semantics in nlp (part 2) mas.s60 rob speer catherine havasi * lots of slides borrowed for lots of...

11
Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! S

Upload: kory-nash

Post on 16-Dec-2015

215 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

Semantics in NLP(part 2)

MAS.S60Rob Speer

Catherine Havasi

* Lots of slides borrowed for lots of sources! See end.

Page 2: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

Are people doing logic?

• Language Log: “Russia sentences”– *More people have been to Russia than I have.

Page 3: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

Are people doing logic?

• Language Log: “Russia sentences”– *More people have been to Russia than I have.– *It just so happens that more people are bitten by

New Yorkers than they are by sharks.

Page 4: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

Are people doing logic?

• The thing is, is people come up with new ways of speaking all the time.

Page 5: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

More lexical semantics

Page 6: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

Quantifiers

• Every/all: \P. \Q. all x. (P(x) -> Q(x))• A/an/some: \P. \Q. exists x. (P(x) & Q(x))• The:– \P. \Q. Q(x)– P(x) goes in the presuppositions

Page 7: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

High-level overview of C&C

• Find the highest-probability result with coherent semantics

• Doesn’t this create billions of parses that need to be checked?

• Yes.

Page 8: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

High-level overview of C&C

• Parses using a Combinatorial Categorial Grammar (CCG)– fancier than a CFG– includes multiple kinds of “slash rules”– lots of grad student time spent transforming

Treebank• MaxEnt “supertagger” tags each word with a

semantic category

Page 9: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

High-level overview of C&C

• Find the highest-probability result with coherent semantics

• Doesn’t this create billions of parses that need to be checked?

Page 10: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

High-level overview of C&C

• Find the highest-probability result with coherent semantics

• Doesn’t this create millions of parses that need to be checked?

• Yes. A typical sentence uses 25 GB of RAM.• That’s where the Beowulf cluster comes in.

Page 11: Semantics in NLP (part 2) MAS.S60 Rob Speer Catherine Havasi * Lots of slides borrowed for lots of sources! See end

Can we do this with NLTK?

• NLTK’s feature-based parser has some machinery for doing semantics