GloVe: Global Vectors for Word Representation
Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014 · paper · AI and machine learning
Introduces GloVe, a log-bilinear regression model for word vectors trained on nonzero entries of a word co-occurrence matrix. It combines matrix factorization and local context window approaches, scoring 75% on a word analogy task and outperforming related models on similarity and entity recognition.
Checked: PDF from nlp.stanford.edu, fetched 4 Oct 2026. Page numbers are those of that PDF.
6 findings
Reference or label In-text citation "information retrieval (Manning et al., 2008)" has no matching entry in the reference list.
- p. 1 Introduction, first paragraph
information retrieval (Manning et al., 2008)
- p. 11 References (alphabetical; goes Luong → Mikolov with no Manning entry)
Minh-Thang Luong, Richard Socher, and Christo-pher D Manning. 2013.
- p. 11 References
References
Judge's note: I read the full reference list on pp. 11-12. It goes from Luong et al. 2013 to Mikolov et al. 2013a, and there is no entry with Manning as first author. Manning appears only as a co-author (Luong 2013, Huang 2012, Socher 2013, Wang and Manning 2013).
Second check: another model agreed
Report a problem with this finding
Reference or label Table 2 caption "(i)vLBL results are from (Mnih et al., 2013)" and footnote 9 "Table 1 of (Mnih et al., 2013)" cite a two-author work that the reference list gives as "Andriy Mnih and Koray Kavukcuoglu. 2013".
- p. 6 Table 2 caption
(i)vLBL results are from (Mnih et al., 2013)
- p. 10 Footnote 9
Table 1 of (Mnih et al., 2013)
- p. 11 References
Andriy Mnih and Koray Kavukcuoglu. 2013.
- p. 10 Footnote 9
In Ta-ble 1 of (Mnih et al., 2013)
Judge's note: The only Mnih entry on p11 has two authors. Section 2 (p2) cites it correctly as 'Mnih and Kavukcuoglu (2013)'. The Table 2 caption (p6) and footnote 9 (p10) use 'Mnih et al., 2013', which matches no entry.
Second check: another model agreed
Report a problem with this finding
Reference or label In-text citation "Stanford NER model (Finkel et al., 2005)" has no matching entry in the reference list.
- p. 7 Section 4.1, Named entity recognition
model (Finkel et al., 2005).
- p. 11 References (between Duchi and Finkelstein)
Lev Finkelstein, Evgenly Gabrilovich
- p. 7 Section 4.1, Named entity recognition
model (Finkel et al., 2005)
- p. 11 Finkelstein et al. reference
Evgenly Gabrilovich
Judge's note: The reference list on pp. 11-12 goes from Duchi 2011 to Finkelstein 2001 to Huang 2012; there is no Finkel 2005 entry. The spelling 'Evgenly Gabrilovich' is also printed in the Finkelstein entry. That point rests on outside knowledge (the author's name is spelled Evgeniy) and is not the basis of this verdict.
Second check: another model agreed
Report a problem with this finding
Reference or label "We used 10 negative samples, which we show in Section 4.6 to be a good choice for this corpus" — the negative-sample analysis is in Section 4.7 (Fig. 4), not Section 4.6.
- p. 8 Section 4.2, first paragraph of left column
We used 10 negative samples, which we show in Section 4.6 to be a good choice for this corpus.
- p. 10 Section 4.7 text on Fig. 4
word2vec’s performance actually decreases if the number of negative samples increases beyond about 10.
- p. 8 Section 4.2/4.3 SVD baselines paragraph
which we show in Section 4.6 to be a good choice for this corpus
Judge's note: Section 4.6 (Model Analysis: Run-time) covers only populating X, per-iteration time and a pointer to the Fig. 4 learning curve. The negative-sample discussion ('performance actually decreases if the number of negative samples increases beyond about 10') is in Section 4.7 on p10.
Second check: another model agreed
Report a problem with this finding
Reference or label The reference reads "Omer Levy, Yoav Goldberg, and Israel Ramat-Gan. 2014.", listing a place name as a third author.
- p. 11 References, Levy et al. 2014
Omer Levy, Yoav Goldberg, and Israel Ramat- Gan.
