Pith. sign in

REVIEW 4 major objections 5 minor 46 references

A Fisher's exact test justification of the TF-IDF term-weighting scheme

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read TF-IDF is a Fisher's exact test p-value, up to corrections.

desk verdict Plausible new expansion linking TF-ICF to the hypergeometric p-value, but Lemma 1 is false as stated and the corollaries overstate the asymptotics; worth a serious referee but needs major repair. read the letter →

arxiv 2507.15742 v2 pith:KPY4AZWE submitted 2025-07-21 cs.CL cs.IRmath.STstat.TH

Paul Sheridan , Zeyad Ahmed , Aitazaz A. Farooque This is my paper · ORCID
classification cs.CLcs.IRmath.STstat.TH MSC 62F0362E1768P20
keywords TF-IDFTF-ICFFisher'sexacttesthypergeometricdistributiontermweightingsignificancetestinginformationretrievaltextanalysis
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

This paper argues that the classic term-weighting formula TF-IDF, a workhorse of information retrieval for over fifty years, is not merely a heuristic but a significance test in disguise. The authors show that the TF-ICF variant equals, up to explicit correction and error terms, the negative logarithm of the one-tailed Fisher's exact test p-value for over-representation of a term in a document. Under idealized conditions, such as equal document lengths and a term that appears with the same fixed count in every document where it appears, TF-IDF itself tracks the same p-value; in the limit of an infinitely large document collection the match becomes exact to order $O(1/d)$. The contribution is a statistical significance-testing rationale for why rare-but-concentrated terms deserve high weight.

What carries the argument

The load-bearing object is the tail probability $H_{ij} = H(n_{ij}; n_i, n_j, n)$ of the hypergeometric distribution, which is the exact p-value of the one-tailed Fisher test, compared term by term with the binomial probability $b(n_{ij}; n_j, p_i)$. Lemma 1 bounds the ratio $Q_{ij} = H(n_{ij}+1; n_i, n_j, n)/b(n_{ij}; n_j, p_i)$ in $(0,1)$ under large-count and small-$p_i$ conditions; that bound licenses the Taylor expansion $\log(1+Q_{ij}) = Q_{ij} + o(Q_{ij})$. The proof then uses Stirling's approximation for binomial coefficients, Chvátal's tail bound for the hypergeometric distribution, and a binomial approximation to the hypergeometric mass with $O(1/n)$ error to isolate TF-ICF as the leading term.

What would settle it

Compute $D_{ij} = \frac{1}{n_j}\log b(n_{ij}; n_j, p_i) - \frac{1}{n_j}\log H(n_{ij}+1; n_i, n_j, n)$ for small parameters such as $n=12{,}500$, $n_i=6$, $n_j=80$, $n_{ij}=2$, a case the paper's Table 4 already tests: if $D_{ij} \le 0$, the claimed $0 < Q_{ij} < 1$ fails and the Lemma 1 expansion is invalid at that point. More generally, a grid search over small $p_i$ and small $n_{ij}$ that locates any nonpositive $D_{ij}$ would settle the matter directly.

Watch

Extended reading notes

Core claim

The central result is Theorem 1: for term $i$ and document $j$, with $H_{ij}$ the one-tailed Fisher's exact test p-value (a hypergeometric tail probability), the relation $-\log H_{ij} = \mathrm{TF\text{-}ICF}(i,j) + n_{ij}\log p_{ij} + (n_j - n_{ij})(p_i - p_{ij}) - Q_{ij} + \text{error terms}$ holds when $p_i$ is small relative to $p_{ij}$ and the counts involved are large. The leading term is exactly $\mathrm{TF\text{-}ICF}(i,j) = n_{ij}\log(n/n_i)$; the remaining terms are finite corrections, the largest coming from the binomial probability of the observed count. Corollary 1 converts TF-ICF into TF-IDF under equal document lengths and a term pattern where the term appears in each document either with the same small count $r$ or not at all; Corollary 2 makes the approximation $-\log H_{ij} = \mathrm{TF\text{-}IDF}(i,j) + O(1/d)$ in the limiting case where the term appears exclusively in the target document. The paper is explicit that the clean corollaries rest on idealized assumptions and should be read as bridges between term weighting and significance testing rather than as general equivalences.

