Semantic role labeling for eventive nominalizations 
Justin Betteridge and Matthew Bilotti 

Traditional semantic role labeling systems such as ASSERT (Pradhan, 
et. al., 2005) extract noun phrase arguments from sentences and link 
them to their target verb using semantic roles.  Here is an example of 
a hypothetical sentence about a Biology faculty member, Professor 
Smith.  In this case, PropBank (Kingsbury, et. al.) roles are shown, 
since that is the type of roles ASSERT is trained to produce. 

Professor Smith earned her Ph.D. from Prestigious University. 

[ARG0 Professor Smith] [TARGET earned ] [ARG1 her Ph.D.] [ARG2 from 
Prestigious University.] 

What happens if the sentence happened to be written this way instead? 

Prestigious University announced that Professor Smith's name was among 
those selected for conferral of the Ph.D. in Molecular Biology. 

ASSERT and similar systems would analyze this sentence as shown: 

[ARG0 Prestigious University] [TARGET announced ] that [ARG1 
Smith's name was among those selected for conferral of the Ph.D. in 
Molecular Biology.] 

Many semantic role labeling systems, including ASSERT, do not provide 
bracketings when a form of have or be is the target verb.  In 
addition, because the fact that the Ph.D. was conferred upon Smith by 
Prestigious University is expressed using a nominalization and not a 
verb, it is not labeled. 

We propose to build a system that can cover this gap in coverage for 
nominalizations.  For the example above, our system would extract: 

EVENT: conferral 
ARG0: Prestigious University 
ARG1: Ph.D. in Molecular Biology 
ARG2: Smith 

When combined with a traditional semantic role labeler, our system 
will greatly improve coverage of event frames extracted from text. 

Training Procedure: 

We will be training on a collection of chunked noun phrases that 
contain eventive nominals, as identified in Nomlex, extracted from the 
AQUAINT corpus.  We will learn a way to identify substructure in the 
noun phrase, such as by identifying that in the phrase "conferral of 
the Ph.D. in Molecular Biology" the argument "the Ph.D. in Molecular 
Biology" appears as the PP-OBJ of a PP headed by "of".  From this 
structural analysis of noun phrases, we will learn a mapping up into 
Nomlex, and from there, can access the verb frame information for the 
corresponding verb by existing cross-references into VerbNet. 

Once the model has been built: 

We are going to package our learned model as a UIMA AnalysisEngine 
(annotator) which will require a prerequisite, noun-phrase chunking 
(which we will add), which requires sentence breaking (which we have 
already).  Our AnalysisEngine will take noun phrases encountered in 
text (iterate over noun phrase annotations already in the CAS) and 
check to see if they contain an eventive nominalization.  It will 
apply the same structural analysis to the noun phrase, then use the 
learned Nomlex mapping to put the nominal's arguments into their 
correct slots in the corresponding verb frame found in VerbNet.  We 
will add this information as annotations on the CAS and send it along 
to the next AnalysisEngine in the UIMA pipeline. 

Here are more examples of the kind of information the system will be 
able to extract: 

[The appointment of Professor Smith as the chair of the Biology 
Department] was widely publicised in the local newspaper. 

EVENT: appointment 
ARG0: Professor Smith 
ARG1: the chair of the Biology Department 

John told Mary about [the publication of Professor Smith's latest article.] 

EVENT: publication 
ARG0: (unspecified; perhaps a conference or a journal publisher) 
ARG1: Professor Smith's latest article 

[Professor Smith's organization of the International Biological 
Sciences Conference] was appreciated by other members of the field. 

EVENT: organization 
ARG0: Professor Smith 
ARG1: the International Biological Sciences Conference 

Issues yet to be solved: 

Issues with chunking accuracy 
Is a full parse going to be necessary? 
Handling prepositional phrase attachment 
Generalizing from nominals present in Nomlex to other nominals that have 
the same type of frame 
How to compile Nomlex and VerbNet down into a unified resource we can 
use efficiently 

References: 

Collin F. Baker, Charles J. Fillmore, and John B. Lowe. The berkeley 
framenet project. In Proceedings of COLING/ACL '98, pages 86­90, 1998. 

Daniel Gildea and Daniel Jurafsky. Automatic labeling of semantic 
roles. Computational Linguistics, 28(3), 2002. 

Paul Kingsbury, Martha Palmer, and Mitch Marcus. Adding semantic 
annotation to the penn treebank. 

Karin Kippm, Hoa Trang Dang, and Martha Palmer. Class-based 
construction of a verb lexicon. 

Catherine Macleod, Ralph Grishman, Adam Meyers, Leslie Barrett, and 
Ruth Reeves. Nomlex: A lexicon of nominalizations. 

Sameer Pradhan, Kadri Hacioglu, Valeri Krugler, Wayne Ward, James 
H. Martin, and Daniel Jurafsky. Support vector learning for semantic 
argument classification. Machine Learning, 60(1):11­39, 2005. 

Sameer Pradhan, Honglin Sun, Wayne Ward, James H. Martin, and Daniel 
Jurafsky. Parsing arguments of nominalizations in english and chinese. 

Mihai Surdeanu, Sanda M. Harabagiu, John Williams, and John 
Aarseth. Using predicate argument structures for information 
extraction. 

Robert S. Swier and Suzanne Stevenson. Exploiting a verb lexicon in 
automatic semantic role labelling. 
File: 10-709-nominalizations.txt