Presentation is loading. Please wait.

Presentation is loading. Please wait.

Working with Discourse Representation Theory Patrick Blackburn & Johan Bos Lecture 3 DRT and Inference.

Similar presentations


Presentation on theme: "Working with Discourse Representation Theory Patrick Blackburn & Johan Bos Lecture 3 DRT and Inference."— Presentation transcript:

1 Working with Discourse Representation Theory Patrick Blackburn & Johan Bos Lecture 3 DRT and Inference

2 This lecture  Now that we know how to build DRSs for English sentences, what do we do with them?  Well, we can use DRSs to draw inferences.  In this lecture we show how to do that, both in theory and in practice.

3 Overview  Inference tasks  Why FOL?  From model theory to proof theory  Inference engines  From DRT to FOL  Adding world knowledge  Doing it locally

4 The inference tasks  The consistency checking task  The informativity checking task

5 Why First-Order Logic?  Why not use higher-order logic?  Better match with formal semantics  But: Undecidable/no fast provers available  Why not use weaker logics?  Modal/description logics (decidable fragments)  But: Can’t cope with all of natural language  Why use first-order logic?  Undecidable, but good inference tools available  DRS translation to first-order logic  Easy to add world knowledge

6 Axioms encode world knowledge  We can write down axioms about the information that we find fundamental  For example, lexical knowledge, world knowledge, information about the structure of time, events, etc.  By the Deduction Theorem  1 …  n |=  iff |=  1 & … &  n    That is, inference reduces to validity of formulas.

7 From model theory to proof theory  The inference tasks were defined semantically  For computational purposes, we need symbolic definitions  We need to move from the concept of |= to |--  In other words, from validity to provability

8 Soundness  If provable then valid: If |--  then |=   Soundness is a `no garbage` condition

9 Completeness  If valid then provable If |=  then |--   Completeness means that proof theory has captured model theory

10 Decidability  A problem is decidable, if a computer is guaranteed to halt in finite time on any input and give you a correct answer  A problem that is not decidable, is undecidable

11 First-order logic is undecidable  What does this mean? It is not possible, to write a program that is guaranteed to halt when given any first-order formula and correctly tell you whether or not that formula is valid.  Sounds pretty bad!

12 Good news  FOL is semi-decidable  What does that mean?  If in fact a formula is valid, it is always possible, to symbolically verify this fact in finite time  That is, things are only going wrong for FOL when it is asked to tackle something that is not valid  On some non-valid input, any algorithm is bound not to terminate

13 Put differently  Half the task, namely determining validity, is fairly reasonable.  The other half of the task, showing non- validity, or equivalenty, satisfiability, is harder.  This duality is reflected in the fact that there are two fundamental computational inference tools for FOL:  theorem provers  and model builders

14 Theorem provers  Basic thing they do is show that a formula is provable/valid.  There are many efficient off-the-shelf provers available for FOL  Theorem proving technology is now nearly 40 years old and extremely sophisticated  Examples: Vampire, Spass, Bliksem, Otter

15 Theorem provers and informativity  Given a formula , a theorem prover will try to prove , that is, to show that it is valid/uninformative  If  is valid/uninformative, in theory, the theorem prover will always succeed So theorem provers are a negative test for informativity  If the formula  is not valid/uninformative, all bets are off.

16 Theorem provers and consistency  Given a formula , a theorem prover will try to prove , that is, to show that  is inconsistent  If  is inconsistent, in theory, the theorem prover will always succeed So theorem provers are also a negative test for consistency  If the formula  is not inconsistent, all bets are off.

17 Model builders  Basic thing that model builders do is try to generate a [usually] finite model for a formula. They do so by iteration over model size.  Model building for FOL is a rather new field, and there are not many model builders available.  It is also an intrinsically hard task; harder than theorem proving.  Examples: Mace, Paradox, Sem.

18 Model builders and consistency  Given a formula , a model builder will try to build a model for , that is, to show that  is consistent  If  is consistent, and satisfiable on a finite model, then, in theory, the model builder will always succeed So model builders are a partial positive test for consistency  If the formula  is not consistent, or it is not satisfiable on a finite model, all bets are off.

19 Finite model property  A logic has the finite model property, if every satisfiable formula is satisfiable on a finite model.  Many decidable logics have this property.  But it is easy to see that FOL lacks this property.