Load-bearing premise

Lemma 1, and therefore the Taylor expansion that produces the main approximation, requires the quotient $Q_{ij}$ to lie strictly between 0 and 1, and the paper's proof of that fact assumes, without a quantified bound, that the term's overall rarity $p_i$ is small enough and the document counts are large enough for one positive term to dominate all negative ones.

Editorial extensions

If this is right

  • Statisticians can explain TF-IDF as a significance test: a high weight means the term is statistically over-represented in the document, measured by a p-value.
  • The correction terms in Theorem 1 give a concrete account of when TF-ICF and TF-IDF mislead, such as when the term appears too rarely in the document or too evenly across the collection.
  • Under equal-length documents and a diagnostic-term pattern, TF-IDF approximates the negative log p-value with error that shrinks as the collection grows.
  • In the idealized limiting case of an exclusive-term collection, ranking by TF-IDF is asymptotically equivalent to ranking by the Fisher exact test p-value.
  • The result places TF-IDF on the same conceptual footing as established enrichment tests used in bioinformatics, where negative log p-values are standard significance scores.

Reading between the lines

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

  • If TF-IDF is an approximate significance score, one could calibrate TF-IDF thresholds to standard p-value cutoffs such as 0.05 instead of tuning arbitrary weight cutoffs; the paper does not propose this.
  • The correction term $\Psi_{ij}$ predicts that TF-IDF will diverge most from the exact test when terms are common across the collection or occupy a large share of the document; testing this on real corpora with unequal document lengths would show whether the significance-test view survives outside the idealized regime.
  • The word-burstiness discussion suggests a testable empirical claim: TF-IDF's practical advantage should be largest on bursty corpora, because that is precisely where the hypergeometric tail and the TF-ICF leading term track each other.
  • The paper's sketch of a Wallenius noncentral hypergeometric weighting could be developed into a concrete test: fit such a weighted sampling model and check whether its p-values still approximate TF-IDF when the sampling is biased toward rare terms.
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 significance-testing justification for TF-IDF. In a bag-of-words corpus it defines the one-tailed Fisher exact test p-value H_ij for over-representation of term i in document j, and claims in Theorem 1 that -log H_ij can be expanded as TF-ICF(i,j) plus a correction term Phi_ij = n_ij log p_ij + (n_j - n_ij)(p_i - p_ij) - Q_ij, where Q_ij = H(n_ij+1)/b(n_ij), with error terms under stated regularity conditions. Corollary 1 gives a TF-IDF version under equal-length documents and constant occurrence counts; Corollary 2 claims -log H_ij = TF-IDF(i,j) + O(1/d) when the term appears exclusively or not at all in each document. Numerical experiments in Section 5 illustrate the approximations and a GitHub repository is provided. The manuscript is clearly organized and carefully situated in the existing TF-IDF justification literature, but the formal claims as stated are not correct, and the proof of the central lemma is not self-contained.

Significance. If established, a Fisher-exact-test justification would be a useful addition to the long list of theoretical rationales for TF-IDF, and it would provide statisticians with an intuitive interpretation of an influential heuristic. The paper has genuine strengths: a clean notational setup, a thorough survey in Sections 2 and 3, reproducible numerical code, and an honest discussion of the restrictiveness of the idealized assumptions. However, the advertised contribution is substantially weaker than it appears: the main expansion relates -log H_ij to TF-ICF plus a correction term that itself contains the hypergeometric tail H(n_ij+1), and the direct TF-IDF results rely on quite restrictive conditions. More importantly, the key lemma supporting the expansion is false as stated, and Corollary 2 contains a false additive asymptotic claim. The intended results may be salvageable, but the manuscript in its current form does not establish them.

