Statistical Phrase-Based Translation
2003 · paper · Computational linguistics
Proposes a phrase-based translation model and decoder to compare ways of learning phrase translations. It finds simple methods work best: heuristics from word alignments and lexical weighting. Phrases longer than three words add little, and syntactically restricted phrases hurt performance.
Checked: PDF from aclanthology.org, fetched 5 Oct 2026. Page numbers are those of that PDF.
1 finding
Reference or label Reference "Seymore, K. and Rosenfeld, R. (1997). Statistical language modeling using the CMU-Cambridge toolkit. In Proceedings of Eurospeech" (cited in Section 2.1 as [Seymore and Rosenfeld, 1997]) appears to pair the wrong authors with this title.
- p. 7 References
Seymore, K. and Rosenfeld, R. (1997)
- p. 2 Section 2.1, last paragraph
trigram language model [Seymore and Rosenfeld, 1997]
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the Eurospeech 1997 paper 'Statistical language modeling using the CMU-Cambridge toolkit' is by Clarkson and Rosenfeld, not Seymore and Rosenfeld. The page 7 image shows the entry exactly as quoted, and the text layer of page 2 has the matching in-text citation, so the paper is internally consistent. The mismatch cannot be verified from the files; the authors should say whether the authors or the title and venue are the wrong part.
Second check: another model agreed
Report a problem with this finding
17 minor (typos and the like; most checked by the judge only)
Wording Introduction: "what constitutes the best heuristic differs from language pair to language pair and varies with the size of the training corpus"; the paper reports the heuristic comparison (Section 4.5, Figure 5) only for German-English, and Table 3 has no heuristic comparison.
- p. 1 Section 1, last paragraph
what constitutes the best heuristic differs from language pair to language pair and varies with the size of
- p. 4 Section 4, second paragraph
In all experiments in Section 4.1-4.6 we translate from German to English
- p. 7 Table 3 and Section 4.7
Table 3 displays some of the results
Only the authors can say what is right here.
Judge's note: Page 1 image shows the sentence. From the text layer: Section 4 says experiments in 4.1-4.6 are German to English; Section 4.5 and Figure 5 compare heuristics by corpus size only; Table 3 has columns Model4, Phrase and Lex only; Section 4.7 says Table 3 shows "some of the results". The language-pair part of the claim is therefore not traceable to reported data, though it is not contradicted. Only the authors can say whether unreported experiments support it. Pages 4 and 7 were checked in the text layer, not as images.
Second check: another model agreed
Report a problem with this finding
Wording Section 2.2 uses n for two quantities: "we only keep a beam of the best n hypotheses" and "only n(n+1)/2 such sequences for a foreign input sentence of length n".
- p. 2 Section 2.2, right column, pruning paragraph
we only keep a beam of the best n hypotheses
- p. 2 Section 2.2, right column, future cost paragraph
foreign input sentence of length n, we can pre-compute
Judge's note: Both uses are on the page 2 image, right column, typeset as the same math variable. The complexity sentence that follows treats beam size and sentence length as separate quantities. Meaning is recoverable from context, so this is a notation overload only.
Second check: another model agreed
Report a problem with this finding
Wording Section 4.4 extended model prints "φ(f̄_i|ē_i) d(a_i − b_{i−1}) p_w(f̄_i|ē_i, a)^λ": a_i is a phrase start position (Section 2.1) while bare a is a word alignment, and a is left unbound although the text has just defined p_w(f̄|ē) = max_a p_w(f̄|ē, a).
- p. 2 Section 2.1, distortion model
denotes the start position of the foreign
- p. 5 Section 4.4, definition of lexical weight
and a word alignment a be-
- p. 5 Section 4.4, extended model equation (image: d(a_i − b_{i−1}) p_w(f̄_i|ē_i, a)^λ)
extended to
- p. 5 Section 4.4, equation for multiple alignments
we use the one with the highest lexical weight
Judge's note: Page 2 image defines a_i as the start position of the foreign phrase. Page 5 image defines a as a word alignment, gives the max_a form, and then prints the extended model with p_w(f̄_i|ē_i, a)^λ. Both the symbol clash and the unbound a are present. The intended meaning (the best alignment) is recoverable, so this is a notation inconsistency and not a wrong formula.
Second check: another model agreed
Report a problem with this finding
Wording "The syntactic restriction eliminates over 80% of all phrase pairs" is unqualified, but in the 10k column of Table 1 (AP 84k, Syn 19k) the reduction is about 77%; the other five columns are 82-88%.
- p. 4 Section 4.1, paragraph on Table 1
striction eliminates over 80% of all phrase pairs
- p. 4 Table 1, 10k column (AP / Syn rows)
84k
- p. 4 Table 1, Syn row, 10k column
19k
- p. 4 Section 4.1, after Table 1
The syntactic re-striction eliminates over 80% of all phrase pairs.
Judge's note: Read Table 1 and the Section 4.1 paragraph on the page 4 image. AP row: 84k, 176k, 370k, 736k, 1536k, 3152k; Syn row: 19k, 24k, 67k, 105k, 217k, 373k. Only the 10k column falls below 80%, and rounding to thousands cannot lift it above 80% (18.5/84.5 still gives about 78%). Pooled over all six sizes the reduction is about 87%, so the sentence holds as a general summary and fails for one column. I found no qualifying statement in the text. Minor overgeneralisation in wording.
Second check: another model agreed
Report a problem with this finding
Wording "The number grows almost linearly with the training corpus size" is said of each method, but the Syn row of Table 1 goes 19k, 24k, 67k (x1.26 then x2.79 per doubling); the 24k entry is out of line with the rest of the row.
