Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Automated Extraction of Acronym-Expansion Pairs from Scientific Papers

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A regex parser plus GPT-4 beats either alone at expanding acronyms in scientific PDFs.

desk verdict A sensible hybrid regex+LLM pipeline whose main claim is undermined by an evaluation that counts unvalidated GPT-4 outputs as correct expansions. read the letter →

arxiv 2412.01093 v1 pith:F7RZAJ4Y submitted 2024-12-02 cs.CL cs.IR

Izhar Ali , Million Haileyesus , Serhiy Hnatyshyn , Jan-Lucas Ott , Vasil Hnatyshin This is my paper · ORCID
classification cs.CLcs.IR
keywords acronymextractionexpansionregularexpressionsGPT-4largelanguagemodelsPDFpreprocessingscientifictextminingnaturalprocessing
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 is trying to establish that acronym-expansion extraction from scientific PDFs is best done by a three-stage pipeline: clean the PDF text aggressively, use a regular-expression parser to identify every acronym-shaped token, and let GPT-4 expand each token using only a two-sentence window of surrounding text. The authors report that across 200 arXiv papers from biochemistry, systems biology, computational linguistics, and numerical analysis, this combined pipeline found expansions for 72.2% to 84.9% of parser-identified acronyms, while the parser alone found only 14.8% to 20.9%. The paper also contends that neither tool alone is adequate: the parser is a reliable detector but a poor expander, and GPT-4 is a fluent expander but an unreliable detector, routinely missing acronyms the parser catches. A sympathetic reader would care because automated acronym expansion improves machine comprehension and readability of scientific text, and the paper proposes a practical division of labor between deterministic pattern matching and a probabilistic language model.

What carries the argument

The load-bearing mechanism is a two-tier division of labor with limited context. First, a regular-expression parser uses a word-boundary pattern that matches capitalized letter sequences (optionally hyphenated, with an optional trailing 's') to enumerate acronym candidates, and then two hand-built patterns extract expansions written either as 'acronym (expansion)' or 'expansion (acronym)', with stopword filtering and length-based refinement to trim false or overlong expansions. Second, for each acronym the parser could not expand, the algorithm sends GPT-4 the acronym together with at most two sentences — the sentence containing it and the preceding sentence — and instructs the model to refine, infer, or discard the expansion, returning JSON. The design choice that carries the argument is limiting GPT-4's context to two sentences: it keeps prompts small enough to avoid the model's token limits and 'lost-in-the-middle' failures, while giving the model enough local evidence to infer non-local expansions.

What would settle it

Take 20 papers from the same four arXiv domains, have a domain expert hand-label every acronym and its correct expansion, then run RegEx, GPT+Pre, and GPT+RegEx+Pre against that gold standard; if most of the expansions GPT-4 supplies are plausible but wrong, or if many acronyms the parser misses are common in the text, the combined pipeline's advantage over a simpler approach largely disappears.

Watch

Extended reading notes

Core claim

The central claim is that acronym extraction should be split by competence: regular expressions carry recall, GPT-4 carries expansion quality, and preprocessing removes the noise that would otherwise be mistaken for acronyms. On a corpus of 50 papers from each of four arXiv domains, the combined GPT+RegEx+Pre approach expanded 84.9% of parser-identified acronyms in biochemistry, 80.9% in systems biology, 80.5% in computational linguistics, and 72.2% in numerical analysis, compared with 14.8% to 20.9% for the regex parser alone. The paper further reports that GPT-4 on its own returns an expansion for nearly everything it is given, but misses many genuine acronyms that the regex parser finds, such as NTP, Cryo-EM, IDT, and TEV. The authors conclude that as of GPT-4o mini, the most accurate and consistent results come from preprocessing plus parser plus GPT-4.

Load-bearing premise

The reported success rates assume the regex parser's acronym list is the correct denominator and that every GPT-4 output counted in Table II is a correct expansion, but neither assumption was checked against a manually built gold standard.

