This is an old revision of the document!
Table of Contents
Unsupervised methods for head assignments, EACL 2009
Federico Sangati and Willem Zuidema
Presented by: Zdeněk Žabokrtský
Report by: Eduard Bejček
Introduction
- The paper describes two methods for unsupervised head assignment and tests them on two corpora (English PennTB/PARC700 and German Tiger/TigerDB) by two means (against gold standard and using it in a constituency parser).
- It is the first successful unsupervised head assignment.
- There is a problem when converting dependency structures into head annotations: their algorithm doesn't guarantee exactly one head assigned among siblings.
What do we dislike about the paper
- Equation 5 is misleading/wrong: entropy is here computed for the distribution p which sums up to <latex>|\mathcal{L}|</latex>.
- We found three possible interpretations of “any” in Section 3.3, second paragraph; their approach is therefore a bit unclear. The interpretations are:
- apply one fixed annotation of the heads (but it can be any annotation)
- apply the (partial) annotation which is consistent with all possible head annotations
- apply all possible head annotations
- It is also unclear, what “transformed to their respective spines” means exactly; Section 3.4, first item in the itemize.
- The greediness of the entropy model could be its main weakness resulting in not so good numbers.
What do we like about the paper
- Things mentioned in the Introduction.
- The correspondence between rule based head-assignment (by Magerman-Collins) and depencencies (in PARC 700 Depencency Bank) is surprisingly low: 85%.