- p. 11 References, Levy et al. 2014
Omer Levy, Yoav Goldberg, and Israel Ramat-Gan. 2014.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: Ramat-Gan, Israel is the location of the authors' institution (Bar-Ilan University), not a person. The page image on p.11 shows the entry exactly as quoted.
Second check: another model agreed
Report a problem with this finding
Reference or label Rohde et al. 2006 is given "Communications of the ACM, 8:627–633", the same journal, volume and pages as the Rubenstein & Goodenough 1965 entry printed directly below it ("8(10):627–633").
- p. 11 References, Rohde et al. 2006
Communications of the ACM, 8:627–633.
- p. 11 References, Rubenstein and Goodenough 1965
munications of the ACM, 8(10):627–633.
- p. 11 References, Rubenstein and Goodenough 1965
Communications of the ACM, 8(10):627–633.
Only the authors can say what is right here.
Judge's note: Checked p.11. The two entries share journal, volume 8 and pages 627–633, yet one is dated 2006 and the other 1965. One volume cannot hold both, so this is an internal inconsistency that points to a copied venue. The correct venue for Rohde et al. cannot be determined from the files.
Second check: another model agreed
Report a problem with this finding
14 minor (typos and the like; most checked by the judge only)
Reference or label Introduction credits "the skip-gram model of Mikolov et al. (2013c)" while Section 2 credits "The skip-gram and continuous bag-of-words (CBOW) models of Mikolov et al. (2013a)"; 2013c is listed as 'Linguistic regularities in continuous space word representations'.
- p. 1 Introduction, paragraph 3
such as the skip-gram model of Mikolov et al. (2013c)
- p. 2 Related Work, Shallow Window-Based Methods
models of Mikolov et al. (2013a) propose a sim-
- p. 11 References
Linguistic regularities in con- tinuous space word representations.
- p. 2 Section 2, Shallow Window-Based Methods
models of Mikolov et al. (2013a) propose a sim-ple single-layer architecture
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that skip-gram was introduced in Mikolov et al. 2013a ('Efficient Estimation...'), not in 2013c. Internally, p1 cites 2013c and p2 cites 2013a for the same model; the p11 reference list confirms 2013c is the HLT-NAACL linguistic-regularities paper. Only the authors can say whether 2013c was intended.
Second check: another model agreed
Report a problem with this finding
Wording Conclusion: "outperforms other models on word analogy, word similarity, and named entity recognition tasks" (the abstract makes a similar claim), yet Table 4 shows HPCA at 88.7 on the CoNLL Test set against GloVe's 88.3.
- p. 10 Section 5 Conclusion, last sentence
outperforms other models on word analogy, word similarity, and named entity recognition tasks
- p. 1 Abstract, last sentence
It also outperforms related models on similarity tasks and named entity recognition.
- p. 9 Table 4, Test column
88.7
Only the authors can say what is right here.
Judge's note: Table 4 (p.9) prints HPCA Test 88.7 in bold and GloVe Test 88.3. Sec. 4.3 (p.8) itself says "except for the CoNLL test set, on which the HPCA method does slightly better". The unqualified wording in the conclusion (p.10) and the abstract overstates the result; the data themselves are fine.
Second check: another model agreed
Report a problem with this finding
Wording Table 2 row "ivLBL 100 1.5B 55.9 50.1 53.2": the total implies a semantic share of about 0.53, while the other rows imply about 0.45.
- p. 6 Table 2, row ivLBL 100 1.5B
55.9 50.1 53.2
Only the authors can say what is right here.
Judge's note: Tot. should be a fixed-weight mean of Sem. and Syn. over the same question set. The GloVe 100, CBOW 300 and other rows give a semantic share of about 0.45. With that share, ivLBL 100 gives about 52.7, and the rounding bounds give at most about 52.8, not 53.2. The values come from Mnih and Kavukcuoglu, so only the authors can say whether a value was transcribed wrongly or a different question subset was used.
Second check: another model agreed
Report a problem with this finding
Typo Subject-verb disagreement: "A number of techniques exist that addresses this shortcoming of HAL" (should be 'address').
- p. 2 Related Work, Matrix Factorization Methods
techniques exist that addresses this shortcoming of
Judge's note: Visible in the p2 image; the relative clause refers to the plural 'techniques'.
