Discourse Checking

>>> from nltk import *
>>> from nltk.sem import logic
>>> logic._counter._value = 0

Introduction

The NLTK discourse module makes it possible to test consistency and redundancy of simple discourses, using theorem-proving and model-building from nltk.inference.

The DiscourseTester constructor takes a list of sentences as a parameter.

>>> dt = DiscourseTester(['a boxer walks', 'every boxer chases a girl'])

The DiscourseTester parses each sentence into a list of logical forms. Once we have created DiscourseTester object, we can inspect various properties of the discourse. First off, we might want to double-check what sentences are currently stored as the discourse.

>>> dt.sentences()
s0: a boxer walks
s1: every boxer chases a girl

As you will see, each sentence receives an identifier si. We might also want to check what grammar the DiscourseTester is using (by default, book_grammars/discourse.fcfg):

>>> dt.grammar() # doctest: +ELLIPSIS
% start S
# Grammar Rules
S[SEM = <app(?subj,?vp)>] -> NP[NUM=?n,SEM=?subj] VP[NUM=?n,SEM=?vp]
NP[NUM=?n,SEM=<app(?det,?nom)> ] -> Det[NUM=?n,SEM=?det]  Nom[NUM=?n,SEM=?nom]
NP[LOC=?l,NUM=?n,SEM=?np] -> PropN[LOC=?l,NUM=?n,SEM=?np]
...

A different grammar can be invoked by using the optional gramfile parameter when a DiscourseTester object is created.

Readings and Threads

Depending on the grammar used, we may find some sentences have more than one logical form. To check this, use the readings() method. Given a sentence identifier of the form si, each reading of that sentence is given an identifier si-rj.

>>> dt.readings()
<BLANKLINE>
s0 readings:
<BLANKLINE>
s0-r0: exists z1.(boxer(z1) & walk(z1))
s0-r1: exists z1.(boxerdog(z1) & walk(z1))
<BLANKLINE>
s1 readings:
<BLANKLINE>
s1-r0: all z2.(boxer(z2) -> exists z3.(girl(z3) & chase(z2,z3)))
s1-r1: all z1.(boxerdog(z1) -> exists z2.(girl(z2) & chase(z1,z2)))

In this case, the only source of ambiguity lies in the word boxer, which receives two translations: boxer and boxerdog. The intention is that one of these corresponds to the person sense and one to the dog sense. In principle, we would also expect to see a quantifier scope ambiguity in s1. However, the simple grammar we are using, namely sem4.fcfg, doesn't support quantifier scope ambiguity.

We can also investigate the readings of a specific sentence:

>>> dt.readings('a boxer walks')
The sentence 'a boxer walks' has these readings:
    exists x.(boxer(x) & walk(x))
    exists x.(boxerdog(x) & walk(x))

Given that each sentence is two-ways ambiguous, we potentially have four different discourse 'threads', taking all combinations of readings. To see these, specify the threaded=True parameter on the readings() method. Again, each thread is assigned an identifier of the form di. Following the identifier is a list of the readings that constitute that thread.

>>> dt.readings(threaded=True) # doctest: +NORMALIZE_WHITESPACE
d0: ['s0-r0', 's1-r0']
d1: ['s0-r0', 's1-r1']
d2: ['s0-r1', 's1-r0']
d3: ['s0-r1', 's1-r1']

Of course, this simple-minded approach doesn't scale: a discourse with, say, three sentences, each of which has 3 readings, will generate 27 different threads. It is an interesting exercise to consider how to manage discourse ambiguity more efficiently.

Checking Consistency

Now, we can check whether some or all of the discourse threads are consistent, using the models() method. With no parameter, this method will try to find a model for every discourse thread in the current discourse. However, we can also specify just one thread, say d1.

>>> dt.models('d1')
--------------------------------------------------------------------------------
Model for Discourse Thread d1
--------------------------------------------------------------------------------
% number = 1
% seconds = 0
<BLANKLINE>
% Interpretation of size 2
<BLANKLINE>
c1 = 0.
<BLANKLINE>
f1(0) = 0.
f1(1) = 0.
<BLANKLINE>
  boxer(0).
- boxer(1).
<BLANKLINE>
- boxerdog(0).
- boxerdog(1).
<BLANKLINE>
- girl(0).
- girl(1).
<BLANKLINE>
  walk(0).
- walk(1).
<BLANKLINE>
- chase(0,0).
- chase(0,1).
- chase(1,0).
- chase(1,1).
<BLANKLINE>
Consistent discourse: d1 ['s0-r0', 's1-r1']:
    s0-r0: exists z1.(boxer(z1) & walk(z1))
    s1-r1: all z1.(boxerdog(z1) -> exists z2.(girl(z2) & chase(z1,z2)))
<BLANKLINE>

There are various formats for rendering Mace4 models --- here, we have used the 'cooked' format (which is intended to be human-readable). There are a number of points to note.

  1. The entities in the domain are all treated as non-negative integers. In this case, there are only two entities, 0 and 1.
  2. The - symbol indicates negation. So 0 is the only boxerdog and the only thing that walks. Nothing is a boxer, or a girl or in the chase relation. Thus the universal sentence is vacuously true.
  3. c1 is an introduced constant that denotes 0.
  4. f1 is a Skolem function, but it plays no significant role in this model.