- p. 4 Section 4.1, paragraph on Table 1
ber grows almost linearly with the training corpus size
- p. 4 Table 1, Syn row, 20k column
24k
Only the authors can say what is right here.
Judge's note: Read Table 1 on the page 4 image. Syn row: 19k, 24k, 67k, 105k, 217k, 373k. The AP row doubles steadily (about 2.0-2.1 per doubling) and the Joint row grows by about 1.76-1.82 per doubling, both near-linear. The Syn row is irregular at 10k to 20k to 40k. Either the 24k entry is a misprint or the linearity statement does not apply well to Syn; only the authors can say which. The claim is loose ('almost'), so this is a minor query.
Second check: another model agreed
Report a problem with this finding
Typo Introduction, first sentence: "Various researchers have improved the quality of statistical machine translation system with the use of phrase translation" has singular "system" where "systems" is needed.
- p. 1 Section 1, first sentence
tical machine translation system with the use of phrase
Judge's note: Read on the page 1 image. The singular is printed with no article; the same paragraph later uses "translation systems". Grammatical slip only.
Second check: none (judge only)
Report a problem with this finding
Typo Spurious comma before a "that" clause: "Our experiments also show, that small phrases of up to three words are sufficient"; the same pattern recurs in "conclude from this, that" and "suspect, that".
- p. 1 Introduction
periments also show, that small phrases
Judge's note: Seen on the page 1 image. The other two instances are in the text layer at Sections 4.2 and 4.6. Punctuation only, no effect on meaning.
Second check: none (judge only)
Report a problem with this finding
Typo Section 2.2: "The beam size, e.g. the maximum number of hypotheses in each stack, is fixed to a certain number" uses "e.g." where "i.e." is meant.
- p. 2 Section 2.2, right column
The beam size, e.g. the maximum number of hypothe-
Judge's note: Seen on the page 2 image. The clause defines the beam size; it is not an example of it. Elsewhere (page 3) the paper uses "i.e." correctly.
Second check: none (judge only)
Report a problem with this finding
Typo Section 2.2: "translating 1755 sentence of length 5-15 words" should read "1755 sentences".
- p. 2 Section 2.2, last paragraph
translating 1755 sentence of length 5-15 words
- p. 2 Section 2.2, last paragraph
With larger beams sizes,
Judge's note: Seen on the page 2 image. Section 4 has "1755 sentences of length 5-15" in the text layer. The "beams sizes" slip also mentioned in this cluster is handled under C010.
Second check: none (judge only)
Report a problem with this finding
Typo Section 2.2: "With larger beams sizes, only few sentences are translated differently" should read "larger beam sizes".
- p. 2 Section 2.2, last paragraph
With larger beams sizes,
Judge's note: Seen on the page 2 image. The paper writes "beam size" everywhere else in the same paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.2: "comes from recent efforts to built syntactic translation models" should read "to build".
- p. 3 Section 3.2, second paragraph
phrases comes from recent efforts to built syntactic trans-
- p. 3 Section 3.2, second paragraph
recent efforts to built syntactic trans-
Judge's note: Seen on the page 3 image, right column.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.3: "generates both the Source and Target sentences in a parallel corpus" capitalises "Source" and "Target" mid-sentence.
- p. 3 Sec. 3.3
generates both the Source and Target sentences
Judge's note: Seen on the page 3 image. The capitals are printed as quoted. This is a style inconsistency with no effect on meaning.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3 announces "three different methods", then has "As the first method" and "The third method for comparison is the joint phrase model"; the syntactic-phrase paragraph between them is never labelled as the second method.
- p. 3 Sec. 3, paragraphs 2-4
The third method for comparison is the joint phrase
Judge's note: Seen on the page 3 image. The paragraph beginning "A number of researchers have proposed" carries no ordinal. The numbering is recoverable from context, so this is cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Reference "Schmidt, H. and Schulte im Walde, S. (2000). Robust German noun chunking with a probabilistic context-free grammar" and both in-text citations "Schmidt and Schulte im Walde, 2000" appear to misspell the first author's surname (Schmid).
- p. 7 References
Schmidt, H. and Schulte im Walde, S. (2000)
- p. 3 Section 3 and Section 3.2
Schmidt and Schulte im Walde,
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the first author of the COLING 2000 paper is Helmut Schmid, not Schmidt. The page 7 image shows 'Schmidt, H.' in the reference list, and the text layer shows 'Schmidt' in both in-text citations on page 3. The spelling is consistent within the paper, so this is not an internal inconsistency and cannot be verified from the files.
Second check: none (judge only)
Report a problem with this finding
Typo "Even the largest models take up only a few hundred megabyte of RAM": 'megabyte' should be plural, 'megabytes'.
- p. 4 Sec. 4.1
only a few hundred megabyte of
Judge's note: Read on the page 4 image, Section 4.1, last paragraph: 'only a few hundred megabyte of RAM'. Singular unit after 'a few hundred'; cosmetic grammatical slip.
Second check: none (judge only)
Report a problem with this finding
Typo "We use the lexical weight p_w during translation as a additional factor": should read 'an additional factor'.
- p. 5 Section 4.4, right column
during translation as a
- p. 5 Sec. 4.4, after lexical weight definition
during translation as a additional factor
Judge's note: Read on the page 5 image, right column of Section 4.4: the line ends 'during translation as a' and the next begins 'additional factor'. Wrong article; cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Reference Brown et al. (1993) prints "Computational Linguistics, 19(2):263–313"; the end page appears to be wrong (311).
- p. 7 References, first entry
Computational Linguistics, 19(2):263–313.
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the standard citation for 'The mathematics of statistical machine translation' is Computational Linguistics 19(2):263-311. The page 7 image shows '19(2):263–313' as quoted. The true page range cannot be verified from the files.
Second check: none (judge only)