Second check: none (judge only)
Report a problem with this finding
Typo Missing word: "the information present the ratio Pik/Pjk in the word vector space" (should be 'present in the ratio').
- p. 3 Section 3, right column top
the information present the ratio Pik/Pjk in the
Judge's note: Confirmed in the p3 image, at the top of the right column.
Second check: none (judge only)
Report a problem with this finding
Typo Table 1 caption: "properties specific to ice, and ... properties specific of steam" (should be 'specific to steam').
- p. 3 Table 1 caption
properties specific of steam
Judge's note: The Table 1 caption image shows the parallel phrase 'specific to ice' followed by 'specific of steam'. This is a preposition slip.
Second check: none (judge only)
Report a problem with this finding
Typo Duplicated word: "aside from the the fact that they attempt to maximize".
- p. 4 Section 3.1, after Eqn. (10)
aside from the the fact that they
- p. 4 Sec 3.1
aside from the the fact
- p. 4
for our purposes, aside from the the fact that they atte
Judge's note: The p4 image after Eqn. (10) shows 'the the' on a single line, so it is not an extraction artefact.
Second check: code check
Report a problem with this finding
Typo The worst-case bound is written "O(|V|^2)" in Section 3.2 but "O(V^2)" at the end of the section, with V defined as the vocabulary size.
- p. 6 Section 3.2, end
much better than the worst case O(V 2)
- p. 5 Section 3.2, first paragraph
O(|V |2)
Judge's note: Seen in the p5 and p6 images. This is a notation inconsistency only; the meaning is clear.
Second check: none (judge only)
Report a problem with this finding
Typo "We train models on CoNLL-03 training data on test on three datasets" (should be 'and test on').
- p. 7 Section 4.1, Named entity recognition
els on CoNLL-03 training data on test on three
- p. 7 Named entity recognition
training data on test on three
- p. 7 Named entity recognition
ConLL-03 testing data
Judge's note: Confirmed in the p7 image. The 'ConLL-03' spelling point is handled under C014.
Second check: none (judge only)
Report a problem with this finding
Typo "1) ConLL-03 testing data" is inconsistent with "CoNLL-03" in the same sentence.
- p. 7 Section 4.1, Named entity recognition
1) ConLL-03 testing data
Judge's note: The p7 image shows 'CoNLL-2003' and 'CoNLL-03' followed by 'ConLL-03' in the same paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo The Apostol 1976 reference repeats its title: "Introduction to Analytic Number Theory. Introduction to Analytic Number Theory.", and no publisher is given.
- p. 11 References, Apostol 1976
Introduction to Analytic Number Theory. Introduction to Analytic Num- ber Theory.
Judge's note: The page image on p.11 shows the title in italics followed by the same title again in roman type. This is a duplicated field. Which publisher should appear (Springer is the reviewer's suggestion) rests on outside knowledge.
Second check: none (judge only)
Report a problem with this finding
Typo The Finkelstein et al. 2001 reference spells a co-author as "Evgenly Gabrilovich".
- p. 11 References, Finkelstein et al. 2001
Evgenly Gabrilovich
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the author's name is spelled Evgeniy Gabrilovich. The page image on p.11 shows "Evgenly".
Second check: none (judge only)
Report a problem with this finding
Typo The Rohde et al. 2006 reference title is misspelled "co-occurence"; the paper spells it "co-occurrence" everywhere else.
- p. 11 References, Rohde et al. 2006
co-occurence.
- p. 11 References, Rohde et al. 2006
co-occurence
- p. 11 References
co-occurence. Communications of the ACM, 8:627–633.
Judge's note: The page image on p.11 shows "co-occurence". The Lund & Burgess and Bullinaria entries on the same page spell it "co-occurrence". Reviewer note 4 (the duplicated venue) is the same defect as C020 and is judged there.
Second check: none (judge only)
Report a problem with this finding
Typo The Tellex et al. 2003 reference has "Development in Informaion Retrieval" where "Information" is meant.
- p. 12 References, Tellex et al. 2003
ment in Informaion Retrieval.
- p. 12 References, Tellex et al. 2003
ment in Informaion Retrieval
- p. 12 Tellex et al. 2003 reference
Development in Informaion Retrieval
Judge's note: The page image on p.12 shows "Informaion", so this is not an extraction artefact.
Second check: none (judge only)