Editorial extensions

If this is right

  • A hybrid detector-plus-LLM pipeline can raise acronym expansion coverage from roughly 15-21% with regex alone to 72-85% of parser-identified acronyms, so practical tools for expanding acronyms in PDFs are feasible today.
  • Because the parser identifies far more acronyms than GPT-4 alone, any LLM-based acronym extractor should run behind a deterministic recall stage rather than asking the model to scan whole documents.
  • Aggressive PDF preprocessing that strips headers, footers, references, equations, roman numerals, and gene-like strings reduces misidentified acronyms by up to 11%, so text cleanup directly improves downstream expansion accuracy.
  • GPT-4 can expand acronyms whose definitions are missing or non-local in the text, which is exactly the failure mode that defeats parenthetical-pattern regexes.
  • Improving preprocessing further should raise the combined pipeline's percentages without changing the model, since GPT-4 currently wastes effort trying to expand misidentified acronyms.

Reading between the lines

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

  • The reported 72-85% figures are computed against the parser's own acronym count, so they are upper-bounded by regex recall; extending the parser to lowercase or digit-leading initialisms (for example 2FA, eLisp) would likely raise the absolute number of acronyms expanded even if the percentage stays similar.
  • The paper's 'lost-in-the-middle' observation suggests a testable extension: shuffle or randomly order the 15-20 acronyms sent in each GPT-4 prompt and check whether acronyms in the middle of the prompt are expanded less consistently than those at the edges.
  • A natural validation experiment the paper does not run is a hand-built gold standard of acronym-expansion pairs on a sample of the 200 papers; with one, the same pipeline could be measured for true precision and recall rather than coverage of regex-identified tokens.
  • The same division of labor — a cheap deterministic detector followed by an LLM resolver over a small context window — likely transfers to other scientific-text extraction tasks such as gene names, chemical entities, or figure references, wherever pattern recall is easy and semantic resolution is hard.
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 / 5 minor

Summary. The paper proposes a hybrid pipeline for extracting acronym-expansion pairs from scientific PDFs, combining document preprocessing, a regular-expression-based parser, and GPT-4 (specifically GPT-4o mini) prompts. The parser identifies acronyms and extracts expansions that follow parenthetical patterns; GPT-4 is then used to verify, correct, or infer expansions for acronyms whose expansions were not found. The method is evaluated on 200 arXiv papers from four domains. The central claim, stated in Sections III-C and VII, is that the integrated approach (GPT+RegEx+Pre) outperforms both the regex parser alone and GPT-4 alone, with Table II reporting expansion-finding rates of 72.2% to 84.9% for the hybrid versus roughly 15% to 21% for the regex parser alone.

Significance. If the performance claims were supported by a rigorous evaluation, the paper would offer a practical and potentially useful engineering contribution: combining regex-based identification with an LLM for expansion inference in a token-efficient, domain-agnostic way. The paper also documents concrete preprocessing challenges for PDF-derived text. However, the current evidence does not substantiate the headline claim. The evaluation in Table II counts GPT-4 outputs as correct without any gold-standard labels or precision measurement, and the denominator of acronyms is defined by the authors' own regex parser rather than by an independently validated set. The paper explicitly acknowledges in Section VI that no parser accuracy statistics were collected. As a result, the claimed advantage of the hybrid method over the regex parser alone is not established. The work is reproducible only in a loose sense: the regex patterns and prompt are given, but the dataset, code, and raw outputs are not provided.