20 Model builders and informativity  Given a formula , a model builder will try to build a model for , that is, to show that  is informative  If  is satisfiable on a finite model, then, in theory, the model builder will always succeed So model builders are a partial positive test for informativity  If the formula  is not satisfiable on a finite model all bets are off.

21 Yin and Yang of Inference  Theorem Proving and Model Building function as opposite forces

22 Doing it in parallel  We have general negative tests [theorem provers], and partial positive tests [model builders]  Why not try to get of both worlds, by running these tests in parallel?  That is, given a formula we wish to test for informativity/consistency, we hand it to both a theorem prover and model builder at once  When one succeeds, we halt the other

23 Parallel Consistency Checking  Suppose we want to test  [representing the latest sentence] for consistency wrto the previous discourse  Then:  If a theorem prover succeeds in finding a proof for PREV  , then it is inconsistent  If a model builder succeeds to construct a model for PREV & , then it is consistent

24 Why is this relevant to natural language?  Testing a discourse for consistency DiscourseTheorem proverModel builder

25 Why is this relevant to natural language?  Testing a discourse for consistency DiscourseTheorem proverModel builder Vincent is a man. ??Model

26 Why is this relevant to natural language?  Testing a discourse for consistency DiscourseTheorem proverModel builder Vincent is a man. ??Model Mia loves every man. ??Model

27 Why is this relevant to natural language?  Testing a discourse for consistency DiscourseTheorem proverModel builder Vincent is a man. ??Model Mia loves every man. ??Model Mia does not love Vincent. Proof??

28 Parallel informativity checking  Suppose we want to test the formula  [representing the latest sentence] for informativity wrto the previous discourse  Then:  If a theorem prover succeeds in finding a proof for PREV  , then it is not informative  If a model builder succeeds to construct a model for PREV & , then it is informative

29 Why is this relevant to natural language?  Testing a discourse for informativity DiscourseTheorem proverModel builder

30 Why is this relevant to natural language?  Testing a discourse for informativity DiscourseTheorem proverModel builder Vincent is a man. ??Model

31 Why is this relevant to natural language?  Testing a discourse for informativity DiscourseTheorem proverModel builder Vincent is a man. ??Model Mia loves every man. ??Model

32 Why is this relevant to natural language?  Testing a discourse for informativity DiscourseTheorem proverModel builder Vincent is a man. ??Model Mia loves every man. ??Model Mia loves Vincent. Proof??

33 Let`s apply this to DRT  Pretty clear what we need to do:  Find efficient theorem provers for DRT  Find efficient model builders for DRT  Run them in parallel  And Bob`s your uncle!  Recall that theorem provers are more established technology than model builders  So let`s start by finding an efficient theorem prover for DRT…

34 Googling theorem provers for DRT

35 Theorem proving in DRT  Oh no! Nothing there, efficient or otherwise.  Let`s build our own one.  One phone call to Voronkov later:  Oops --- does it take that long to build one from scratch?  Oh dear.

36 Googling theorem provers for FOL

37 Use FOL inference technology for DRT  There are a lot FOL provers available and they are extremely efficient  There are also some interesting freely available model builders for FOL  We have said several times, that DRT is FOL in disguise, so lets get precise about this and put this observation to work

38 From DRT to FOL  Compile DRS into standard FOL syntax  Use off-the-shelf inference engines for FOL  Okay --- how do we do this?  Translation function (…) fo

39 Translating DRT to FOL: DRSs x 1 …x n C1...CnC1...Cn ( ) fo = x 1 … x n ((C 1 ) fo &…&(C n ) fo )

40 Translating DRT to FOL: Conditions (R(x 1 …x n )) fo = R(x 1 …x n ) (x 1 =x 2 ) fo = x 1 =x 2 (B) fo = (B) fo (B 1 B 2 ) fo = (B 1 ) fo  (B 2 ) fo

41 Translating DRT to FOL: Implicative DRS-conditions x 1 …x m C1...CnC1...Cn ( B ) fo = x 1 …x m (((C 1 ) fo &…&(C n ) fo )(B) fo )