major comments (4)
  1. [Section 4, Lemma 1] Lemma 1 (0 < Q_ij < 1) is false as stated. Take n = 10^9, n_i = 10^6, n_j = 10^6, and n_ij = 10^3, so that p_i = p_ij = 10^{-3}. Then b(n_ij) = b(1000; 10^6, 10^{-3}) is approximately 0.0126 and H(n_ij+1) = P(X >= 1001) is approximately 0.5, giving Q_ij approximately 40, which is larger than 1. All stated conditions of the lemma (p_i sufficiently small and n_j, n_ij, n_j - n_ij sufficiently large) are satisfied. The failure persists if p_i is made arbitrarily small while n_j p_i is kept large, so no threshold-based reading of 'sufficiently small' repairs the statement. The proof is also not self-contained: the key simplification of W(p_ij) is relegated to 'supplementary materials' that are not included with the manuscript, and the assertion that the negative terms 'none possess sufficient magnitude' to overcome the positive term is not quantified. In addition, the proof silently uses the stronger condition p_i much smaller than p_ij, which is not an assumption of the lemma.
  2. [Section 4, Theorem 1] Theorem 1 inherits the failure of Lemma 1: the proof requires the expansion log(1 + Q_ij + O(1/n)) = Q_ij + o(Q_ij), which presupposes that Q_ij + O(1/n) is small, but Lemma 1 only purports to prove Q_ij < 1 and does not do so. Even if Lemma 1 were corrected to assume p_i << p_ij, Eq. (10)-(11) contain an unjustified step: after writing the hypergeometric probability h as b + O(1/n), the proof divides by b and absorbs 1/b into the O(1/n) term on the grounds that b is 'constant with respect to n.' In the regime of Theorem 1, where n_j and n_ij are large and p_i and p_ij tend to 0, b(n_ij; n_j, p_i) can be exponentially small, so the relative error O(1/n)/b is not negligible. The theorem also mixes several asymptotic limits (n -> infinity, n_j -> infinity, p_i -> 0, p_ij -> 0) without quantifying rates or uniformity, so 'sufficiently large' and 'sufficiently small' are not enough to make Eq. (8) a well-defined asymptotic statement.
  3. [Section 4, Corollary 2] The claimed additive error O(1/d) in Eq. (14) is false for fixed R > 1 and fixed b_i. Under the hypotheses of the corollary, n = R d and n_i = R b_i, so -log H_ij = log C(R d, R) - log C(R b_i, R). Expanding this as d -> infinity gives -log H_ij - TF-IDF(i,j) = (R - 1)/(2 b_i) + o(1), which does not vanish. For example, with R = 20 and b_i = 8 the limiting difference is 19/16, approximately 1.19; the numerical entries in Table 3 (errors 1.0477 at d = 50 and 2.3766 at d = 125 for different R) are consistent with a nonvanishing constant rather than O(1/d). The corollary should be restated as an asymptotic equivalence, or with an error term of the form O(1/d + 1/b_i).
  4. [Section 5, Table 3] The numerical validation of Theorem 1 compares -log H_ij with TF-ICF(i,j) + Phi_ij, where Phi_ij is computed from the true hypergeometric tail. Since Phi_ij contains -Q_ij = -H(n_ij+1)/b(n_ij), the agreement largely reflects the identity -log H_ij = -log(b + H(n_ij+1) + O(1/n)) by construction; it does not test whether the TF-ICF term alone approximates the p-value. This is not circular in a logical sense, but it substantially weakens the advertised conclusion that TF-ICF is 'closely related' to the negative logarithm of the p-value. The paper should either provide numerical evidence in a regime where -log H_ij and TF-ICF are close without the Phi correction, or explicitly qualify the claim to state that TF-ICF arises inside an expansion with a non-negligible correction.
minor comments (5)
  1. [Section 4, Eq. (5)] In the displayed expansion of (1/n_j) log C(n_j, n_ij), the term '- (1/n_ij) log n_ij!' should read '- (1/n_j) log n_ij!'.
  2. [Section 4, Lemma 1 proof] The proof refers to 'supplementary materials' for the simplification of W(p_ij), but no supplementary materials are included with the submitted text; the missing material should either be supplied or the proof should be completed in the main text.
  3. [Section 4, Theorem 1 proof] The proof states that the regularity conditions require n_i sufficiently large, while the theorem statement only lists n_j and n_ij as sufficiently large; these conditions should be aligned.
  4. [Throughout] There are several typos and infelicities, for example 'are are unknown' and 'going from from list'. A careful proofread would improve readability.
  5. [Table 3] In the 'Thrm. 1 verification' columns, the row 'TF-IDF(i,j) + Psi_ij' evaluates a formula that is only claimed under Corollary 1, not under Theorem 1; labeling this row in a Theorem 1 verification table may confuse readers.

