CSA4050: Advanced Topics in NLP

CSA4050:Advanced Topics in NLP

Semantics IV• Partial Execution• Proper Noun• Transitive Verb Phrase

October 2004 CSA4050 Advanced NLP 2

Partial Execution

• Partial execution involves the replacing of certain runtime computations with changes to the source of the Prolog program itself.

• For example, we can replace the rule:s(S) --> np(NP), vp(VP) { reduce(NP,VP,S) }.withs(S) --> np(VP^S), vp(S).

• This is because the computation of reduce involves only the mutual binding of several variables.

October 2004 CSA4050 Advanced NLP 3

How it works

s(S) --> np(NP), vp(VP) { reduce(NP,VP,S)}.

1. match reduce(A^F,A,F)

2. rewrite s(F) --> np(A^F), vp(A).

3. rename s(S) --> np(VP^S), vp(VP).

October 2004 CSA4050 Advanced NLP 4

Exercise

• What is the result of eliminating the reduce clause by partial execution in the following rule?

np(NP) --> d(D), n(N), {reduce(D,N,NP)}.

October 2004 CSA4050 Advanced NLP 5

Answer

np(NP) --> d(D), n(N) { reduce(D,N,NP)}.

1. match reduce(A^F,A,F)

2. rewrite np(F) --> d(A^F), n(A).

3. rename np(NP) --> d(N^NP), n(N).

October 2004 CSA4050 Advanced NLP 6

DCG with QuantificationProgram 4

% grammars(S) --> np(VP^S), vp(VP).np(NP) --> d(N^NP), n(N).vp(VP) --> v(VP).

% lexiconv(X^walk(X)) --> [walks].n(X^man(X)) --> [man].n(suzie) --> [‘Suzie’].det(RL^SL^all(X,R,S) --> [every], {reduce(RL,X,R), reduce(SL,X,S) }.

October 2004 CSA4050 Advanced NLP 7

Handling Proper Nouns

• The grammar handles every man walksX = all(_G, man(_G), walk(_G))

• Will this grammar parse Suzie walks?

• Let’s try it!

•?- s(X,['Suzie',walks],[ ]).

October 2004 8

?- s(X,['Suzie',walks],[ ]).

Call: (8) s(_G492, ['Suzie', walks], []) ? Call: (9) np(_L183, ['Suzie', walks], _L184) ? Call: (10) pn(_L183, ['Suzie', walks], _L184) ? Exit: (10) pn(suzie, ['Suzie', walks], [walks]) ? Exit: (9) np(suzie, ['Suzie', walks], [walks]) ? Call: (9) vp(_L185, [walks], _L186) ? Call: (10) iv(_L185, [walks], _L186) ? Exit: (10) iv(_G556^walk(_G556), [walks], []) ? Exit: (9) vp(_G556^walk(_G556), [walks], []) ? Call: (9) reduce(suzie, _G556^walk(_G556), _G492) ? Fail: (9) reduce(suzie, _G556^walk(_G556), _G492)? ........

suzie is not a function

October 2004 CSA4050 Advanced NLP 9

Handling Proper Nouns

• Problem is with the “type” of LF of Suzie.

• We require that LF of Suzie has the same type as any other NP - such as every man, i.e.– As a lambda expression it would be

λp.p(suzie). – In Prolog we can regard this as a function

from [VP] to [S] such that reduce(VP,john,S) holds.

October 2004 CSA4050 Advanced NLP 10

DCG with QuantificationProgram 4

% grammars(S) --> np(VP^S), vp(VP).np(NP) --> n(NP).np(NP) --> d(N^NP), n(N).vp(VP) --> v(VP).

% lexiconv(X^walk(X)) --> [walks].n(X^man(X)) --> [man].n(VP^S) --> [‘Suzie’],

{reduce(VP,suzie,S)}.det(RL^SL^all(X,R => S) --> [every],

{reduce(RL,X,R), reduce(SL,X,S) }.

October 2004 CSA4050 Advanced NLP 11

Exercise 2

Using partial execution, eliminate the reduce clause in

pn(VP^S) --> [‘Suzie’],{reduce(VP,suzie,S)}.

