Pith. sign in

REVIEW 2 major objections 117 references

Two simple normalisations turn raw LambdaG scores into well-calibrated likelihood ratios for forensic authorship verification without any separate calibration model.

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 · grok-4.5

2026-07-13 02:33 UTC pith:IFPPDIRC

load-bearing objection Parameter-free normalisers that replace logistic-regression calibration for LambdaG, with solid Cllr evidence across 15 corpora. the 2 major comments →

arxiv 2607.09501 v1 pith:IFPPDIRC submitted 2026-07-10 cs.CL stat.AP

Normalisation-Based Likelihood Ratio Estimation for Forensic Authorship Verification

Sadie Barlow , Andrea Nini , Edoardo Manino This is my paper
classification cs.CL stat.AP
keywords Authorship VerificationLikelihood RatiosCalibrationLambdaGHapax CorrectionSquare Root CorrectionForensic LinguisticsNormalisation
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.

Most authorship-verification systems output scores that must be turned into likelihood ratios by a trained calibrator. That calibrator needs large amounts of case-matched labelled pairs, which are hard to collect, preprocess and justify in court. This paper shows that two closed-form normalisations applied to the LambdaG score already produce usable likelihood ratios. The Square Root Correction divides by the square root of the questioned-text length; the Hapax Correction multiplies by the proportion of once-occurring tokens. Across fifteen corpora and lengths from 100 to 9 500 tokens the Hapax Correction beats logistic-regression calibration in roughly 45 percent of tests and stays within 5 percent more often when it loses. Removing the calibration step therefore cuts data demand, time and opacity while keeping forensic-grade performance.

Core claim

The Square Root Correction and the Hapax Correction convert uncalibrated LambdaG scores into likelihood ratios whose log-likelihood-ratio cost (Cllr) is comparable to that of logistic-regression calibration; the Hapax Correction outperforms logistic regression in approximately 45 percent of corpus-weighted tests and, when it underperforms, remains within 5 percent more frequently than the reverse comparison.

What carries the argument

The Hapax Correction multiplies the raw LambdaG score by the ratio of hapax legomena to total tokens in the questioned text; the Square Root Correction divides the same score by the square root of that token count. Both counteract the systematic overstatement that arises when repeated stylistic patterns are treated as independent evidence.

Load-bearing premise

The claim stands only if the main source of score inflation is the naïve independence assumption applied to repeated tokens, and if scaling by square-root length or by the hapax ratio correctly compensates for that redundancy.

What would settle it

On a fresh multi-register corpus never used in the original experiments, compute mean Cllr for Hapax-normalised LambdaG versus logistic regression trained on matched same-size calibration sets; if Hapax is systematically more than 10–15 percent worse, or produces Cllr greater than 1 while logistic regression stays well below 1, the claim of comparable calibration fails.

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

If this is right

  • Forensic linguists can report numerical likelihood ratios using only the case texts and a reference corpus, without assembling a separate labelled calibration set.
  • Preparation time and cost drop because the labour of matching, anchoring and validating calibration data disappears.
  • The normalisations are transparent algebraic formulas rather than fitted parameters, simplifying explanation to courts and opposing experts.
  • Reliable performance extends to short texts (a few hundred tokens) where suitable calibration data are scarcest.
  • Authorship verification moves closer to the generative-modelling style already standard in DNA evidence evaluation.

Where Pith is reading between the lines

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

  • The same length- and diversity-based scalings may calibrate other additive, token-level AV scorers that currently require logistic regression.
  • Because the Hapax Correction is exactly Baayen’s productivity measure, lexical-diversity statistics may be reinterpreted as direct estimators of evidential weight under usage-based idiolect models.
  • If the independence assumption is the true culprit, replacing the additive sum with a non-linear aggregation of n-gram evidence could eliminate post-hoc correction entirely.
  • The methods invite targeted stress tests on deliberate impersonation, where an adversary repeats a known feature to inflate the raw score.