major comments (4)
  1. [Section VI, Table II] The numerator 'Expansions Found' counts any non-empty GPT-4 output as a correct expansion without validation. The prompt in Listing 4 explicitly instructs GPT-4 to 'use your best judgment to provide the most accurate and concise full form' when the full form cannot be determined from context. This means hallucinated or plausible-but-wrong expansions are counted as successes. The paper reports only three cases where GPT-4 failed to produce any expansion and no cases where an produced expansion was wrong. Without a precision measurement on a sample of GPT-4 outputs, the reported 72.2% to 84.9% success rates for the hybrid method are not interpretable as accuracy figures.
  2. [Section VI, Table II] The denominator 'Total Acronyms' for the RegEx and hybrid rows is the set of strings flagged by the authors' own regex parser after preprocessing (Listing 1), not a gold-standard set of acronyms in the papers. The paper acknowledges in Section VI that it was 'unable to collect statistics regarding the parser's accuracy' and that the manual review of 40 papers only confirmed that the parser finds all acronyms matching its pattern. The parser can admit false positives (e.g., chemical names, headings) and false negatives (e.g., acronyms starting with lowercase letters or numbers, as noted in Section VII). Consequently, all percentages in Table II are relative to an unvalidated denominator, and the comparison between the hybrid method and GPT-alone rows is not apples-to-apples because GPT-alone uses a different denominator (GPT's own detections).
  3. [Section VII and Section III-C] The conclusion that 'the combination of our preprocessing steps and parser, alongside GPT-4, yields the best results' outruns the evidence. There are no significance tests, no confidence intervals, and no error analysis linking the reported percentages to actual correctness. Because both the numerator and denominator in Table II are unvalidated, the large gap between the hybrid method (80.5% to 84.9%) and the regex-only method (14.8% to 21.2%) could shrink substantially or disappear if GPT-4's expansions are often wrong. At minimum, the authors need to annotate a stratified random sample of GPT-4 outputs and of parser-identified acronyms against a gold standard, and report precision, recall, and inter-annotator agreement.
  4. [Section VI, preprocessing discussion] The paper reports that preprocessing reduces the number of GPT-identified acronyms (e.g., from 823 to 729 in BC, from 865 to 622 in SB) but does not analyze whether preprocessing removes genuine acronyms along with noise. Since preprocessing is a core component of the proposed method, false negatives introduced by steps such as heading removal, stopword filtering, or the exclusion of acronyms longer than 10 characters should be quantified separately. Otherwise, the reported improvement of GPT+RegEx+Pre over GPT+Pre could be partly due to preprocessing removing hard cases that GPT would have handled correctly.
minor comments (5)
  1. [Section VI-A] In the sentence 'presenting an estimate of the average number of acronyms per paper is each domain', the word 'is' should be 'in'.
  2. [Table II] The column header 'Total *' is unclear; the asterisk appears to refer to the note below the table, but the note does not explain the multiplication symbol or the exact meaning of the column.
  3. [Section IV-B] The explanation of code snippet (1) states that 's?' optionally matches an 's' at the end, but the pattern is written as `[A-Z]s?\b`; the literal 's' is not in a character class, which is correct, but the notation in the explanation could be clearer for readers unfamiliar with regex syntax.
  4. [Section V] The paper alternates between saying 'GPT-4' and 'GPT-4o mini' and states they are used interchangeably; since GPT-4o mini is a distinct model with different behavior, the paper should consistently report the exact model version and date for each experiment.
  5. [Figure 1] Figure 1 is referenced in Section III-C but is not included in the manuscript text provided; please ensure the figure is present in the final version and that it clearly shows the data flow through preprocessing, regex parsing, and GPT-4 refinement.

Circularity Check

2 steps flagged · score 6.0 of 10

Table II's success metric counts unvalidated GPT-4 outputs over a regex-defined acronym set, so the claimed superiority of GPT+RegEx+Pre is partially built into the evaluation definitions.

  1. self definitional [Section VI, Table II (note) and Section IV-C, Listing 4]
    "GPT excels at coming up with expansions, as it is designed to always provide an answer. ... If the full form cannot be determined from the context, use your best judgment to provide the most accurate and concise full form."

    The metric 'Expansions Found' in Table II is the count of GPT-4 outputs; no gold-standard verification of the 3100, 2098, 1239, or 909 counted expansions is reported. Because the prompt instructs GPT-4 to provide a best-judgment expansion when uncertain, and GPT-4 is 'designed to always provide an answer,' the high percentages for GPT+RegEx+Pre are guaranteed by the prompt and output convention rather than by extraction accuracy. The conclusion that the integrated approach 'yielded the best performance' therefore reduces to the observation that GPT-4 returns strings for most parser-identified acronyms.

  2. self definitional [Section VI, parser-accuracy validation paragraph]
    "By manually examining 40 randomly selected papers, we were able to confirm that the parser does identify all acronyms in the paper that follow the regular expression pattern defined in code snippet (1)."

    The denominator 'Total Acronyms' for RegEx, RegEx+Pre, and GPT+RegEx+Pre is the set of strings matched by the authors' own regular expression. The manual validation checks recall only against that same pattern, so the parser's completeness is true by definition. Consequently, the headline figures (e.g., 84.9% of 'total acronyms' expanded) are relative to a self-selected set, not to an independently annotated acronym inventory; the system's denominator and its success rate share the same regex origin. The claim of best performance is therefore not measured against any external ground truth.

