004

    2004

  27. [27]

    METEOR : An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments

    Banerjee, Satanjeev and Lavie, Alon. METEOR : An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization. 2005

    2005

  28. [28]

    B leu: a Method for Automatic Evaluation of Machine Translation

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing. B leu: a Method for Automatic Evaluation of Machine Translation. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics. 2002. doi:10.3115/1073083.1073135

    arXiv 2002

  29. [29]

    2022 , eprint=

    A Contrastive Framework for Neural Text Generation , author=. 2022 , eprint=

    2022

  30. [30]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  31. [31]

    2025 , eprint=

    NEXT-EVAL: Next Evaluation of Traditional and LLM Web Data Record Extraction , author=. 2025 , eprint=

    2025

  32. [32]

    FreeDOM: A Transferable Neural Architecture for Structured Information Extraction on Web Documents , url=

    Lin, Bill Yuchen and Sheng, Ying and Vo, Nguyen and Tata, Sandeep , year=. FreeDOM: A Transferable Neural Architecture for Structured Information Extraction on Web Documents , url=. doi:10.1145/3394486.3403153 , booktitle=

    arXiv

  33. [33]

    2022 , eprint=

    DOM-LM: Learning Generalizable Representations for HTML Documents , author=. 2022 , eprint=

    2022

  34. [34]

    2022 , eprint=

    Don't read, just look: Main content extraction from web pages using visual features , author=. 2022 , eprint=

    2022

  35. [35]

    Fundus: A Simple-to-Use News Scraper Optimized for High Quality Extractions

    Dallabetta, Max and Dobberstein, Conrad and Breiding, Adrian and Akbik, Alan. Fundus: A Simple-to-Use News Scraper Optimized for High Quality Extractions. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). 2024. doi:10.18653/v1/2024.acl-demos.29

    work page doi:10.18653/v1/2024.acl-demos.29 2024

  36. [36]

    and Grimm, Peter and Chiang, Michael F

    Gupta, Suhit and Kaiser, Gail E. and Grimm, Peter and Chiang, Michael F. and Starren, Justin , title =. World Wide Web , month = jun, pages =. 2005 , issue_date =. doi:10.1007/s11280-004-4873-3 , abstract =

    work page doi:10.1007/s11280-004-4873-3 2005

  37. [37]

    DELOS Workshops / Conferences , year=

    Fact or Fiction: Content Classification for Digital Libraries , author=. DELOS Workshops / Conferences , year=

  38. [38]

    Boilerplate detection using shallow text features , year =

    Kohlsch\". Boilerplate detection using shallow text features , year =. Proceedings of the Third ACM International Conference on Web Search and Data Mining , pages =. doi:10.1145/1718487.1718542 , abstract =

    arXiv

  39. [39]

    Building a 70 billion word corpus of E nglish from C lue W eb

    Pomik \'a lek, Jan and Jakub \'i c ek, Milo s and Rychl \'y , Pavel. Building a 70 billion word corpus of E nglish from C lue W eb. Proceedings of the Eighth International Conference on Language Resources and Evaluation ( LREC '12). 2012

    2012

  40. [40]

    2017 , booktitle =

    Hamborg, Felix and Meuschke, Norman and Breitinger, Corinna and Gipp, Bela , title =. 2017 , booktitle =. doi:10.5281/zenodo.4120316 , pages =

    doi:10.5281/zenodo.4120316 2017

  41. [41]

    2020 , isbn =

    Leonhardt, Jurek and Anand, Avishek and Khosla, Megha , title =. 2020 , isbn =. doi:10.1145/3366424.3383547 , booktitle =

    arXiv 2020

  42. [42]

    2024 , eprint=

    Cleaner Pretraining Corpus Curation with Neural Web Scraping , author=. 2024 , eprint=

    2024

  43. [43]

    2025 , eprint=

    WebLists: Extracting Structured Information From Complex Interactive Websites Using Executable LLM Agents , author=. 2025 , eprint=

    2025