We might want to now add another sentence to the discourse, and there is method add_sentence() for doing just this.

>>> dt.add_sentence('John is a boxer')
>>> dt.sentences()
s0: a boxer walks
s1: every boxer chases a girl
s2: John is a boxer

We can now test all the properties as before; here, we just show a couple of them.

>>> dt.readings()
<BLANKLINE>
s0 readings:
<BLANKLINE>
s0-r0: exists z1.(boxer(z1) & walk(z1))
s0-r1: exists z1.(boxerdog(z1) & walk(z1))
<BLANKLINE>
s1 readings:
<BLANKLINE>
s1-r0: all z1.(boxer(z1) -> exists z2.(girl(z2) & chase(z1,z2)))
s1-r1: all z1.(boxerdog(z1) -> exists z2.(girl(z2) & chase(z1,z2)))
<BLANKLINE>
s2 readings:
<BLANKLINE>
s2-r0: boxer(John)
s2-r1: boxerdog(John)
>>> dt.readings(threaded=True) # doctest: +NORMALIZE_WHITESPACE
d0: ['s0-r0', 's1-r0', 's2-r0']
d1: ['s0-r0', 's1-r0', 's2-r1']
d2: ['s0-r0', 's1-r1', 's2-r0']
d3: ['s0-r0', 's1-r1', 's2-r1']
d4: ['s0-r1', 's1-r0', 's2-r0']
d5: ['s0-r1', 's1-r0', 's2-r1']
d6: ['s0-r1', 's1-r1', 's2-r0']
d7: ['s0-r1', 's1-r1', 's2-r1']

If you are interested in a particular thread, the expand_threads() method will remind you of what readings it consists of:

>>> thread = dt.expand_threads('d1')
>>> for rid, reading in thread:
...     print(rid, str(reading.normalize()))
s0-r0 exists z1.(boxer(z1) & walk(z1))
s1-r0 all z1.(boxer(z1) -> exists z2.(girl(z2) & chase(z1,z2)))
s2-r1 boxerdog(John)

Suppose we have already defined a discourse, as follows:

>>> dt = DiscourseTester(['A student dances', 'Every student is a person'])

Now, when we add a new sentence, is it consistent with what we already have? The `` consistchk=True`` parameter of add_sentence() allows us to check:

>>> dt.add_sentence('No person dances', consistchk=True)
Inconsistent discourse: d0 ['s0-r0', 's1-r0', 's2-r0']:
    s0-r0: exists z1.(student(z1) & dance(z1))
    s1-r0: all z1.(student(z1) -> person(z1))
    s2-r0: -exists z1.(person(z1) & dance(z1))
<BLANKLINE>
>>> dt.readings()
<BLANKLINE>
s0 readings:
<BLANKLINE>
s0-r0: exists z1.(student(z1) & dance(z1))
<BLANKLINE>
s1 readings:
<BLANKLINE>
s1-r0: all z1.(student(z1) -> person(z1))
<BLANKLINE>
s2 readings:
<BLANKLINE>
s2-r0: -exists z1.(person(z1) & dance(z1))

So let's retract the inconsistent sentence:

>>> dt.retract_sentence('No person dances', verbose=True) # doctest: +NORMALIZE_WHITESPACE
Current sentences are
s0: A student dances
s1: Every student is a person

We can now verify that result is consistent.

>>> dt.models()
--------------------------------------------------------------------------------
Model for Discourse Thread d0
--------------------------------------------------------------------------------
% number = 1
% seconds = 0
<BLANKLINE>
% Interpretation of size 2
<BLANKLINE>
c1 = 0.
<BLANKLINE>
  dance(0).
- dance(1).
<BLANKLINE>
  person(0).
- person(1).
<BLANKLINE>
  student(0).
- student(1).
<BLANKLINE>
Consistent discourse: d0 ['s0-r0', 's1-r0']:
    s0-r0: exists z1.(student(z1) & dance(z1))
    s1-r0: all z1.(student(z1) -> person(z1))
<BLANKLINE>

Checking Informativity

Let's assume that we are still trying to extend the discourse A student dances. Every student is a person. We add a new sentence, but this time, we check whether it is informative with respect to what has gone before.

>>> dt.add_sentence('A person dances', informchk=True)
Sentence 'A person dances' under reading 'exists x.(person(x) & dance(x))':
Not informative relative to thread 'd0'

In fact, we are just checking whether the new sentence is entailed by the preceding discourse.

