MaltParser: A Data-Driven Parser-Generator for Dependency Parsing
Proceedings of the Language Resources and Evaluation Conference, 2006 · paper · Computational linguistics
Presents MaltParser, a system that builds a dependency parser for a language from a treebank, with selectable parsing algorithms, learning algorithms and user-defined features. It reports dependency accuracy of 80–90% across Swedish, English, Czech, Danish and Bulgarian.
Checked: PDF from aclanthology.org, fetched 5 Oct 2026. Page numbers are those of that PDF.
3 findings
Substance In the Malt-TAB example "an DT 4 DET" and "old JJ 4 NMOD" give head 4, but "old" is itself token 4 (a token as its own head); both should point to "story", token 5. The Malt-XML version repeats it (head="4" for "an" and "old").
- p. 4 Section 6, Malt-TAB example, first sentence
old JJ 4 NMOD
- p. 4 Section 6, Malt-TAB example, first sentence
an DT 4 DET
- p. 4 Section 6, Malt-XML example
<word form="old" postag="JJ" head="4" deprel="NMOD"/>
- p. 4 Malt-TAB example and Malt-XML example
an DT 4 DET / old JJ 4 NMOD
Judge's note: Page 4 image. Tokens of the first sentence in order: This=1, is=2, an=3, old=4, story=5, .=6. The other heads fix the numbering as 1-based with 0 for the root: "is" has head 0 / ROOT, and "This", "story" and "." have head 2 ("is"); the second sentence is consistent with the same scheme, and Section 3 says HEAD[i] = 0 means not attached. Under that numbering "old" with head 4 is its own head, which is not a valid dependency structure, and "an" is attached to the adjective. Under 0-based numbering the other rows would be wrong instead. The same values appear in the XML listing, so the slip is in the example itself.
Second check: another model agreed
Report a problem with this finding
Reference or label The Malt-XML example opens with "<sentence>" immediately followed by "</sentence>", so the six <word> elements of the first sentence lie outside any sentence element; the closing tag belongs after the line for "." and before the second "<sentence>".
- p. 4 Section 6, Malt-XML example, first two lines
<sentence> </sentence>
- p. 4 Section 6, Malt-XML example, first two lines
</sentence>
- p. 4 Section 6, sentence introducing the example
same sentences represented in Malt-XML would look as
- p. 4 Malt-XML listing, between sentences
<sentence>
Judge's note: Page 4 image: the listing's line order is <sentence>, </sentence>, six <word> lines, <sentence>, four <word> lines, </sentence>. This is the printed order, not an extraction reordering. The second sentence is enclosed correctly. The listing is introduced as "The same sentences represented in Malt-XML", so it should show two enclosed sentences. The total count of opening and closing tags is not balanced as intended: the first element is empty and the first sentence's words are unenclosed.
Second check: another model agreed
Report a problem with this finding
Wording Section 6 says that in parsing mode the system "constructs a projective dependency graph for each sentence", while Section 3 says Covington's algorithm can be run in "a non-projective mode, allowing non-projective (but acyclic) dependency structures".
- p. 4 Section 6, parsing mode bullet
constructs a projective dependency graph
- p. 2 Section 3, Covington's algorithm bullet
a non-projective mode, allowing non-projective (but
- p. 2 Section 3, Covington's algorithm bullet
quadratic-time algorithm for unrestricted dependency
- p. 2 Section 3, Covington's algorithm
allowing non-projective (but acyclic) dependency structures
Only the authors can say what is right here.
Judge's note: Page 4 image, first bullet, and page 2 image, Covington bullet in Section 3. Section 3 describes Covington's algorithm as being "for unrestricted dependency structures" and says the CONTEXT stack exists for algorithms that allow non-projective structures; Section 4 repeats that CONTEXT is relevant with Covington's algorithm in non-projective mode. The Section 6 sentence is unqualified and refers to "the same parsing algorithm ... used during learning", so it covers that mode too. No statement elsewhere restricts the output to projective graphs. Probably written with Nivre's algorithm in mind; the wording overstates.
Second check: another model agreed
Report a problem with this finding
9 minor (typos and the like; most checked by the judge only)
Wording The text says "Combining these features gives the following standard model", but the listing contains "POS STACK 1", which is not covered by the bullet "Part-of-speech features for TOP and NEXT, as well as a lookahead of 1–3 tokens".