Circularity Check

1 steps flagged · score 6.0 of 10

Theorem 1's approximation embeds the p-value's own tail in its correction term, making the numerical agreement partly by construction.

  1. self definitional [Section 4, Theorem 1 / Lemma 1, Eqs. (8)-(10); Section 5, Table 3]
    "Q_ij := H(n_ij + 1; n_i, n_j, n)/b(n_ij; n_j, p_i) satisfies 0 < Q_ij < 1 ... −logH_ij = −log (h(n_ij; n_i, n_j, n) + H(n_ij + 1; n_i, n_j, n)) = −log (b(n_ij; n_j, p_i) + H(n_ij + 1; n_i, n_j, n) + O(1/n)) ... −logH_ij ≈ TF-ICF(i, j) + Φ_ij (9) with Φ_ij := n_ij logp_ij + (n_j − n_ij)(p_i − p_ij) − Q_ij"

    Q_ij is the tail H(n_ij+1) of the very p-value being approximated, normalized by the binomial mass; since H_ij = h(n_ij) + H(n_ij+1), placing Q_ij inside Φ_ij means the right-hand side of Eq. (9) contains the target quantity. The residual is only the o(Q_ij) plus Stirling/Taylor remainder terms, so the closeness of TF-ICF+Φ to -logH_ij is not an independent finding. Table 3 computes Φ_ij from the true hypergeometric tail, so the small |Δ| values are forced by construction; indeed the raw TF-ICF values in Table 3 are far from -logH (e.g., 40.2359 vs 5.5429 in the small-n Theorem-1 column). Corollary 2 is independent because Q_ij vanishes there, but the main Theorem-1 relation and its numerical validation are partially self-definitional.

full rationale

The central Theorem-1 relation is not a free-standing approximation: its correction term Φ_ij includes Q_ij = H(n_ij+1)/b(n_ij), i.e., the normalized tail of the same p-value being approximated, because H_ij = h(n_ij) + H(n_ij+1). Consequently, the numerical agreement in Table 3 between TF-ICF+Φ and -logH is induced by construction, and the raw TF-ICF values in that table are far from -logH, so the paper's headline 'closeness' is carried by a correction containing the very quantity being explained. This is a genuine self-definitional element. The derivation is not entirely empty: expanding -log b(n_ij; n_j, p_i) into TF-ICF plus interpretable p_ij/p_i terms is real algebra, and Corollary 2, where Q_ij = 0 identically, provides an independent limiting connection to TF-IDF. The Sheridan-Onsjö citation is motivational only and is not load-bearing. Separately, Lemma 1's proof is incomplete as stated (the dominance of (p_ij - ˇp_ij) log p_i over the bounded negative terms is asserted without a quantified bound, and the proof refers to absent supplementary materials); that is a correctness/rigor gap rather than a circularity, so it does not itself increase the circularity score. Overall, one central 'prediction/validation' reduces by construction, while independent content remains in the corollaries, hence a partial circularity score of 6.

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

The central claim rests on standard asymptotic tools (Stirling's formula, Chvátal's tail inequality, the binomial approximation to the hypergeometric), the bag-of-words modeling assumption, and the unproven Lemma 1, which supplies the 0<Q_ij<1 condition used to expand log(1+Q_ij). No free parameters are fitted, and no new entities are introduced.

assumptions (5)
  • standard math Stirling's approximation for factorials is used to expand log C(n_j, n_ij) and log b(n_ij).
    Invoked in the proof of Lemma 1 (Eq. 5) and Theorem 1 (Stirling step before Eq. 8).
  • standard math Chvátal's tail inequality for the hypergeometric distribution bounds H(n_ij+1).
    Used to derive W_H(p_ij) in the proof of Lemma 1; cited as Chvátal (1979).
  • standard math The hypergeometric probability mass function is approximated by the binomial with O(1/n) error (Jaioun and Teerapabolarn 2014).
    This is the starting point of Eq. (10) in Theorem 1; the bound is cited but not reproduced.
  • domain assumption The bag-of-words model represents documents as multisets of terms, and term occurrences follow a hypergeometric sampling scheme under the null hypothesis of no association.
    Stated in Section 2.1 and used to define the contingency table and Fisher's exact test in Section 2.3.
  • ad hoc to paper Lemma 1: Q_ij < 1 under 'sufficiently large' n_j, n_ij, n_j - n_ij and 'sufficiently small' p_i.
    This unproven lemma is load-bearing for the log(1+Q) expansion in Theorem 1; its proof is sketched and depends on absent supplementary material.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Fisher's exact test justification of the TF-IDF term-weighting scheme." pith.science (2026). https://pith.science/paper/KPY4AZWE