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

2 major / 0 minor

Summary. The paper proposes two parameter-free normalisation techniques—the Square Root Correction (ΛG = λG / √N(Q)) and the Hapax Correction (ΛG = λG × V1(Q)/N(Q))—that convert raw LambdaG scores into usable log-likelihood ratios without a separate logistic-regression calibration model. The corrections are motivated by the claim that naïve-Bayes independence assumptions cause over-inflation of evidential strength for long or repetitive texts. Performance is measured by Cllr against logistic-regression calibration on fifteen English corpora spanning multiple registers, with systematic length sweeps (100–9 500 tokens) on three of them and five random seeds. The Hapax Correction outperforms the baseline in ~45 % of corpus-weighted tests and, when it loses, is more often within 5 % of the baseline than the reverse; the Square Root Correction is weaker but still competitive on some registers. The authors conclude that the practical advantages (reduced data, time and opacity) support forensic adoption for English LambdaG scores.

Significance. If the empirical results hold, the work removes a major practical bottleneck in forensic authorship verification: the need for case-matched calibration data and the associated expert decisions about anchoring and validation. Parameter-free normalisers that still achieve Cllr comparable to logistic regression would make well-calibrated LLRs more accessible and transparent for courts. The evaluation protocol itself is a strength—fifteen diverse corpora, length-controlled experiments, multi-seed reporting of means and standard deviations, and direct head-to-head comparison with the established baseline. The paper also situates the corrections within usage-based linguistic theory and Baayen’s P, giving them more interpretability than a black-box calibrator. These contributions are of clear interest to forensic linguistics and to the broader community working on score-to-LR conversion for behavioural biometrics.

major comments (2)
  1. Section 5 and Equations 7–8: the central theoretical claim—that over-inflation arises from the naïve-Bayes independence assumption applied to repeated stylistic patterns, and that √N(Q) and V1(Q)/N(Q) correctly compensate for that redundancy—is presented only heuristically. No derivation or even a simple generative simulation is supplied to show that these particular functional forms recover the true LLR under the stated independence violation. Because the paper’s strongest claim is empirical (comparable Cllr), this gap does not invalidate the reported numbers, but it does leave the corrections as ad-hoc inventions whose domain of validity is unknown. A short analytic or simulation argument, or at least an explicit statement that optimality is not claimed, would strengthen the manuscript.
  2. Section 9 and the Bolt results (Figure 4): the Hapax Correction produces Cllr > 1 when K ≈ 100 tokens. The authors correctly note that such short samples are generally insufficient for AV, yet the paper still recommends the method for forensic use. A clearer operational boundary (minimum reliable N(Q) or V1(Q)/N(Q) range) should be stated so that practitioners know when the normaliser is expected to fail.

Circularity Check

1 steps flagged

No significant circularity: parameter-free normalisers evaluated empirically against an external logistic-regression baseline on held-out case data; minor self-citation of the base LambdaG scorer is not load-bearing for the calibration claim.

specific steps
  1. self citation load bearing [Abstract; Section 4.2; Equation (6)]
    "This study proposes two novel normalisation techniques, the Square Root Correction and the Hapax Correction, for deriving likelihood ratios from the AV method LambdaG without the need of a calibration model (Nini et al. 2026)."

    LambdaG (the source of the raw scores λ_G) is defined and validated only in the overlapping-author citation Nini et al. (2026). The present paper takes those scores as given and never re-derives them. The citation is therefore load-bearing for the existence of the input, yet it is not load-bearing for the novel claim that the two normalisers produce well-calibrated Λ_G; that claim rests on independent Cllr measurements against logistic regression. Hence only a minor, non-central circularity of degree 1.

full rationale

The paper’s central claim is empirical: two fixed, parameter-free transformations (Λ_G = λ_G / √N(Q) and Λ_G = λ_G × V1(Q)/N(Q)) applied to raw LambdaG scores yield Cllr values comparable to logistic-regression calibration across 15 corpora and length conditions (Tables 2–3, Figures 2–4). These normalisers contain no fitted coefficients; performance is measured on held-out case problems whose ground-truth labels are independent of the transformations. The only self-citation is to Nini et al. (2026) for the definition of the base scorer λ_G itself; that citation supplies the input scores but does not force or define the subsequent Cllr results of the proposed corrections. No equation equates a “prediction” to a fitted quantity by construction, no uniqueness theorem is imported, and no ansatz is smuggled in. The acknowledged theoretical gap (Section 9) concerns the missing formal derivation of why the chosen normalisers compensate for naïve-Bayes redundancy; that is a correctness/completeness issue, not circularity. The derivation chain is therefore self-contained against the external baseline.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on the LambdaG scorer (prior work), the standard Cllr metric, and two ad-hoc normalisers motivated by known lexical-statistics regularities. No free parameters are fitted to the evaluation data; the only free choices are the default LambdaG hyperparameters and the decision to count hapaxes on content-masked text.

