Sequence to Sequence Learning with Neural Networks
arXiv (Cornell University), 2014 · paper · AI and machine learning
Presents an approach to mapping sequences to sequences using one multilayer LSTM to encode the input and another to decode the output. On English-to-French translation it scored 34.8 BLEU versus 33.3 for a phrase-based system; reversing source word order helped.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
1 finding
Reference or label Introduction names the technique of [11] as "The Connectionist Sequence Classification", whereas reference [11] is titled "Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks".
- p. 2 Introduction, third paragraph
The Connectionist Sequence Classification is another popular
- p. 9 Reference [11]
Connectionist temporal classification: labelling
Judge's note: Page 2 image shows "The Connectionist Sequence Classification is another popular technique ... [11]". Page 9 image shows reference [11] (Graves et al., ICML 2006) titled "Connectionist temporal classification". The capitalised name in the body does not match the paper's own reference entry, so no outside knowledge is needed.
Second check: another model agreed
Report a problem with this finding
17 minor (typos and the like; most checked by the judge only)
Wording Figure 1 caption says "Note that the LSTM reads the input sentence in reverse", but the drawing feeds A, B, C in forward order and the Conclusion cites it as "the non-reversed translation problem (shown in fig. 1)".
- p. 2 Figure 1 caption
Note that the LSTM reads the input sentence in reverse
- p. 2 Figure 1 caption
Our model reads an input sentence “ABC”
- p. 8 Conclusion, second paragraph
the non-reversed translation problem (shown in fig. 1)
- p. 8 Conclusion, second paragraph
on the non-reversed translation problem (shown in fig. 1)
Only the authors can say what is right here.
Judge's note: On the page 2 image the Figure 1 diagram has inputs A, B, C, <EOS> left to right for an input sentence "ABC", which is the forward order. The caption's note says the LSTM reads the input in reverse. On the page 8 image the Conclusion refers to fig. 1 as showing the non-reversed problem. The caption's note may be describing the actual system and not the drawing, but as printed the two descriptions of the figure disagree. Only the authors can say which was intended.
Second check: another model agreed
Report a problem with this finding
Wording Section 3.3 says "LSTMs trained on reversed source sentences did much better on long sentences than LSTMs trained on the raw source sentences (see sec. 3.7)", but Section 3.7 and Figure 3 compare only the LSTM (34.8) with the SMT baseline (33.3).
- p. 5 Section 3.3, last sentence (continued from page 4)
trained on the raw source sentences (see sec. 3.7)
- p. 6 Section 3.7
which is shown quantitatively in figure 3
- p. 4 Section 3.3, last paragraph (continues on PDF page 5)
LSTMs trained on reversed source sentences did much better on long sentences than LSTMs
- p. 6 Section 3.7
the LSTM did well on long sentences, which is shown quantita-
Only the authors can say what is right here.
Judge's note: Pages 4-5 images: the claim and the pointer to sec. 3.7 are as quoted. Page 6 image: Section 3.7 is two sentences pointing to Figure 3 and Table 3. Page 7 image: Figure 3 plots only "LSTM (34.8)" against "baseline (33.3)", and Table 3 gives example translations. No by-length result for a forward (raw-source) LSTM appears anywhere in the paper. The pointer does reach the long-sentence section, but the cited section does not show the reversed-versus-raw comparison the sentence claims.
Second check: another model agreed
Report a problem with this finding
Typo Missing articles in the Introduction: "to obtain large fixed-dimensional vector representation" and "the first to map the entire input sentence to vector".
- p. 2 Introduction, second paragraph on page
to obtain large fixed-
- p. 2 Introduction, paragraph 'There have been a number of related attempts'
map the entire input sentence to vector
Judge's note: Both phrases are on the page 2 image as quoted; the article 'a' is missing in each. Cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3: "without using a reference SMT system and we it to rescore the n-best lists of an SMT baseline" has a missing verb (presumably 'used').
- p. 3 Section 3, first paragraph
and we it to rescore the
- p. 3
without using a reference SMT system and we it to rescore the
Judge's note: The page 3 image shows "and we it to rescore the" exactly as quoted. It is not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Dropped words: "alignment between the inputs the outputs" (p3), "increases the number model parameters" (p3), "divided it the size of the batch" (p5), footnote "There several variants of the BLEU score" (p5), "in way that produces smooth translations" (p8).
- p. 3 Section 2, paragraph after the RNN equations
alignment between the inputs the
- p. 5 Section 3.4, third bullet
divided it the size of the batch
- p. 5 Footnote 1
There several variants of the BLEU score
- p. 3 Section 2, paragraph 'Our actual models'
so increases the number model parameters
Judge's note: All five phrases appear as quoted on the page 3, 5 and 8 images. The missing words are 'and', 'of', 'by', 'are' and 'a' respectively. Cosmetic copy-editing slips, bundled here; C008, C010 and C014 repeat individual items.
Second check: none (judge only)
Report a problem with this finding
Typo End of Introduction (pages 2-3): "while different sentences meanings will be far" is ungrammatical.
