G22.2590 - Natural Language Processing - Spring 2008 Prof. Grishman

Lecture 11 Outline

April 17, 2008

Syntax-Driven Semantic Analysis  (J&M Chapter 15))

Complete oour discussion of semantic representations (Lecture 10 notes)

Review examples of logical forms

Discuss rules for producing semantic representations of sentences using syntax-driven semantic analysis

Scoping resolution:  converting quasi-logical form to logical form (with conventional quantifier scope)

Factors in resolving scope:

Strategies for data base query applications

Assuming queries are full grammatical sentences: (we will consider how to handle fragmentary queries as part of discourse processing)

A Canonical Representation?  (Lexical Semantics - J&M 16.1 - 16.2)

Logical form brings us a step closer to a 'canonical representation', where there is one representation for a meaning (and one meaning for a representation).  For example, active and passive forms will translate to the same logical form.

The predicates in our logical form are words, which is not satisfactory:
OntoNotes also incorporates the TreeBank annotation as well as the PropositionBank ("PropBank"), which provides standardized 'frames' (lists of arguments) for each verb, as we would expect a good logical form to do.

But this far from exhausts the range of paraphrases (different ways of conveying the same meaning).  Beyond synonyms (words with the same meaning) and verb alternations (different argument structures with the same meaning), there are many cases in which there are larger semantically equivalent structures.  Capturing these is a fundamental challenge of NLP.

For narrow technical domains it seems feasible to collect them by hand, but for broader domains ('political news') the set is so vast that we have to look towards unsupervised discovery procedures, as have been used for binary relations between names.  If we know that two texts present the same material (e.g., two translations of the same text into the same languages;  two news stories about the same topic from the same day), we can try to align the texts and find paraphrases.  This approach has met with some limited success, but building larger inventories of paraphrases remains a very difficult problem for NLP.