full rationale

No self-citation is load-bearing; the paper's references are external. The circularity lies in the evaluation design. The central claim that GPT+RegEx+Pre is best rests on Table II, where 'Expansions Found' counts GPT-4's produced strings without an external correctness check, and 'Total Acronyms' is the authors' own regex parser output, validated only against its own pattern. Under these definitions, the high expansion rates for the combined method are partly by construction: GPT-4 is prompted to return a best-judgment answer for every parser-identified acronym, so the percentage of non-empty outputs is high regardless of whether the expansions are correct. The comparison of GPT alone at 100% of its self-selected set versus RegEx at about 20% of a different set further obscures the result. Still, the paper contains substantial non-circular engineering content: preprocessing rules, regex design, and chunking are concrete artifacts, and the rank-ordering of RegEx versus RegEx+Pre is not purely definitional. The circularity is partial, affecting the headline superiority claim, not the entire method.

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

The central claim rests on several hand-chosen settings and domain assumptions rather than on a derivation. The most consequential free parameters are the regex pattern itself, the context window (current plus preceding sentence), the 10-character acronym cutoff, and the stopword-filter threshold; none are derived from theory or tuned against a public benchmark. The evaluation also assumes that GPT-4's output is a reliable oracle for expansions.

free parameters (4)
  • Acronym length cutoff = 10 characters
    Hand-chosen to avoid gene sequences and other non-acronym strings; Section IV-A.
  • Context window for GPT-4 = Current sentence plus preceding sentence
    Selected through trial and error; the authors found the preceding sentence more useful than the following sentence; Section III-C.
  • Chunk size for GPT-4 calls = 15-20 acronyms per call
    Chosen to balance minimizing API calls against input limits and quality; Section V.
  • Stopword proportion threshold = Preset threshold, not specified
    Used to discard parenthetical expansions with a high proportion of stopwords, such as 'which is usually used in practice'; Section IV-B.
assumptions (4)
  • domain assumption GPT-4 returns correct expansions when given a two-sentence context
    The method relies on the model's ability to infer expansions from context; Section III-C and the prompt in Listing 4 assume this.
  • ad hoc to paper The regex pattern in Listing 1 captures the intended notion of an acronym
    The definition of 'total acronyms' and thus all performance percentages depend on this hand-written pattern; Section IV-B.
  • ad hoc to paper Preprocessing removes non-acronym text without removing genuine acronyms
    Filters for Roman numerals, gene sequences, headings, etc. are applied to all documents; Section IV-A. The paper notes these may still occasionally misclassify.
  • ad hoc to paper The metric 'expansions found' without precision measurement is a valid performance indicator
    Table II counts any expansion returned by GPT-4 as success; the paper does not measure false expansions. Section VI.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Automated Extraction of Acronym-Expansion Pairs from Scientific Papers." pith.science (2026). https://pith.science/paper/F7RZAJ4Y

@misc{pith2026241201093,
  author       = {Pith},
  title        = {Pith review of: Automated Extraction of Acronym-Expansion Pairs from Scientific Papers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7RZAJ4Y}},
  note         = {Machine review of arXiv:2412.01093}
}
read the original abstract

This project addresses challenges posed by the widespread use of abbreviations and acronyms in digital texts. We propose a novel method that combines document preprocessing, regular expressions, and a large language model to identify abbreviations and map them to their corresponding expansions. The regular expressions alone are often insufficient to extract expansions, at which point our approach leverages GPT-4 to analyze the text surrounding the acronyms. By limiting the analysis to only a small portion of the surrounding text, we mitigate the risk of obtaining incorrect or multiple expansions for an acronym. There are several known challenges in processing text with acronyms, including polysemous acronyms, non-local and ambiguous acronyms. Our approach enhances the precision and efficiency of NLP techniques by addressing these issues with automated acronym identification and disambiguation. This study highlights the challenges of working with PDF files and the importance of document preprocessing. Furthermore, the results of this work show that neither regular expressions nor GPT-4 alone can perform well. Regular expressions are suitable for identifying acronyms but have limitations in finding their expansions within the paper due to a variety of formats used for expressing acronym-expansion pairs and the tendency of authors to omit expansions within the text. GPT-4, on the other hand, is an excellent tool for obtaining expansions but struggles with correctly identifying all relevant acronyms. Additionally, GPT-4 poses challenges due to its probabilistic nature, which may lead to slightly different results for the same input. Our algorithm employs preprocessing to eliminate irrelevant information from the text, regular expressions for identifying acronyms, and a large language model to help find acronym expansions to provide the most accurate and consistent results.