free parameters (2)
  • LambdaG order N = 10
    Default N=10 taken from the idiolect package; not re-tuned on the evaluation corpora.
  • number of reference models r = 30
    Default r=30; paper states robustness but does not re-optimise.
axioms (4)
  • domain assumption Token-level contributions to LambdaG are approximately independent, so raw scores grow linearly with length while true evidential value grows sub-linearly.
    Stated in Section 5 as the motivation for both corrections; never proved.
  • ad hoc to paper Hapax ratio V1(Q)/N(Q) (Baayen’s P) is a sufficient statistic for the redundancy that inflates LambdaG.
    Introduced in Eq. 8 without derivation that it is optimal among diversity measures.
  • ad hoc to paper Square-root scaling of length is the correct functional form for diminishing returns of additional tokens.
    Justified by analogy to variance of weakly dependent sums and attention scaling (Section 5.2); not derived from the language model.
  • domain assumption Content-masking (POSNoise) removes topic bias without destroying the stylistic signal needed for calibration.
    Inherited from LambdaG; used throughout the experiments.
invented entities (2)
  • Square Root Correction no independent evidence
    purpose: Scale raw LambdaG by 1/√N(Q) to counteract length-driven score inflation.
    Defined in Eq. 7; no independent theoretical derivation or prior use in AV.
  • Hapax Correction no independent evidence
    purpose: Scale raw LambdaG by the hapax-to-token ratio to penalise repetitive texts.
    Defined in Eq. 8; re-uses Baayen’s P but applies it as a multiplicative normaliser for LRs for the first time.

pith-pipeline@v1.1.0-grok45 · 30514 in / 2702 out tokens · 30834 ms · 2026-07-13T02:33:51.849262+00:00 · methodology

0 comments
read the original abstract

Authorship verification (AV) is the task of determining whether two texts were written by the same author. In a forensic context, the strength of AV evidence can be quantified using likelihood ratios. Most AV methods are score-based and deriving well-calibrated likelihood ratios from these scores requires a separate calibration model. This, in turn, requires additional amounts of case-relevant data, which is often time-consuming to obtain and prepare. This study proposes two novel normalisation techniques, the Square Root Correction and the Hapax Correction, for deriving likelihood ratios from the AV method LambdaG without the need of a calibration model (Nini et al. 2026). These corrections are designed to mitigate the overestimation of evidential strength that may result from long or highly repetitive texts. Performance is evaluated against logistic regression calibration across fifteen corpora and a range of text lengths (100-9,500 tokens), using the log-likelihood ratio cost (Cllr). The proposed methods achieve performance comparable to logistic regression calibration, with the Hapax Correction outperforming it in approximately 45% of tests (weighted by corpora). Furthermore, performance was more frequently close (within 5%) when the Hapax Correction was outperformed by logistic regression calibration, compared with the reverse comparison. Eliminating the need to train a calibration model reduces data-requirements, time and complexity, thereby increasing the accessibility and transparency of forensic text comparison. This combination of empirical performance and practical advantages supports the adoption of the proposed methods in forensic settings.

Figures

Figures reproduced from arXiv: 2607.09501 by Andrea Nini, Edoardo Manino, Sadie Barlow.

