===== Beyond NomBank: A Study of Implicit Arguments for Nominal Predicates =====
Matthew Gerber and Joyce Y. Chai
[[http://aclweb.org/anthology-new/P/P10/P10-1160.pdf]]

===== Comments =====

  * When reading the paper it is beneficial to have some basic knowledge about [[http://nlp.cs.nyu.edu/meyers/NomBank.html|NomBank]] and [[http://en.wikipedia.org/wiki/PropBank|PropBank]] (and don't be confused by other resources: VerbNet, FrameNet, WordNet,...).

  * The paper (& PropBank & NomBank) is about **valency** and **dependency structure** of a sentence, although the paper does not use these terms.

  * Although it is not crucial for understanding the paper, we are not sure about terminology of SRL (Semantic Role Labeling), for example in a sentence //"Companies produce paper for customers."// (predicate //produce// in PropBank) or "There are three companies. Their production of paper for customers is growing." (predicate //production// in NomBank):

^ A | companies | paper | customers   |
^ B | arg0      | arg1  | arg2        |
^ C | Agent     | Theme | Beneficiary |
The question is how to call A, B, C using terms **(theta) role, argument, argument position?** My tip is A or B=argument, C=role. What about your tips?

  * Note that word "case" has in linguistics two meanings (not speaking of "test case" meaning): a) **morphological case**, i.g. nominative, genitive, dative and b) **semantic case** aka theta role, thematic case, deep case. We think that the second paragraph of section 2 (case-marked expressions in Japanese) speaks about a), although Charles Fillmore is mentioned in the next paragraph.

  * Section 3.1 states that "We limited our attention to nominal predicates with unambiguous role sets". At first glance, I considered this as too restricting given that most frequent predicates are ambiguous. Now, I have discovered that just 595 out of 4705 predicates in NomBank are considered ambiguous (e.g. way.01="ability-with-agent/attribute", way.02="path", way.03="issue",... way.13="idiom/nomadvlike-backward"). Well,...

  * The description of feature 10 in Table 2 ("Head word of //p//'s right sibling node") is unclear. First, according to examples //price:drop// and //price:index//, there is missing "//p//:" in the description. Second, what does "Head word of //p//'s right sibling node" mean? We thought, siblings should have same head word. 

===== What do we like about the paper =====

  * clever way of exploiting coreference chains
  * [[http://links.cse.msu.edu:8000/lair/projects/semanticrole.html|link]] to annotations and the data needed to replicate the experimets
  * considering IAA (inter annotator agreement) and measuring [[http://en.wikipedia.org/wiki/Cohen%27s_kappa|Cohen's kappa]]

Written by --- //[[popel@ufal.mff.cuni.cz|Martin Popel]] 2010/10/14 14:38//