42 Two example translations  Example 1  Example 2 x man(x) walk(x) y woman(y)  x man(x) e adore(e) agent(e,x) theme(e,y)

43 Example 1 x man(x) walk(x)

44 Example 1 x man(x) walk(x) ) fo (

45 Example 1  x ( ( man(x) ) fo & ( walk(x) ) fo )

46 Example 1  x ( man(x) & ( walk(x) ) fo )

47 Example 1  x ( man(x) & walk(x) )

48 Example 2 y woman(y)  x man(x) e adore(e) agent(e,x) theme(e,y)

49 Example 2 y woman(y)  x man(x) e adore(e) agent(e,x) theme(e,y) ) fo (

50 Example 2 x man(x) e adore(e) agent(e,x) theme(e,y)  y ( ) ( woman(y) ) fo & (  ) fo

51 Example 2 x man(x) e adore(e) agent(e,x) theme(e,y)  y ( ) woman(y) & (  ) fo

52 Example 2 e adore(e) agent(e,x) theme(e,y) )  y (woman(y) &x ( ( man(x) ) fo  ( ) fo )

53 Example 2 e adore(e) agent(e,x) theme(e,y) )  y (woman(y) &x (man(x)  ( ) fo )

54 Example 2  y (woman(y) &x (man(x)   e ( ( adore(e) ) fo & ( agent(e,x) ) fo & ( theme(e,y) ) fo )))

55 Example 2  y (woman(y) &x (man(x)   e (adore(e) & ( agent(e,x) ) fo & ( theme(e,y) ) fo )))

56 Example 2  y (woman(y) &x (man(x)   e (adore(e) & agent(e,x) & ( theme(e,y) ) fo )))

57 Example 2  y (woman(y) &x (man(x)   e (adore(e) & agent(e,x) & theme(e,y))))

58 Basic setup  DRS: x y vincent(x) mia(y) love(x,y)

59 Basic setup  DRS:  FOL:  x  y(vincent(x) & mia(y) & love(x,y)) x y vincent(x) mia(y) love(x,y)

60 Basic setup  DRS:  FOL:  x  y(vincent(x) & mia(y) & love(x,y))  Model: D = {d1} F(vincent)={d1} F(mia)={d1} F(love)={(d1,d1)} x y vincent(x) mia(y) love(x,y)

61 Background Knowledge (BK)  Need to incorporate BK  Formulate BK in terms of first-order axioms  Rather than just giving  to the theorem prover (or model builder), we give it: BK &  or BK  

62 Basic setup  DRS: x y vincent(x) mia(y) love(x,y)

63 Basic setup  DRS:  FOL:  x  y(vincent(x) & mia(y) & love(x,y)) x y vincent(x) mia(y) love(x,y)

64 Basic setup  DRS:  FOL:  x  y(vincent(x) & mia(y) & love(x,y))  BK:  x (vincent(x)  man(x))  x (mia(x)  woman(x))  x (man(x)   woman(x)) x y vincent(x) mia(y) love(x,y)

65 Basic setup  DRS:  FOL:  x  y(vincent(x) & mia(y) & love(x,y))  BK:  x (vincent(x)  man(x))  x (mia(x)  woman(x))  x (man(x)   woman(x))  Model: D = {d1,d2} F(vincent)={d1} F(mia)={d2} F(love)={(d1,d2)} x y vincent(x) mia(y) love(x,y)

66 Local informativity  Example:  Mia is the wife of Marsellus.  If Mia is the wife of Marsellus, Vincent will be disappointed.  The second sentence is informative with respect to the first. But…

67 x y mia(x) marsellus(y) wife-of(x,y) Local informativity

68 x y z mia(x) marsellus(y) wife-of(x,y) vincent(z) wife-of(x,y)  disappointed(z) Local informativity

69 Local consistency  Example:  Jules likes big kahuna burgers.  If Jules does not like big kahuna burgers, Vincent will order a whopper.  The second sentence is consistent with respect to the first. But…

70 x y jules(x) big-kahuna-burgers(y) like(x,y) Local consistency

71 x y z jules(x) big-kahuna-burgers(y) like(x,y) vincent(z)   u order(z,u) whopper(u) Local consistency like(x,y)

72 DRT and local inference  Because DRS groups information into contexts, we now have natural means to check not only global, but also local consistency and informativity.  Important for dealing with presupposition.  Presupposition is not about strange logic. But about using classical logic in new ways.

73 Tomorrow  Presupposition and Anaphora in DRT


Download ppt "Working with Discourse Representation Theory Patrick Blackburn & Johan Bos Lecture 3 DRT and Inference."

Similar presentations


Ads by Google