Figures

Figures reproduced from arXiv: 2412.01093 by the authors.

Figure 1
Figure 1. Approach for extracting acronym-expansion pairs. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages

  1. [1]

    A Review of Text Corpus-Based Tourism Big Data Mining,

    Q. Li, S. Li, S. Zhang, J. Hu, and J. Hu, “A Review of Text Corpus-Based Tourism Big Data Mining,” Applied Sciences , vol. 9, no. 16, p. 3300, Aug. 2019. [Online]. Available: https: //www.mdpi.com/2076-3417/9/16/3300

    work page 2019

  2. [2]

    Commentaries on Big Data’s Importance for Accounting and Auditing,

    P. A. Griffin and A. M. Wright, “Commentaries on Big Data’s Importance for Accounting and Auditing,” Accounting Horizons, vol. 29, no. 2, pp. 377–379, Jun. 2015. [Online]. Available: https://publications.aaahq.org/accounting-horizons/article/29/ 2/377/2182/Commentaries-on-Big-Data-s-Importance-for

    work page 2015

  3. [3]

    The growth of acronyms in the scientific literature,

    A. Barnett and Z. Doubleday, “The growth of acronyms in the scientific literature,” eLife, vol. 9, p. e60080, Jul. 2020. [Online]. Available: https://elifesciences.org/articles/60080

    work page 2020

  4. [4]

    Abbreviations, initialism and acronyms: their use in medical physics (THUMP),

    M. Caon, “Abbreviations, initialism and acronyms: their use in medical physics (THUMP),” Australasian Physical & Engineering Sciences in Medicine, vol. 39, no. 1, pp. 11–12, Mar. 2016. [Online]. Available: http://link.springer.com/10.1007/s13246-016-0423-4

    work page doi:10.1007/s13246-016-0423-4 2016

  5. [5]

    So sick or so cool? The language of youth on the internet,

    S. A. Tagliamonte and In collaboration with Dylan Uscher, Lawrence Kwok, and students from HUM199Y 2009 and 2010, “So sick or so cool? The language of youth on the internet,” Language in Society , vol. 45, no. 1, pp. 1–32, Feb. 2016. [Online]. Available: https://www.cambridge.org/core/product/identifier/ S0047404515000780/type/journal article

    work page 2009

  6. [6]

    Using compression to identify acronyms in text,

    S. A. Yeates, D. Bainbridge, and I. H. Witten, “Using compression to identify acronyms in text,” CoRR, vol. cs.DL/0007003, 2000. [Online]. Available: https://arxiv.org/abs/cs/0007003

    work page arXiv 2000

  7. [7]

    Creating an Online Dictionary of Abbreviations from MEDLINE,

    J. T. Chang, “Creating an Online Dictionary of Abbreviations from MEDLINE,” Journal of the American Medical Informatics Association , vol. 9, no. 6, pp. 612–620, Nov. 2002. [Online]. Available: https: //academic.oup.com/jamia/article-lookup/doi/10.1197/jamia.M1139

    work page doi:10.1197/jamia.m1139 2002

  8. [8]

    Acronyms: identification, expansion and disambiguation,

    K. Jacobs, A. Itai, and S. Wintner, “Acronyms: identification, expansion and disambiguation,” Annals of Mathematics and Artificial Intelligence , vol. 88, no. 5-6, pp. 517–532, Jun. 2020. [Online]. Available: http://link.springer.com/10.1007/s10472-018-9608-8

    work page doi:10.1007/s10472-018-9608-8 2020

Show all 40 references
  1. [9]

    Available: https://openai.com/research/gpt-4

    “GPT-4.” [Online]. Available: https://openai.com/research/gpt-4

  2. [10]

    Richard Nordquist, Ph.D

    “Richard Nordquist, Ph.D.” [Online]. Available: https://www.npr.org/sections/memmos/2015/01/06/605393666/ do-you-suffer-from-ras-syndrome

    2015

  3. [11]

    Automatic extraction of acronym-meaning pairs from medline databases,

    J. Pustejovsky, J. M. Casta ˜no, B. Cochran, M. Kotecki, and M. Morrell, “Automatic extraction of acronym-meaning pairs from medline databases,” Studies in health technology and informatics , vol. 84 Pt 1, pp. 371–5, 2001. [Online]. Available: https://api. semanticscholar.org/...

