A Fast and Accurate Dependency Parser using Neural Networks
Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014 · paper · Computational linguistics
Proposes a neural network classifier using a small number of dense features for a greedy, transition-based dependency parser. It improves attachment scores by about 2% on English and Chinese, parsing over 1000 sentences per second at 92.2% unlabeled accuracy on English.
Checked: PDF from aclanthology.org, fetched 5 Oct 2026. Page numbers are those of that PDF.
2 findings
Reference or label In-text citation "arc-standard system (Nivre, 2004)" (p. 2, Sec. 2) has no matching entry in the reference list; the only Nivre-first-authored entry is "Joakim Nivre, Johan Hall, and Jens Nilsson. 2006. Maltparser...".
- p. 2 Section 2, paragraph 2
arc-standard system (Nivre, 2004)
- p. 10 References
Joakim Nivre, Johan Hall, and Jens Nilsson. 2006.
- p. 2 Section 2, paragraph 2
we employ the arc-standard system (Nivre, 2004)
Judge's note: Viewed the C003-a crop and the full reference list on p. 10. A grep of the text layer for 'Nivre' finds the in-text 2004 citation, Zhang and Nivre 2011, Kübler, McDonald and Nivre 2009, and Nivre, Hall and Nilsson 2006. There is no 2004 entry. The list is alphabetical, and Nivre et al. 2006 is followed directly by Socher, so no Nivre 2004 entry sits anywhere else. The 2006 entry is cited separately on p. 6 as 'Nivre et al., 2006' (MaltParser), so the 2004 citation does not refer to it.
Second check: another model agreed
Report a problem with this finding
Reference or label Section 5 says "(Titov and Henderson, 2007) applied Incremental Sigmoid Belief Networks to constituency parsing", but the only Titov and Henderson 2007 entry in the reference list is "Fast and robust multilingual dependency parsing with a generative latent variable model. In EMNLP-CoNLL.", a dependency-parsing paper.
- p. 8 Section 5 Related Work
(Titov and Henderson, 2007) applied Incremental Sigmoid Belief Networks to constituency parsing
- p. 10 References
Fast and robust multilingual dependency parsing with a generative latent variable model. In EMNLP-CoNLL.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that a separate 2007 Titov and Henderson paper on constituent parsing with ISBNs exists and may be the intended citation. Within the paper, the page 8 text (crop C016-a) describes constituency parsing with ISBNs. The reference list on page 10 contains only the EMNLP-CoNLL multilingual dependency-parsing entry under Titov and Henderson 2007. Crop C016-b does not show that entry, so I checked it on the full page 10. The claim and the cited work do not match, which suggests a wrong or missing reference entry. Only the authors can say which work was intended.
Second check: another model agreed
Report a problem with this finding
11 minor (typos and the like; most checked by the judge only)
Wording "our parser achieves around 2% improvement in UAS and LAS on all datasets" (also the abstract's "about 2%... on both English and Chinese"); on CTB test the gain over the arc-standard baseline is 1.2 UAS and 1.2 LAS (Table 6).
- p. 6 Section 4.2 text after Table 4
parser achieves around 2% improvement in UAS and LAS on all datasets, while running about 20 times faster
- p. 7 Table 6
standard 82.4 80.9 82.7 81.2 72
- p. 6 Section 4.2, paragraph after Table 4
parser achieves around 2% improvement in UAS and LAS on all datasets
- p. 1 Abstract
ing an about 2% improvement in unla-
Only the authors can say what is right here.
Judge's note: Tables 4-6 test scores. On PTB:CD and PTB:SD the gains over both baselines are 2.1 to 2.4. On CTB they are 1.2 over arc-standard and 3.6 to 3.7 over arc-eager. The wording overstates the CTB arc-standard comparison.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 4.3 says "cube can achieve 0.8% ∼ 1.2% improvement in UAS over tanh and other functions", but in Figure 4 (left) cube exceeds tanh on CTB by about 1.8 points and identity by about 3.6 to 4 points.
- p. 7 Section 4.3, Cube activation function
cube can achieve 0.8% ∼ 1.2% improvement in UAS over tanh and other functions
- p. 9 Figure 4 left
cube tanh sigmoid identity
Only the authors can say what is right here.
Judge's note: Bar heights were read from the Figure 4 (left) image. The y-axis runs 80 to 85 over about 83 px. On CTB, cube is about 83.6, tanh about 81.8, sigmoid about 82.4 and identity about 80.0. On PTB:CD, cube is about 91.4, tanh about 90.7 and identity about 87.3. The stated 0.8–1.2 range matches cube's gain over the strongest competitor on each dataset: about 0.8 over tanh on PTB:CD and about 1.2 over sigmoid on CTB. It does not match 'over tanh' on CTB (about 1.8) or over identity (about 4). The sentence is therefore ambiguous or understated; it is not a data error. Values are approximate, read by eye.