Figure 1
Figure 1. Figure 1: Distribution of raw LambdaG scores for each corpus adopted from Nini et al. (2026), [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of three approaches on the Bolt dataset showing the relationship be [PITH_FULL_IMAGE:figures/full_fig_p024_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

117 extracted references · 68 canonical work pages · 1 internal anchor

  1. [1]

    Computational

    Argamon, Shlomo Engelson , year = 2019, month = may, journal =. Computational

    2019

  2. [2]

    Interpreting

    Argamon, Shlomo , year = 2008, month = jun, journal =. Interpreting. doi:10.1093/llc/fqn003 , urldate =

    work page doi:10.1093/llc/fqn003 2008

  3. [3]

    Harald , year = 2001, series =

    Baayen, R. Harald , year = 2001, series =. Word. doi:10.1007/978-94-010-0844-0 , urldate =

    work page doi:10.1007/978-94-010-0844-0 2001

  4. [4]

    Bagnall, Douglas , year = 2015, month = jun, doi =. Author

    2015

  5. [5]

    and Nichols, Richard A

    Balding, David J. and Nichols, Richard A. , year = 1994, month = feb, journal =. doi:10.1016/0379-0738(94)90222-4 , urldate =

    work page doi:10.1016/0379-0738(94)90222-4 1994

  6. [6]

    Genetica , volume =

    A Method for Quantifying Differentiation between Populations at Multi-Allelic Loci and Its Implications for Investigating Identity and Paternity , author =. Genetica , volume =. doi:10.1007/BF01441146 , urldate =

    work page doi:10.1007/bf01441146

  7. [7]

    International Journal of Corpus Linguistics , volume =

    Individual Differences and Usage-Based Grammar , author =. International Journal of Corpus Linguistics , volume =. doi:10.1075/ijcl.18.4.01bar , urldate =

    work page doi:10.1075/ijcl.18.4.01bar

  8. [8]

    Benoit, Kenneth and Matsuo, Akitaka and Gruber, Johannes , year = 2023, month = dec, doi =. Spacyr:

    2023

  9. [9]

    doi:10.1515/cllt-2012-0002 , urldate =

    Register as a Predictor of Linguistic Variation , author =. doi:10.1515/cllt-2012-0002 , urldate =

    doi:10.1515/cllt-2012-0002 2012

  10. [10]

    and Aitken, Colin Graeme Girdwood , year = 2016, month = sep, journal =

    Biedermann, Alex and Bozza, Silvia and Taroni, Franco T. and Aitken, Colin Graeme Girdwood , year = 2016, month = sep, journal =. Reframing the Debate:. doi:10.1016/j.scijus.2016.05.008 , urldate =

    work page doi:10.1016/j.scijus.2016.05.008 2016

  11. [11]

    Evaluation of

    Biosa, Giulia and Giurghita, Diana and Vincenti, Marco and Neocleous, Tereza , year = 2020, month = oct, journal =. Evaluation of. doi:10.3389/fchem.2020.00738 , urldate =

    work page doi:10.3389/fchem.2020.00738 2020

  12. [12]

    Similarity

    Boenninghoff, Benedikt and Nickel, Robert and Zeiler, Steffen and Kolossa, Dorothea , year = 2019, doi =. Similarity. International

    2019

  13. [13]

    Stylometric

    Brousalis, Georgios and Markopoulos, George and Mikros, George , year = 2012, month = jan, abstract =. Stylometric

    2012

  14. [14]

    Computer Speech & Language , series =

    Application-Independent Evaluation of Speaker Detection , author =. Computer Speech & Language , series =. doi:10.1016/j.csl.2005.08.001 , urldate =

    work page doi:10.1016/j.csl.2005.08.001 2005

  15. [15]

    Likelihood-Ratio Calibration Using Prior-Weighted Proper Scoring Rules , booktitle =

    Br. Likelihood-Ratio Calibration Using Prior-Weighted Proper Scoring Rules , booktitle =. doi:10.21437/Interspeech.2013-470 , urldate =

    work page doi:10.21437/interspeech.2013-470 2013

  16. [16]

    and Bright, Jo-Anne and Gittelson, Simone and Moretti, Tamyra R

    Buckleton, John S. and Bright, Jo-Anne and Gittelson, Simone and Moretti, Tamyra R. and Onorato, Anthony J. and Bieber, Frederick R. and Budowle, Bruce and Taylor, Duncan A. , year = 2019, journal =. The. doi:10.1111/1556-4029.13898 , urldate =

    work page doi:10.1111/1556-4029.13898 2019

  17. [17]

