[Probcogsci] Chris Brew talk Wednesday 2 December, 2pm-3:30pm AP&M 2351

[Roger Levy]

Chris Brew is visiting from Ohio State and giving a talk in my lab  
meeting tomorrow about computational lexical semantics.  All are  
welcome!

***

Adventures in Computational Verb Semantics

Lexical semantics is the subfield of linguistics concerned with  
describing the denotations of words -- i.e., how and what words refer  
to in some possible world.  Semantic role labeling, on the other hand,  
is concerned with identifying and classifying the participants in a  
process described by some predicate (usually a verb), without  
necessarily supplying an informative label as to what the predicate  
and participants denote.  For example, a lexical semantic analyzer  
might label an instance of 'give' as a TRANSFER verb, whereas a  
semantic role labeler will identify the verb's subject as AGENT and  
its objects as PATIENT and THEME, without saying what sort of event  
they are agents, patients and themes in (viz., a transfer of  
something, by someone, to someone else).
In this talk we summarize our recent work in verbal lexical semantics  
and semantic role labeling, and sketch how these two views on verb  
semantics can be combined.

Our approach to verb lexical semantics is based on Levin's (1993) hand- 
crafted verb taxonomy.  Here we exploit the close connection between  
syntax and semantics in Levin's taxonomy to bootstrap a large corpus  
of labeled verb instances using a syntactic parser.   Our results show  
that a reasonably accurate lexical semantic analyzer can be trained  
without explicitly hand-labeled data.

Our verb semantic role labeler, Brutus, is trained on a merged version  
of two hand-annotated resources: the CCGbank corpus (Hockenmeier and  
Steedman, 2005) of Combinatory Categorial Grammar (CCG; Steedman,  
2000) parses and the Propbank corpus of semantic roles (Palmer et al.,  
2005), yielding semantic role labeling performance on a par with the  
state of the art (Boxwell et al., 2009).

We conclude our talk with some (very) recent efforts to combine these  
two approaches by using the output of our Levin-based classifier as a  
source of information for Brutus.  Preliminary results show a trend of  
improvement in Brutus's role-labeling performance, which suggests that  
large amounts of bootstrapped data can improve semantic role labeling  
performance in a supervised learning setting, even across semantic  
frameworks.

--

Roger Levy                      Email: rlevy at ling.ucsd.edu
Assistant Professor             Phone: 858-534-7219
Department of Linguistics       Fax:   858-534-4789
UC San Diego                    Web:   http://ling.ucsd.edu/~rlevy