Second check: another model agreed
Report a problem with this finding
Typo "the use of many feature templates cause a less studied problem" should read "causes".
- p. 1 Introduction
the use of many feature templates cause a less studied problem
Judge's note: The crop shows this text. The grammatical subject is "the use", which is singular.
Second check: none (judge only)
Report a problem with this finding
Typo Table 1 subheading reads "Three-word feaures (8)"; should be "features".
- p. 2 Table 1
Three-word feaures (8)
- p. 8 Section 4.4 bullet list
Feautre 2
- p. 6 footnote 10
sightly more accurate
Judge's note: The C004-a crop shows 'Three-word feaures (8)' in the Table 1 header, so this is not an extraction artefact. One reviewer also raised 'Feautre' and 'sightly'; those defects are judged under C015 and C010, and this cluster covers the 'feaures' typo.
Second check: none (judge only)
Report a problem with this finding
Typo Table 1 word-pair template printed as "s1.wts2.t"; it lacks the ◦ operator and should read s1.wt ◦ s2.t.
- p. 2 Table 1, word-pair features
s1.wts2.t
Judge's note: The crop shows no ◦ between s1.wt and s2.t. Every other template in the table uses ◦.
Second check: none (judge only)
Report a problem with this finding
Typo "where Nl is number of different arc labels" is missing an article ("is the number").
- p. 2 Section 2, after the transition list
where Nl is number of different arc labels
Judge's note: The crop confirms the wording. This is a missing word only; the meaning is unchanged.
Second check: none (judge only)
Report a problem with this finding
Typo The hidden-layer size is defined as d_h in Sec. 3.1, where h denotes the hidden activation vector, but Secs. 4.2 and 4.4 write "hidden layer size h = 200" and "h = 200 learned dense features".
- p. 4 Section 3.1
We map the input layer to a hidden layer with
- p. 6 Section 4.2
hidden layer size h = 200
- p. 8 Section 4.4
Since we currently only have h = 200 learned
- p. 4 Section 3.1
dh nodes through a cube activation function
Judge's note: Sec. 3.1 reads "d_h nodes" and "h = (...)^3". Sec. 4.2 and Sec. 4.4 use h for the size. The notation is inconsistent, but the intent is clear from context.
Second check: none (judge only)
Report a problem with this finding
Typo Footnote 10 reads "known to be sightly more accurate" (should be "slightly"); Section 5 also reads "Temporal Restricted Boltzman Machine" (should be "Boltzmann").
- p. 6 Footnote 10
known to be sightly more accurate
- p. 8 Section 5
Temporal Restricted Boltzman Machine
- p. 6
which is known to be sightly more accurate, but orders of magnitude
Judge's note: The C010-a crop shows footnote 10 on p. 6 with 'sightly more accurate'. The C010-b crop shows 'Boltzman Machine' on p. 8. The reference list (p. 10) spells it 'boltzmann' in the Garg and Henderson 2011 entry. Both are cosmetic misspellings.
Second check: none (judge only)
Report a problem with this finding
Typo Weights listed as W1^w, W1^t, W1^l are said to be reshaped to "d × nt, d × nw, d × nl matrices"; the order should be d × nw, d × nt, d × nl.
- p. 8 Section 4.4, 'What do W1 capture?'
d × nt, d × nw, d × nl matrices
- p. 8 Section 4.4
reshape them to d × nt, d × nw, d × nl matrices
Judge's note: The crop confirms the order printed in Sec. 4.4. The meaning can be recovered from context.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4.4 list item reads "Feautre 2: rc(s1).t, s1.t, b1.t."; should be "Feature 2".
- p. 8 Section 4.4 bullet list
Feautre 2: rc(s1).t, s1.t, b1.t.
- p. 8 Section 4.4 bullet list
Feautre 2
Judge's note: The C015-a crop (p. 8) shows 'Feautre 2' between 'Feature 1' and 'Feature 3'. It is a plain misspelling.
Second check: none (judge only)
Report a problem with this finding
Typo "Temporal Restricted Boltzman Machine" should read "Boltzmann"; the reference list itself has "boltzmann".
- p. 8 Section 5 Related Work
Temporal Restricted Boltzman Machine
Judge's note: The crop of Sec. 5 shows 'Boltzman'. The Garg and Henderson 2011 reference entry spells it 'boltzmann'.
Second check: none (judge only)