- p. 3 Top of page (end of Introduction)
sentences meanings will be far
Judge's note: Page 2 ends with "while different" and the page 3 image begins "sentences meanings will be far." A possessive is missing or the words are misordered, and 'apart' may also be missing. The meaning is recoverable.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.4: "After 5 epochs, we begun halving the learning rate every half epoch" should read 'began'.
- p. 5 Section 3.4, second bullet
After 5 epochs, we begun halving
Judge's note: The page 5 image shows "we begun halving" as quoted. Wrong verb form; cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.5 reads "on a single GPU processes a speed of approximately 1,700 words per second"; the phrase "processes a speed of" is garbled (e.g. "achieves a speed of" or "processes approximately").
- p. 5 Section 3.5, first sentence
processes a speed of approximately 1,700 words per second
- p. 5 Section 3.5, first sentence
gle GPU processes a speed of approximately 1,700 words per second
- p. 5 Section 3.5
processes a speed of approximately
Judge's note: Read on the page 5 image: the sentence is printed exactly as quoted. Later in the same paragraph the paper writes "achieved a speed of 6,300", which shows the intended construction.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.5 reads "Training took about a ten days with this implementation"; there is a stray article "a" before "ten days".
- p. 5 Section 3.5, last sentence
Training took about a ten
- p. 5 footnote 1 and Section 3.5
There several variants of the BLEU score
- p. 5 Section 3.5
Training took about a ten days
Judge's note: Read on the page 5 image; the text is printed as quoted, with a line break after "a ten". One reviewer note also raises the footnote's missing verb; that is the same defect as C019 and is judged there.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.6 reads "This way of evaluating the BELU score is consistent with [5] and [2]"; "BELU" should be "BLEU".
- p. 5 Section 3.6, first paragraph
of evaluating the BELU score
- p. 5 Section 3.6, first paragraph
of evaluating the BELU score is consistent with [5] and [2]
- p. 5 Section 3.6
evaluating the BELU score
- p. 5
evaluating the BELU score is consistent with [5]
Judge's note: Read on the page 5 image: "BELU" is printed. The same paragraph, the footnote and the table headers all spell it "BLEU".
Second check: none (judge only)
Report a problem with this finding
Typo Footnote 1 reads "There several variants of the BLEU score, and each variant is defined with a perl script."; the verb "are" is missing.
- p. 5 Footnote 1
There several variants of the BLEU score
Judge's note: Read on the page 5 image; the footnote is printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.6 prints the web address twice as "statmt.org\matrix" with a backslash; a URL path separator is a forward slash ("statmt.org/matrix").
- p. 5 Section 3.6, first paragraph (appears twice)
statmt.org\matrix
Judge's note: Read on the page 5 image: both occurrences are typeset in monospace with a backslash, so this is not an extraction artefact. Cosmetic; the intended address is evident.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.4 reads "After 5 epochs, we begun halving the learning rate every half epoch"; the simple past should be "began".
- p. 5 Section 3.4
we begun halving the learning rate
Judge's note: Read on the page 5 image; the text is printed as quoted. It is a grammatical slip (past participle used as simple past).
Second check: none (judge only)
Report a problem with this finding
Typo Table 1 caption reads "an ensemble of 5 LSTMs with a beam of size 2 is cheaper than of a single LSTM with a beam of size 12"; there is a stray "of" (or a missing "that").
- p. 6 Table 1 caption
is cheaper than of a single LSTM
Judge's note: Read on the page 6 image; the caption is printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4 reads "by translating pieces of the source sentence in way that produces smooth translations"; the article "a" is missing ("in a way").
- p. 8 Section 4, third paragraph
source sentence in way that produces
Judge's note: Read on the page 8 image; the text is printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4 cites "Hermann et al. [12]", but reference [12] lists only two authors, "K. M. Hermann and P. Blunsom"; it should read "Hermann and Blunsom [12]".
- p. 8 Section 4, fourth paragraph
Hermann et al. [12]
- p. 9 Reference [12]
K. M. Hermann and P. Blunsom.
- p. 8 Related work, last paragraph
End-to-end training is also the focus of Hermann et al. [12]
Judge's note: Page 8 image shows "End-to-end training is also the focus of Hermann et al. [12]". Page 9 image shows reference [12] as "K. M. Hermann and P. Blunsom. Multilingual distributed representations without word alignment. In ICLR, 2014." The paper writes its other two-author work in full as "Kalchbrenner and Blunsom [18]". The citation points to the right entry, so this is cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo The reference list names a co-author as "B. Merrienboer" in [5] but "B. van Merrienboer" in [26]; if this is the same person, the surname particle "van" is missing in [5].
- p. 9 Reference [5]
B. Merrienboer
- p. 9 Reference [26]
B. van Merrienboer
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that the co-author of the two cited works is the same person and that the surname is "van Merrienboer". Both forms are printed as quoted on the page 9 image. The files show the inconsistency between the two entries but cannot establish that they refer to one person.
Second check: none (judge only)