- p. 3 Section 4, bullet list of useful features
Part-of-speech features for TOP and NEXT, as well as
- p. 3 Section 4, model 7 listing, sixth line
POS STACK 1
- p. 3 Section 4, sentence introducing model 7
Combining these features gives the following standard
Judge's note: Page 3 image. The model 7 listing has six POS lines: STACK, INPUT, INPUT 1, INPUT 2, INPUT 3 and STACK 1. The bullet covers TOP (STACK), NEXT (INPUT) and lookahead (INPUT 1-3), which is five. STACK 1 is the token below TOP, as the paper itself glosses it in the suffix example ("the token immediately below TOP"). The DEP and LEX lines do match their bullets (LEX STACK 0 0 1 is the head of TOP; LEX INPUT 1 is one lookahead token). The bullets are introduced with "In particular", so they can be read as a non-exhaustive summary; hence a minor wording point.
Second check: another model agreed
Report a problem with this finding
Typo Section 1 prints "we need metods that can give good accuracy"; "metods" should be "methods".
- p. 1 Section 1, first paragraph
we need metods that can give
- p. 3 Section 4
more succintly
- p. 3 Section 4
useful for all language investigated so far
- p. 3 Section 4
lexical features be defined as suffix features
Judge's note: Read on the page 1 image, left column, end of the first paragraph: the word is printed "metods". This cluster also bundles "succintly", "all language", "lexical features be defined" and "function,which" on page 3; those are real too but are judged in C005, C006, C007 and C008, so this verdict covers "metods" only.
Second check: none (judge only)
Report a problem with this finding
Typo The abstract prints "a data-driven parser generator for dependency parsing" without a hyphen, while the title ("A Data-Driven Parser-Generator"), Section 2 and the conclusion print "parser-generator".
- p. 1 Abstract
a data-driven parser generator for dependency parsing
- p. 1 Title / Section 2
A Data-Driven Parser-Generator
Judge's note: Page 1 image: the abstract has "parser generator" mid-line, so it is not a line-break effect; the title and Section 2 ("a traditional parser-generator constructs") have the hyphen, as does Section 7 on page 4. A hyphenation inconsistency only; either form is acceptable.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3 prints "(with LC[i] = 0 if i has not left children)" and "(with RC[i] = 0 if i has not right children)"; "not" should be "no", as in the parallel "has no left siblings" and "has no right siblings".
- p. 2 Section 3, bullet for LC
0 if i has not left children
- p. 2 Section 3, bullet for RC
= 0 if i has not right children
- p. 2 Section 3, LC definition
if i has not left children
- p. 2 Section 3, RC definition
if i has not right children
Judge's note: Page 2 image, left column: the LC and RC bullets read "has not left children" and "has not right children", while the LS and RS bullets directly below read "has no left siblings" and "has no right siblings". The meaning is unaffected.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4 prints "these five features can also be specified more succintly:"; "succintly" should be "succinctly".
- p. 3 Section 4, before the truncated feature listing
specified more succintly
- p. 3 Section 4, before second feature list
more succintly:
- p. 3 Section 4
more succintly
Judge's note: Page 3 image, left column, the line above the truncated five-feature listing reads "succintly". C001 also mentions this word, but that cluster is scoped to "metods", so this is not a duplicate.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4 prints "certain features that have proven useful for all language investigated so far"; "language" should be "languages".
- p. 3 Section 4, paragraph before the bullet list
for all language investigated so far
- p. 3 Section 4, paragraph after suffix example
useful for all language investigated so far
- p. 3 Section 4
all language investigated so far
- p. 3 seventh column bullet
function,which
Judge's note: Page 3 image, left column, the paragraph before the bullet list reads "for all language investigated so far (Nivre and Hall, 2005)". The missing space in "function,which", which one reviewer added here, is judged in C008.
Second check: none (judge only)
Report a problem with this finding
Typo Section 4 prints "Alternatively, lexical features be defined as suffix features"; a word is missing ("can be defined").
- p. 3 Section 4, paragraph after model 7
Alternatively, lexical features be defined as suffix features
- p. 3 Section 4, after model 7
lexical features be defined as suffix features
Judge's note: Page 3 image, right column, the paragraph after the model 7 listing. The sentence sits within one line, so no word was lost at a line break or in extraction.
Second check: none (judge only)
Report a problem with this finding
Typo The seventh-column bullet prints "LS or RS function,which can be negative" with no space after the comma.
- p. 3 Section 4, bullet for the seventh column
LS or RS function,which can be negative
Judge's note: Page 3 image, top of the left column: the comma is set tight against "which". The parallel sixth-column bullet on page 2 has "LC or RC function, which can be negative" with the space.
Second check: none (judge only)
Report a problem with this finding
Typo The reference entry "Chih-Chung Chang and Chih-Jen Lin, 2001." has a comma between the authors and the year, where every other entry has a full stop (for example "Michael A. Covington. 2001.").
- p. 4 References
Chih-Chung Chang and Chih-Jen Lin, 2001.
Judge's note: Page 4 image, reference list: the Chang and Lin entry is the only one with a comma before the year; all eleven others use a full stop. Punctuation inconsistency only.
Second check: none (judge only)