    Computer Speech & Language , volume =

    An Empirical Study of Smoothing Techniques for Language Modeling , author =. Computer Speech & Language , volume =. doi:10.1006/csla.1999.0128 , abstract =

    doi:10.1006/csla.1999.0128 1999

  18. [18]

    doi:10.35111/HKFC-7865 , abstract =

    work page doi:10.35111/hkfc-7865

  19. [19]

    and Chater, Nick , year = 2016, journal =

    Christiansen, Morten H. and Chater, Nick , year = 2016, journal =. The. doi:10.1017/S0140525X1500031X , urldate =

    doi:10.1017/s0140525x1500031x 2016

  20. [20]

    , year = 1994, month = jul, journal =

    Collins, Andrew and Morton, Newton E. , year = 1994, month = jul, journal =. Likelihood Ratios for. doi:10.1073/pnas.91.13.6007 , abstract =

    work page doi:10.1073/pnas.91.13.6007 1994

  21. [21]

    Coulthard, Malcolm , year = 2004, month = dec, journal =. Author. doi:10.1093/applin/25.4.431 , urldate =

    work page doi:10.1093/applin/25.4.431 2004

  22. [22]

    Dawid, A. P. , year = 1982, month = sep, journal =. The. doi:10.1080/01621459.1982.10477856 , urldate =

    doi:10.1080/01621459.1982.10477856 1982

  23. [23]

    and Fienberg, Stephen E

    DeGroot, Morris H. and Fienberg, Stephen E. , year = 1983, journal =. The. doi:10.2307/2987588 , urldate =. 2987588 , eprinttype =

    doi:10.2307/2987588 1983

  24. [24]

    Text & Talk , volume =

    The Idiom Principle and the Open Choice Principle , author =. Text & Talk , volume =. doi:10.1515/text.1.2000.20.1.29 , urldate =

    work page doi:10.1515/text.1.2000.20.1.29 2000

  25. [25]

    Understanding and Explaining

    Evert, Stefan and Proisl, Thomas and Jannidis, Fotis and Reger, Isabella and Pielstr. Understanding and Explaining. Digital Scholarship in the Humanities , volume =. doi:10.1093/llc/fqx023 , urldate =

    work page doi:10.1093/llc/fqx023

  26. [26]

    Journal of Experimental Psychology: Learning, Memory, and Cognition , volume =

    Repetition Priming and Frequency Attenuation in Lexical Access , author =. Journal of Experimental Psychology: Learning, Memory, and Cognition , volume =. doi:10.1037/0278-7393.10.4.680 , abstract =

    work page doi:10.1037/0278-7393.10.4.680

  27. [27]

    Proceedings of the National Academy of Sciences , volume =

    Interpretable Algorithmic Forensics , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.2301842120 , abstract =

    work page doi:10.1073/pnas.2301842120

  28. [28]

    Statistical

    Geoffrey, Stewart Morrison and Ewald, Enzinger and Ramos, Daniel and. Statistical. Handbook of

  29. [29]

    Gill, Peter and Benschop, Corina and Buckleton, John and Bleka,. A. Genes , volume =. doi:10.3390/genes12101559 , urldate =

    work page doi:10.3390/genes12101559

  30. [30]

    Journal of the Royal Statistical Society , series =

    Probabilistic Forecasts, Calibration and Sharpness , author =. Journal of the Royal Statistical Society , series =

  31. [31]

    Trends in Cognitive Sciences , volume =

    Chunking Mechanisms in Human Learning , author =. Trends in Cognitive Sciences , volume =. doi:10.1016/S1364-6613(00)01662-4 , urldate =

    work page doi:10.1016/s1364-6613(00)01662-4

  32. [32]

    IEEE Transactions on Audio, Speech, and Language Processing , volume =

    Emulating. IEEE Transactions on Audio, Speech, and Language Processing , volume =. doi:10.1109/TASL.2007.902747 , urldate =

    work page doi:10.1109/tasl.2007.902747 2007

  33. [33]