    2001

  4. [12]

    A simple algorithm for identifying abbreviation definitions in biomedical text,

    A. S. Schwartz and M. A. Hearst, “A simple algorithm for identifying abbreviation definitions in biomedical text,” Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing , pp. 451–62, 2002. [Online]. Available: https://api.semanticscholar.org/CorpusID:28503121

    2002

  5. [13]

    Learning domain ontologies from document warehouses and dedicated web sites,

    R. Navigli and P. Velardi, “Learning domain ontologies from document warehouses and dedicated web sites,” vol. 30, no. 2, p. 151–179, jun

  6. [14]

    A Supervised Learning Approach to Acronym Identification,

    D. Nadeau and P. D. Turney, “A Supervised Learning Approach to Acronym Identification,” in Advances in Artificial Intelligence , D. Hutchison, T. Kanade, J. Kittler, J. M. Kleinberg, F. Mattern, J. C. Mitchell, M. Naor, O. Nierstrasz, C. Pandu Rangan, B. Steffen, M. Sudan, D. ...

    2005 doi

  7. [15]

    BERT-based Acronym Disambiguation with Multiple Training Strategies,

    C. Pan, B. Song, S. Wang, and Z. Luo, “BERT-based Acronym Disambiguation with Multiple Training Strategies,” Mar. 2021, arXiv:2103.00488 [cs]. [Online]. Available: http://arxiv.org/abs/2103. 00488

    2021 arXiv

  8. [16]

    A convolutional route to abbreviation disambiguation in clinical text,

    V . Joopudi, B. Dandala, and M. Devarakonda, “A convolutional route to abbreviation disambiguation in clinical text,” Journal of Biomedical Informatics , vol. 86, pp. 71–78, 2018. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1532046418301552

    2018

  9. [17]

    Pronunciation-based acronym recognition,

    J. Cao, E. Shareghi, and N. Collier, “Pronunciation-based acronym recognition,” in Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) . New Orleans, Louisiana: A...

    2018

  10. [18]

    Improving Structured Text Recognition with Regular Expression Biasing,

    B. Shi, W. Cheng, Y . Lu, C. Zhang, and D. Florencio, “Improving Structured Text Recognition with Regular Expression Biasing,” Nov. 2021, arXiv:2111.06738 [cs]. [Online]. Available: http://arxiv.org/abs/ 2111.06738

    2021 arXiv

  11. [19]

    Unsupervised Abbreviation Disambiguation Contextual disambiguation using word embeddings,

    M. Ciosici, T. Sommer, and I. Assent, “Unsupervised Abbreviation Disambiguation Contextual disambiguation using word embeddings,” May 2019, arXiv:1904.00929 [cs]. [Online]. Available: http://arxiv.org/ abs/1904.00929

    2019 arXiv

  12. [20]

    Transformer Language Models without Positional Encodings Still Learn Positional Information,

    A. Haviv, O. Ram, O. Press, P. Izsak, and O. Levy, “Transformer Language Models without Positional Encodings Still Learn Positional Information,” in Findings of the Association for Computational Linguistics: EMNLP 2022 . Abu Dhabi, United Arab Emirates: Association for Computa...

    2022

  13. [21]

    Jurafsky and J

    D. Jurafsky and J. H. Martin, Speech and language processing: an introduction to natural language processing, computational linguistics, and speech recognition, ser. Prentice Hall series in artificial intelligence. Upper Saddle River, N.J: Prentice Hall, 2000

    2000

  14. [22]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schulman, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe, “Training language models to follow instructions...

    2022 arXiv

  15. [23]

    Large language models in medicine,

    A. J. Thirunavukarasu, D. S. J. Ting, K. Elangovan, L. Gutierrez, T. F. Tan, and D. S. W. Ting, “Large language models in medicine,” Nature Medicine, vol. 29, no. 8, pp. 1930–1940, Aug. 2023. [Online]. Available: https://www.nature.com/articles/s41591-023-02448-8

    1930

  16. [24]

    Comparative analysis of large language models in the Royal College of Ophthalmologists fellowship exams,

    R. Raimondi, N. Tzoumas, T. Salisbury, S. Di Simplicio, M. R. Romano, North East Trainee Research in Ophthalmology Network (NETRiON), T. Bommireddy, H. Chawla, Y . Chen, S. Connolly, S. El Omda, M. Gough, L. Kishikova, T. McNally, S. N. Sadiq, S. Simpson, B. L. Teh, S. Toh, V ...

