CopestakeTalkAbstrat < Main

Tags: view all tags
Ann Copestake<br /> University of Cambridge<br /> <br /> Wednesday, 5/13 1-2pm<br /> EEB 403

<br /> Slacker semantics: why superficiality, dependency and avoidance of<br /> commitment can be the right way to go.<br /> <br /> Abstract<br /> <br /> Standard linguistic approaches to compositional semantics need<br /> adaptation for use in broad-coverage computational processing. I'll<br /> argue that an ideal compositional semantic representation should<br /> capture all and only the information available from syntax and<br /> productive morphology, allowing for shallow as well as deep syntactic<br /> analysis and avoiding over-commitment when ambiguity cannot be<br /> resolved. This leads to `surfacy' semantic representations which must<br /> nevertheless allow for enrichment by deeper analysis (including<br /> lexical semantics and anaphora resolution), support (robust) inference<br /> and have a logically-sound interpretation. Other desiderata are<br /> compatibility with generation, statistical ranking of<br /> analyses/realisations and distributional semantics, ease of use for<br /> consumers of the representation and for human annotators, and<br /> cross-linguistic adequacy.<br /> <br /> Several current approaches to computational compositional semantics<br /> meet at least some of these goals, but in this talk I'll concentrate<br /> on work within the DELPH-IN community using Minimal Recursion<br /> Semantics (MRS) and Robust MRS (RMRS). I'll give a brief overview and<br /> I'll then try to illustrate how the slacker perspective applies in the<br /> grammars by talking about argument labelling (in contrast to semantic<br /> role assignment). I'll then describe some new work on formalism and<br /> show that, under certain assumptions about the grammar, an<br /> interconversion is possible between MRS and a variable-free semantic<br /> dependency notation (DMRS), which removes redundancy, supports<br /> additional forms of underspecification and makes the representation<br /> easier for humans and machine learning algorithms to process. Like<br /> RMRS, DMRS output can be produced from fast, shallow(er) parsers<br /> technology, and if time permits, I'll briefly discuss planned work on<br /> its application to Web-scale parsing.

 -- Main.ebender - 14 May 2009
   * [[%ATTACHURL%/eaclslides.pdf][eaclslides.pdf]]: Slides from Ann Copestake's talk
Attachments
Topic attachments
I	Attachment	History	Action	Size	Date	Who	Comment
pdf	eaclslides.pdf	r1	manage	867.9 K	2009-05-14 - 22:04	UnknownUser	Slides from Ann Copestake's talk
Topic revision: r1 - 2009-05-14 - ebender