    Grant, Tim , year = 2022, series =. The. doi:10.1017/9781108974714 , urldate =

    work page doi:10.1017/9781108974714 2022

  34. [34]

    Attributing the

    Grieve, Jack and Clarke, Isobelle and Chiang, Emily and Gideon, Hannah and Heini, Annina and Nini, Andrea and Waibel, Emily , year = 2019, month = sep, journal =. Attributing the. doi:10.1093/llc/fqy042 , urldate =

    work page doi:10.1093/llc/fqy042 2019

  35. [35]

    Frontiers

    Grieve, Jack and Montgomery, Chris and Nini, Andrea and Murakami, Akira and Guo, Diansheng , year = 2019, doi =. Frontiers

    2019

  36. [36]

    , year = 2017, month = jun, journal =

    Guo, Chuan and Pleiss, Geoff and Sun, Yu and Weinberger, Kilian Q. , year = 2017, month = jun, journal =. On. doi:10.48550/arXiv.1706.04599 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.04599 2017

  37. [37]

    Proceedings of the 16th

    Halvani, Oren and Graner, Lukas , year = 2021, month = aug, series =. Proceedings of the 16th. doi:10.1145/3465481.3470050 , abstract =

    doi:10.1145/3465481.3470050 2021

  38. [38]

    Practice-

    Halvani, Oren , year = 2021, address =. Practice-

    2021

  39. [39]

    Halvani, Oren and Winter, Christian and Graner, Lukas , year = 2017, month = aug, series =. On the. Proceedings of the 12th. doi:10.1145/3098954.3104050 , urldate =

    work page doi:10.1145/3098954.3104050 2017

  40. [40]

    Heaps, H. S. , year = 1978, publisher =. Information

    1978

  41. [41]

    Forensic Science International , volume =

    Score-Based Likelihood Ratios for Handwriting Evidence , author =. Forensic Science International , volume =. doi:10.1016/j.forsciint.2011.12.009 , abstract =

    work page doi:10.1016/j.forsciint.2011.12.009 2011

  42. [42]

    Type-Token

    Herdan, Gustav , year = 1960, publisher =. Type-Token

    1960

  43. [43]

    Knowledge-Based Systems , volume =

    Understanding Writing Style in Social Media with a Supervised Contrastively Pre-Trained Transformer , author =. Knowledge-Based Systems , volume =. doi:10.1016/j.knosys.2024.111867 , abstract =

    doi:10.1016/j.knosys.2024.111867 2024

  44. [44]

    Estimating the

    Ishihara, Shunichi and Tsuge, Satoru and Inaba, Mitsuyuki and Zaitsu, Wataru , editor =. Estimating the. Proceedings of the 20th

  45. [45]

    Ishihara, Shunichi , editor =. A. Proceedings of the

  46. [46]

    Forensic Science International , volume =

    Score-Based Likelihood Ratios for Linguistic Text Evidence with a Bag-of-Words Model , author =. Forensic Science International , volume =. doi:10.1016/j.forsciint.2021.110980 , urldate =

    work page doi:10.1016/j.forsciint.2021.110980 2021

  47. [47]

    Strength of Forensic Text Comparison Evidence from Stylometric Features:

    Ishihara, Shunichi , year = 2017, month = may, journal =. Strength of Forensic Text Comparison Evidence from Stylometric Features:. doi:10.1558/ijsll.30305 , abstract =

    work page doi:10.1558/ijsll.30305 2017

  48. [48]

    Weight of Authorship Evidence with Multiple Categories of Stylometric Features:

    Ishihara, Shunichi , year = 2023, month = mar, journal =. Weight of Authorship Evidence with Multiple Categories of Stylometric Features:. doi:10.1016/j.scijus.2022.12.007 , urldate =

    work page doi:10.1016/j.scijus.2022.12.007 2023

  49. [49]

    Overview of the Author Identification Task at

    Juola, Patrick and Stamatatos, Efstathios , year = 2013, journal =. Overview of the Author Identification Task at

    2013

  50. [50]

    Verifying Authorship for Forensic Purposes:

    Juola, Patrick , year = 2021, month = aug, journal =. Verifying Authorship for Forensic Purposes:. doi:10.1016/j.forsciint.2021.110824 , urldate =

    work page doi:10.1016/j.forsciint.2021.110824 2021

  51. [51]

    , year = 2026, edition =

    Jurafsky, Daniel and Martin, James H. , year = 2026, edition =. Speech and

    2026

  52. [52]

    A Sensitivity Analysis to Determine the Robustness of

    Kelly, Hannah and Bright, Jo-Anne and Kruijver, Maarten and Cooper, Stuart and Taylor, Duncan and Duke, Kyle and Strong, Melissa and Beamer, Vickie and Buettner, Christina and Buckleton, John , year = 2018, month = jul, journal =. A Sensitivity Analysis to Determine the Robustness of. doi:10.1016/j.fsigen.2018.04.009 , urldate =

    work page doi:10.1016/j.fsigen.2018.04.009 2018

  53. [53]

    Overview of the

    Kestemont, Mike and Manjavacas, Enrique and Markov, Ilia and Bevendorff, Janek and Wiegmann, Matti and Stamatatos, Efstathios and Stein, Benno and Potthast, Martin , year = 2021, month = sep, abstract =. Overview of the

    2021

  54. [54]

    Improved Backing-off for

    Kneser,. Improved Backing-off for. 1995. doi:10.1109/ICASSP.1995.479394 , urldate =

    doi:10.1109/icassp.1995.479394 1995

  55. [55]

    Authorship Verification as a One-Class Classification Problem , booktitle =

    Koppel, Moshe and Schler, Jonathan , year = 2004, month = jul, series =. Authorship Verification as a One-Class Classification Problem , booktitle =. doi:10.1145/1015330.1015448 , urldate =

    work page doi:10.1145/1015330.1015448 2004

  56. [56]

    Journal of the Association for Information Science and Technology , volume =

    Determining If Two Documents Are Written by the Same Author , author =. Journal of the Association for Information Science and Technology , volume =. doi:10.1002/asi.22954 , abstract =

    doi:10.1002/asi.22954

  57. [57]

    Koppel, Moshe and Schler, Jonathan and Argamon, Shlomo and Winter, Yaron , year = 2012, month = may, journal =. The ``. doi:10.1080/0013838X.2012.668794 , urldate =

    work page doi:10.1080/0013838x.2012.668794 2012

  58. [58]

    , year = 1987, month = jun, publisher =

    Langacker, Ronald W. , year = 1987, month = jun, publisher =. Foundations of

    1987

  59. [59]

    Forensic Science International: Synergy , volume =

    Improved Likelihood Ratios for Face Recognition in Surveillance Video by Multimodal Feature Pairing , author =. Forensic Science International: Synergy , volume =. doi:10.1016/j.fsisyn.2024.100458 , urldate =

    work page doi:10.1016/j.fsisyn.2024.100458 2024

  60. [60]

    , year = 1956, journal =

    Miller, George A. , year = 1956, journal =. The Magical Number Seven, plus or Minus Two:. doi:10.1037/h0043158 , abstract =

    doi:10.1037/h0043158 1956

  61. [61]

    and Cheney, Nick , year = 2025, month = jun, journal =

    Mitchell, Kevin J. and Cheney, Nick , year = 2025, month = jun, journal =. The. doi:10.1016/j.tig.2025.01.008 , urldate =

    work page doi:10.1016/j.tig.2025.01.008 2025

  62. [62]

    Mollin, Sandra , year = 2009, month = aug, publisher =. ``. doi:10.1075/ijcl.14.3.04mol , abstract =

    work page doi:10.1075/ijcl.14.3.04mol 2009

  63. [63]

    Morrison, Geoffrey Stewart , year = 2024, month = jan, journal =. Bi-. doi:10.1093/lpr/mgae004 , urldate =

    work page doi:10.1093/lpr/mgae004 2024

  64. [64]

    Forensic Science International: Synergy , volume =

    In the Context of Forensic Casework, Are There Meaningful Metrics of the Degree of Calibration? , author =. Forensic Science International: Synergy , volume =. doi:10.1016/j.fsisyn.2021.100157 , urldate =

    work page doi:10.1016/j.fsisyn.2021.100157 2021

  65. [65]