@misc{pith2026250715742,
  author       = {Pith},
  title        = {Pith review of: A Fisher's exact test justification of the TF-IDF term-weighting scheme},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KPY4AZWE}},
  note         = {Machine review of arXiv:2507.15742}
}
abstract

Term frequency-inverse document frequency, or TF-IDF for short, is arguably the most celebrated mathematical expression in the history of information retrieval. Conceived as a simple heuristic quantifying the extent to which a given term's occurrences are concentrated in any one given document out of many, TF-IDF and its many variants are routinely used as term-weighting schemes in diverse text analysis applications. There is a growing body of scholarship dedicated to placing TF-IDF on a sound theoretical foundation. Building on that tradition, this paper justifies the use of TF-IDF to the statistics community by demonstrating how the famed expression can be understood from a significance testing perspective. We show that the common TF-IDF variant TF-ICF is, under mild regularity conditions, closely related to the negative logarithm of the $p$-value from a one-tailed version of Fisher's exact test of statistical significance. As a corollary, we establish a connection between TF-IDF and the said negative log-transformed $p$-value under certain idealized assumptions. We further demonstrate, as a limiting case, that this same quantity converges to TF-IDF in the limit of an infinitely large document collection. The Fisher's exact test justification of TF-IDF equips the working statistician with a ready explanation of the term-weighting scheme's long-established effectiveness.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 45 canonical work pages

  1. [1]

    Aizawa, A. (2003). An information-theoretic perspective of tf-idf measures. Information Processing and Management\/ 39\/ (1), 45–65

    work page 2003

  2. [2]

    Alshehri, A. and A. Algarni (2023). TF - TDA : A novel supervised term weighting scheme for sentiment analysis. Electronics\/ 12\/ (7), 1632

    work page 2023

  3. [3]

    Amati, G. and C. J. Van Rijsbergen (2002). Probabilistic models of information retrieval based on measuring the divergence from randomness. ACM Transactions on Information Systems\/ 20\/ (4), 357–389

    work page 2002

  4. [4]

    Cao, J. (2017). Bayesian functional enrichment analysis for the reactome database. Statistical Theory and Related Fields\/ 1\/ (2), 185--193

    work page 2017

  5. [5]

    Chv\' a tal, V. (1979). The tail of the hypergeometric distribution. Discrete Mathematics\/ 25\/ (3), 285–287

    work page 1979

  6. [6]

    de Vries, A. P. and T. Roelleke (2005). Relevance information: A loss of entropy but a gain for IDF ? In Proceedings of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval , SIGIR ’05, New York, NY, USA, pp.\ 282–289. Association for Computing Machinery

    work page 2005

  7. [7]

    Dogan, T. and A. K. Uysal (2019). Improved inverse gravity moment term weighting for text classification. Expert Systems with Applications\/ 130 , 45--59

    work page 2019

  8. [8]

    Dunning, T. (1994). Accurate methods for the statistics of surprise and coincidence. Computational Linguistics\/ 19\/ (1), 61--74

    work page 1994

Show all 46 references
  1. [9]

    Elkan, C. (2005). Deriving TF-IDF as a Fisher kernel. In Proceedings of the 12th International Conference on String Processing and Information Retrieval , SPIRE'05, Berlin, Heidelberg, pp.\ 295–300. Springer-Verlag

    2005

  2. [10]

    Elkan, C. (2007). Method and System for Selecting Documents by Measuring Document Quality ( U.S. Patent No. 7,200,606 ). U.S. Patent and Trademark Office

    2007

  3. [11]

    Chovanec, and P

    Fields, J., K. Chovanec, and P. Madiraju (2024). A survey of text classification with transformers: H ow wide? H ow large? H ow long? H ow accurate? H ow expensive? H ow safe? IEEE Access\/ 12 , 6518--6531

    2024

  4. [12]

    Fisher, R. A. (1934). Statistical Methods for Research Workers . Oliver and Boyd. (Originally published 1925, 14th ed. 1970)