October 2004 CSA4050 Advanced NLP 12

s(X, ['Suzie', walks], [ ]) ?

Call: (7) s(_G292, ['Suzie', walks], []) ? ↓Exit: (9) pn((suzie^_G357)^_G357, ['Suzie', walks], [walks]) ? ↓Exit: (9) iv(_G362^walk(_G362), [walks], []) ?Exit: (8) vp(_G362^walk(_G362), [walks], []) ?:Call: (8) reduce((suzie^_G357)^_G357, _G362^walk(_G362), _G292) ?:Exit: (8) reduce((suzie^walk(suzie))^walk(suzie), suzie^walk(suzie),

walk(suzie)) ? Exit: (7) s(walk(suzie), ['Suzie', walks], []) ? creep

X = walk(suzie)

October 2004 CSA4050 Advanced NLP 13

Transitive Verb Phases

• Transitive verb phrases take an object, e.g.chased a cat

• To handle transitive VPs we need another VP rulevp(VP) --> v(V), np(NP).

• In this case we have:V = λx. λy. chased(x,y)NP = λs.some(w,cat(w),s(w))

• Issues:– What is the LF of the resulting VP and– How do the component LFs combine?

October 2004 CSA4050 Advanced NLP 14

Resulting VP

λz.some(y,cat(y),chased(z,y))

• Note that this has the form of a standard VP that is waiting for a subject, so is compatible with the rest of the grammar.

October 2004 CSA4050 Advanced NLP 15

Making VP

V

λx. λy. chased(x,y)

chased

NP

λs.some(w,cat(w),s(w))

some cat

VP

λz.some(y,cat(y),chased(z,y))

October 2004 CSA4050 Advanced NLP 16

Apply NP to Vdoesn't work!

λs.some(w,cat(w), s(w)) (λx. λy. chased(x,y))

β

some(w,cat(w), (λy. λx. chased(x,y))(w))β

some(w,cat(w), λx.chased(x,w))

NP V

October 2004 CSA4050 Advanced NLP 17

Solution

• Need a more complex way of combining V and NP: λNP.λV.λz.NP(V(z))

λs.some(w,cat(w), s(w)) ← NP arg

λx. λy. chased(x,y)) ← V arg

October 2004 CSA4050 Advanced NLP 18

How it works

λNP.λV.λz.NP(V(z)) λs.some(w,cat(w), s(w)) λx. λy. chased(x,y))

λ z.λs.some(w,cat(w), s(w)) λx.λy.chased(x,y)(z) β

λ z.λs.some(w,cat(w), s(w)) (λy. chased(z,y)) β

λ z. some(cat(w), (λy.chased(z,y))(w)) β

λ z. some(cat(w), chased(z,w))

October 2004 CSA4050 Advanced NLP 19

VP Rule Revisited

vp(VP) --> v(V), np(NP),

{instructions}.• The instructions encode the combination

λNP.λV.λz.NP(V(z)) vp(Z^P) --> v(V2), np(NP),

{reduce(V2,Z,V1),

reduce(NP,V1,P)}.

October 2004 CSA4050 Advanced NLP 20

Program 5% grammars(S) --> np(NP), vp(VP), {reduce(NP,VP,S)}.np(NP) --> n(NP).np(NP) --> d(D), n(N), {reduce(D,N,NP) }.vp(VP) --> v(VP).vp(Z^P)--> v(V2), np(NP),

{reduce(V2,Z,V1), reduce(NP,V1,P)}.

% lexiconv(X^walk(X)) --> [walks].v(X^Y^chased(X,Y)) --> [chased].n(X^cat(X)) --> [cat].n(VP^S) --> [suzie], {reduce(VP,suzie,S)}.d(RL^SL^some(X,R,S)) --> [a],

{reduce(RL,X,R),reduce(SL,X,S) }.

October 2004 CSA4050 Advanced NLP 21

Test

?- s(X,[suzie,chased,a,cat],[]).

X = some(_G245, cat(_G245), chased(suzie, _G245)) ?- s(X,[a,cat,chased,a,cat],[]).

X = some(_G245, cat(_G245), some(_G266, cat(_G266), chased(_G245, _G266))) ;