    Forensic Science International , volume =

    An Empirical Estimate of the Precision of Likelihood Ratios from a Forensic-Voice-Comparison System , author =. Forensic Science International , volume =. doi:10.1016/j.forsciint.2010.11.001 , urldate =

    work page doi:10.1016/j.forsciint.2010.11.001 2010

  66. [66]

    Science & Justice , volume =

    Measuring the Validity and Reliability of Forensic Likelihood-Ratio Systems , author =. Science & Justice , volume =. doi:10.1016/j.scijus.2011.03.002 , urldate =

    work page doi:10.1016/j.scijus.2011.03.002 2011

  67. [67]

    Journal of Forensic Sciences , volume =

    Source-Anchored, Trace-Anchored, and General Match Score-Based Likelihood Ratios for Camera Device Identification , author =. Journal of Forensic Sciences , volume =. doi:10.1111/1556-4029.14991 , urldate =

    work page doi:10.1111/1556-4029.14991

  68. [68]

    Forensic

    Roemling, Dana and Grieve, Jack , year = 2024, month = mar, journal =. Forensic

    2024

  69. [69]

    Grant, Tim and Grieve, Jack , year = 2022, pages =. The. Methodologies and. doi:10.1002/9781394266661.ch2 , urldate =

    work page doi:10.1002/9781394266661.ch2 2022

  70. [70]

    Proceedings of the 2021

    Learning. Proceedings of the 2021. doi:10.18653/v1/2021.emnlp-main.70 , urldate =

    doi:10.18653/v1/2021.emnlp-main.70 2021

  71. [71]

    Idiolect:

    Nini, Andrea , year = 2026, volume =. Idiolect:

    2026

  72. [72]

    Score Based Procedures for the Calculation of Forensic Likelihood Ratios --

    Morrison, Geoffrey Stewart and Enzinger, Ewald , year = 2018, month = jan, journal =. Score Based Procedures for the Calculation of Forensic Likelihood Ratios --. doi:10.1016/j.scijus.2017.06.005 , urldate =

    doi:10.1016/j.scijus.2017.06.005 2018

  73. [73]

    Tutorial on Logistic-Regression Calibration and Fusion:Converting a Score to a Likelihood Ratio , shorttitle =

    Morrison, Geoffrey Stewart , year = 2013, month = jun, journal =. Tutorial on Logistic-Regression Calibration and Fusion:Converting a Score to a Likelihood Ratio , shorttitle =. doi:10.1080/00450618.2012.733025 , urldate =

    work page doi:10.1080/00450618.2012.733025 2013

  74. [74]

    Minhash and Locality-Sensitive Hashing , author =

  75. [75]

    Grammar as a Behavioral Biometric: Using Cognitively Motivated Grammar Models for Authorship Verification

    Nini, Andrea and Halvani, Oren and Graner, Lukas and Gherardi, Valerio and Ishihara, Shunichi , year =. Grammar as a. doi:10.48550/arXiv.2403.08462 , urldate =. arXiv , keywords =:2403.08462 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2403.08462

  76. [76]

    Idiolect:

    Nini, Andrea , year = 2024, doi =. Idiolect:

    2024

  77. [77]

    The International Journal of Speech, Language and the Law , volume =

    Register Variation in Malicious Forensic Texts , author =. The International Journal of Speech, Language and the Law , volume =. doi:10.1558/ijsll.30173 , urldate =

    work page doi:10.1558/ijsll.30173

  78. [78]

    Nini, Andrea , year = 2023, series =. A. doi:10.1017/9781108974851 , abstract =

    work page doi:10.1017/9781108974851 2023

  79. [79]

    doi:10.35111/WQT6-JG60 , abstract =

    Avocado. doi:10.35111/WQT6-JG60 , abstract =

    work page doi:10.35111/wqt6-jg60

  80. [80]

    , year = 1999, journal =

    Platt, John C. , year = 1999, journal =. Probabilistic

    1999

Showing first 80 references.