    1934

  5. [13]

    Fisher, R. A. (1935). The logic of inductive inference. Journal of the Royal Statistical Society\/ 98\/ (1), 39--82

    1935

  6. [14]

    Diera, B

    Galke, L., A. Diera, B. X. Lin, B. Khera, T. Meuser, T. Singhal, F. Karl, and A. Scherp (2023). Are we really making much progress in text classification? A comparative review. arXiv preprint arXiv:2204.03954

    2023 arXiv

  7. [15]

    L \'o pez-Dom \' nguez, J

    Garcia-Moreno, A., R. L \'o pez-Dom \' nguez, J. A. Villatoro-Garc \' a, A. Ramirez-Mena, E. Aparicio-Puerta, M. Hackenberg, A. Pascual-Montano, and P. Carmona-Saez (2022). Functional enrichment analysis of regulatory elements. Biomedicines\/ 10\/ (3), 590

    2022

  8. [16]

    Zhang, J

    Guo, B., C. Zhang, J. Liu, and X. Ma (2019). Improving text classification with weighted word embeddings via a multi-channel T ext CNN model. Neurocomputing\/ 363 , 366--374

    2019

  9. [17]

    Havrlant, L. and V. Kreinovich (2017). A simple probabilistic explanation of term frequency-inverse document frequency (tf-idf) heuristic (and variations motivated by this explanation). International Journal of General Systems\/ 46\/ (1), 27--36

    2017

  10. [18]

    Hiemstra, D. (2000). A probabilistic justification for using tf.idf term weighting in information retrieval. International Journal on Digital Libraries\/ 3\/ (2), 131–139

    2000

  11. [19]

    Huang, D. W., B. T. Sherman, and R. A. Lempicki (2009). Bioinformatics enrichment tools: P aths toward the comprehensive functional analysis of large gene lists. Nucleic Acids Research\/ 37\/ (1), 1--13

    2009

  12. [20]

    Urli, and L

    Incitti, F., F. Urli, and L. Snidaro (2023). Beyond word embeddings: A survey. Information Fusion\/ 89 , 418--436

    2023

  13. [21]

    Elmekki, A

    Islam, S., H. Elmekki, A. Elsebai, J. Bentahar, N. Drawel, G. Rjoub, and W. Pedrycz (2023). A comprehensive survey on applications of transformers for deep learning tasks. Expert Systems with Applications\/ 241 , 122666

    2023

  14. [22]

    Jaioun, K. and K. Teerapabolarn (2014). An improved binomial approximation for the hypergeometric distribution. Applied Mathematical Sciences\/ 8\/ (13), 613--7

    2014

  15. [23]

    Joachims, T. (1997). A probabilistic analysis of the R occhio algorithm with TFIDF for text categorization. In Proceedings of the Fourteenth International Conference on Machine Learning , ICML '97, San Francisco, CA, USA, pp.\ 143–151. Morgan Kaufmann Publishers Inc

    1997

  16. [24]

    Evans, A

    Jumper, J., R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. Z \'i dek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A. Kohl, A. J. Ballard, A. Cowie, B. Romera-Paredes, S. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen, D. Reiman,...

    2021

  17. [25]

    Kim, D., D. Seo, S. Cho, and P. Kang (2019). Multi-co-training for document classification using various document representations: TF -- IDF , LDA , and D oc2 V ec. Information Sciences\/ 477 , 15--29

    2019

  18. [26]

    Kwok, K. (1990). Experiments with a component theory of probabilistic information retrieval based on single terms as document components. ACM Transactions on Information Systems (TOIS)\/ 8\/ (4), 363--386

    1990

  19. [27]

    Li, Q., H. Peng, J. Li, C. Xia, R. Yang, L. Sun, P. S. Yu, and L. He (2022). A survey on text classification: From traditional to deep learning. ACM Transactions on Intelligent Systems and Technology\/ 13\/ (2), 1--41

    2022

  20. [28]

    Luhn, H. P. (1957). A statistical approach to mechanized encoding and searching of literary information. IBM Journal of Research and Development\/ 1\/ (4), 309–317

    1957

  21. [29]

    Ovens, D

    Maleki, F., K. Ovens, D. J. Hogan, and A. J. Kusalik (2020). Gene set analysis: C hallenges, opportunities, and future research. Frontiers in Genetics\/ 11 , 531777