    2023

  17. [25]

    Predictability and Surprise in Large Generative Models,

    D. Ganguli, D. Hernandez, L. Lovitt, N. DasSarma, T. Henighan, A. Jones, N. Joseph, J. Kernion, B. Mann, A. Askell, Y . Bai, A. Chen, T. Conerly, D. Drain, N. Elhage, S. E. Showk, S. Fort, Z. Hatfield-Dodds, S. Johnston, S. Kravec, N. Nanda, K. Ndousse, C. Olsson, D. Amodei, D...

    2022 arXiv

  18. [26]

    Apache Tika – Apache Tika

    “Apache Tika – Apache Tika.” [Online]. Available: https://tika.apache. org/

  19. [27]

    Available: https://pypdf.readthedocs.io/en/stable/ index.html

    “Pypdf.” [Online]. Available: https://pypdf.readthedocs.io/en/stable/ index.html

  20. [28]

    Pypdf benchmarking

    “Pypdf benchmarking.” [Online]. Available: https://github.com/py-pdf/ benchmarks

  21. [29]

    Text Preprocessing and its Implications in a Digital Humanities Project

    M. Kunilovskaya and A. Plum, “Text Preprocessing and its Implications in a Digital Humanities Project.” Online: INCOMA Ltd., Sep. 2021, pp. 85–93. [Online]. Available: https://aclanthology.org/2021.ranlp-srw.13

    2021

  22. [30]

    C. D. Manning, P. Raghavan, and H. Sch ¨utze, Introduction to informa- tion retrieval. New York: Cambridge University Press, 2008, oCLC: ocn190786122

    2008

  23. [31]

    S. Bird, E. Klein, and E. Loper, Natural Language Processing with Python, 01 2009

    2009

  24. [32]

    A scoping review of preprocessing methods for unstructured text data to assess data quality,

    M. Nesca, A. Katz, C. K.-S. Leung, and L. M. Lix, “A scoping review of preprocessing methods for unstructured text data to assess data quality,” International Journal of Population Data Science, vol. 7, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:252739682

    2022

  25. [33]

    GPT-4o-mini Technical Report,

    OpenAI, “GPT-4o-mini Technical Report,” Jul

  26. [34]

    Generating Wikipedia by Summarizing Long Sequences,

    P. J. Liu, M. Saleh, E. Pot, B. Goodrich, R. Sepassi, L. Kaiser, and N. Shazeer, “Generating Wikipedia by Summarizing Long Sequences,” Jan. 2018, arXiv:1801.10198 [cs]. [Online]. Available: http://arxiv.org/abs/1801.10198

    2018 arXiv

  27. [35]

    Chatgpt: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope,

    P. P. Ray, “Chatgpt: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope,” Internet of Things and Cyber-Physical Systems , vol. 3, pp. 121– 154, 2023. [Online]. Available: https://www.sciencedirect.com/science/ article...

    2023

  28. [36]

    Is a prompt and a few samples all you need? Using GPT-4 for data augmentation in low-resource classification tasks,

    A. G. Møller, J. A. Dalsgaard, A. Pera, and L. M. Aiello, “Is a prompt and a few samples all you need? Using GPT-4 for data augmentation in low-resource classification tasks,” Apr. 2023, arXiv:2304.13861 [physics]. [Online]. Available: http://arxiv.org/abs/2304.13861

    2023 arXiv

  29. [37]

    DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models,

    B. Wang, W. Chen, H. Pei, C. Xie, M. Kang, C. Zhang, C. Xu, Z. Xiong, R. Dutta, R. Schaeffer, S. T. Truong, S. Arora, M. Mazeika, D. Hendrycks, Z. Lin, Y . Cheng, S. Koyejo, D. Song, and B. Li, “DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models,” Jun. ...

    2023 arXiv

  30. [38]

    arXiv.org e-Print archive

    “arXiv.org e-Print archive.” [Online]. Available: https://arxiv.org/

  31. [2004]

    Available: https://doi.org/10.1162/089120104323093276

    [Online]. Available: https://doi.org/10.1162/089120104323093276

    doi

  32. [2024]

    Available: https://openai.com/index/ gpt-4o-mini-advancing-cost-efficient-intelligence/

    [Online]. Available: https://openai.com/index/ gpt-4o-mini-advancing-cost-efficient-intelligence/

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.