    2020

  22. [30]

    Medvedev, K. E., R. D. Schaeffer, K. S. Chen, and N. V. Grishin (2023). Pan-cancer structurome reveals overrepresentation of beta sandwiches and underrepresentation of alpha helical domains. Scientific Reports\/ 13\/ (1), 11988

    2023

  23. [31]

    Szymczak, T

    Piron, A., F. Szymczak, T. Papadopoulou, M. I. Alvelos, M. Defrance, T. Lenaerts, D. L. Eizirik, and M. Cnop (2024). Red R ibbon: A new rank--rank hypergeometric overlap for gene and transcript expression signatures. Life Science Alliance\/ 7\/ (2), e202302203

    2024

  24. [32]

    Rathi, R. and A. Mustafi (2023). The importance of term weighting in semantic understanding of text: A review of techniques. Multimedia Tools and Applications\/ 82\/ (7), 9761--9783

    2023

  25. [33]

    Stevens, J

    Reusens, M., A. Stevens, J. Tonglet, J. De Smedt, W. Verbeke, S. Vanden Broucke, and B. Baesens (2024). Evaluating text classification: A benchmark study. Expert Systems with Applications\/ 254 , 124302

    2024

  26. [34]

    Personnaz, L

    Rivals, I., L. Personnaz, L. Taing, and M.-C. Potier (2007). Enrichment or depletion of a GO category within a class of genes: W hich test? Bioinformatics\/ 23\/ (4), 401--407

    2007

  27. [35]

    Zaragoza, et al

    Robertson, S., H. Zaragoza, et al. (2009). The probabilistic relevance framework: BM 25 and beyond. Foundations and Trends in Information Retrieval\/ 3\/ (4), 333--389

    2009

  28. [36]

    Robertson, S. E. and K. S. Jones (1976). Relevance weighting of search terms. Journal of the American Society for Information Science\/ 27\/ (3), 129--146

    1976

  29. [37]

    Roelleke, T. (2013). Information Retrieval Models: Foundations and Relationships . Synthesis lectures on information concepts, retrieval, and services. Morgan & Claypool Publishers

    2013

  30. [38]

    Roelleke, T. and J. Wang (2006). A parallel derivation of probabilistic information retrieval models. In Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval , SIGIR ’06, New York, NY, USA, pp.\ 107–114. Associa...

    2006

  31. [39]

    Roelleke, T. and J. Wang (2008). TF-IDF uncovered: A study of theories and probabilities. In Proceedings of the 31st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval , SIGIR '08, New York, NY, USA, pp.\ 435–442. ACM

    2008

  32. [40]

    Salton, G. and C. S. Yang (1973). On the specification of term values in automatic indexing. Journal of Documentation\/ 29\/ (4), 351--372

    1973

  33. [41]

    Sheridan, P. and M. Onsj \"o (2024). The hypergeometric test performs comparably to TF - IDF on standard text analysis tasks. Multimedia Tools and Applications\/ 83\/ (10), 28875--28890

    2024

  34. [42]

    Sp \"a rck Jones, K. (1972). A statistical interpretation of term specificity and its application in retrieval. Journal of Documentation\/ 28 , 11–21

    1972

  35. [43]

    Sunehag, P. (2007). Using two-stage conditional word frequency models to model word burstiness and motivating tf-idf. In Proceedings of the 11th International Conference on Artificial Intelligence and Statistics, 2007 , pp.\ 8--16

    2007

  36. [44]

    Madhavji, and J

    Wahba, Y., N. Madhavji, and J. Steinbacher (2023). Attention is not always what you need: Towards efficient classification of domain-specific text: Case-study: IT support tickets. In Science and Information Conference , pp.\ 1159--1166. Springer

    2023

  37. [45]

    Wu, H. C., R. W. P. Luk, K. F. Wong, and K. L. Kwok (2008). Interpreting TF-IDF term weights as making relevance decisions. ACM Transactions on Information Systems\/ 26\/ (3), 13:1–13:37

    2008

  38. [46]

    Xie, Y. and D. Song (2022). Text classification in shipping industry using unsupervised models and transformer based supervised models. arXiv preprint arXiv:2212.12407

    2022 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.