>>> dt.models()
--------------------------------------------------------------------------------
Model for Discourse Thread d0
--------------------------------------------------------------------------------
% number = 1
% seconds = 0
<BLANKLINE>
% Interpretation of size 2
<BLANKLINE>
c1 = 0.
<BLANKLINE>
c2 = 0.
<BLANKLINE>
  dance(0).
- dance(1).
<BLANKLINE>
  person(0).
- person(1).
<BLANKLINE>
  student(0).
- student(1).
<BLANKLINE>
Consistent discourse: d0 ['s0-r0', 's1-r0', 's2-r0']:
    s0-r0: exists z1.(student(z1) & dance(z1))
    s1-r0: all z1.(student(z1) -> person(z1))
    s2-r0: exists z1.(person(z1) & dance(z1))
<BLANKLINE>

Adding Background Knowledge

Let's build a new discourse, and look at the readings of the component sentences:

>>> dt = DiscourseTester(['Vincent is a boxer', 'Fido is a boxer', 'Vincent is married', 'Fido barks'])
>>> dt.readings()
<BLANKLINE>
s0 readings:
<BLANKLINE>
s0-r0: boxer(Vincent)
s0-r1: boxerdog(Vincent)
<BLANKLINE>
s1 readings:
<BLANKLINE>
s1-r0: boxer(Fido)
s1-r1: boxerdog(Fido)
<BLANKLINE>
s2 readings:
<BLANKLINE>
s2-r0: married(Vincent)
<BLANKLINE>
s3 readings:
<BLANKLINE>
s3-r0: bark(Fido)

This gives us a lot of threads:

>>> dt.readings(threaded=True) # doctest: +NORMALIZE_WHITESPACE
d0: ['s0-r0', 's1-r0', 's2-r0', 's3-r0']
d1: ['s0-r0', 's1-r1', 's2-r0', 's3-r0']
d2: ['s0-r1', 's1-r0', 's2-r0', 's3-r0']
d3: ['s0-r1', 's1-r1', 's2-r0', 's3-r0']

We can eliminate some of the readings, and hence some of the threads, by adding background information.

>>> import nltk.data
>>> bg = nltk.data.load('grammars/book_grammars/background.fol')
>>> dt.add_background(bg)
>>> dt.background()
all x.(boxerdog(x) -> dog(x))
all x.(boxer(x) -> person(x))
all x.-(dog(x) & person(x))
all x.(married(x) <-> exists y.marry(x,y))
all x.(bark(x) -> dog(x))
all x y.(marry(x,y) -> (person(x) & person(y)))
-(Vincent = Mia)
-(Vincent = Fido)
-(Mia = Fido)

The background information allows us to reject three of the threads as inconsistent. To see what remains, use the filter=True parameter on readings().

>>> dt.readings(filter=True) # doctest: +NORMALIZE_WHITESPACE
d1: ['s0-r0', 's1-r1', 's2-r0', 's3-r0']

The models() method gives us more information about the surviving thread.

>>> dt.models()
--------------------------------------------------------------------------------
Model for Discourse Thread d0
--------------------------------------------------------------------------------
No model found!
<BLANKLINE>
--------------------------------------------------------------------------------
Model for Discourse Thread d1
--------------------------------------------------------------------------------
% number = 1
% seconds = 0
<BLANKLINE>
% Interpretation of size 3
<BLANKLINE>
Fido = 0.
<BLANKLINE>
Mia = 1.
<BLANKLINE>
Vincent = 2.
<BLANKLINE>
f1(0) = 0.
f1(1) = 0.
f1(2) = 2.
<BLANKLINE>
  bark(0).
- bark(1).
- bark(2).
<BLANKLINE>
- boxer(0).
- boxer(1).
  boxer(2).
<BLANKLINE>
  boxerdog(0).
- boxerdog(1).
- boxerdog(2).
<BLANKLINE>
  dog(0).
- dog(1).
- dog(2).
<BLANKLINE>
- married(0).
- married(1).
  married(2).
<BLANKLINE>
- person(0).
- person(1).
  person(2).
<BLANKLINE>
- marry(0,0).
- marry(0,1).
- marry(0,2).
- marry(1,0).
- marry(1,1).
- marry(1,2).
- marry(2,0).
- marry(2,1).
  marry(2,2).
<BLANKLINE>
--------------------------------------------------------------------------------
Model for Discourse Thread d2
--------------------------------------------------------------------------------
No model found!
<BLANKLINE>
--------------------------------------------------------------------------------
Model for Discourse Thread d3
--------------------------------------------------------------------------------
No model found!
<BLANKLINE>
Inconsistent discourse: d0 ['s0-r0', 's1-r0', 's2-r0', 's3-r0']:
    s0-r0: boxer(Vincent)
    s1-r0: boxer(Fido)
    s2-r0: married(Vincent)
    s3-r0: bark(Fido)
<BLANKLINE>
Consistent discourse: d1 ['s0-r0', 's1-r1', 's2-r0', 's3-r0']:
    s0-r0: boxer(Vincent)
    s1-r1: boxerdog(Fido)
    s2-r0: married(Vincent)
    s3-r0: bark(Fido)
<BLANKLINE>
Inconsistent discourse: d2 ['s0-r1', 's1-r0', 's2-r0', 's3-r0']:
    s0-r1: boxerdog(Vincent)
    s1-r0: boxer(Fido)
    s2-r0: married(Vincent)
    s3-r0: bark(Fido)
<BLANKLINE>
Inconsistent discourse: d3 ['s0-r1', 's1-r1', 's2-r0', 's3-r0']:
    s0-r1: boxerdog(Vincent)
    s1-r1: boxerdog(Fido)
    s2-r0: married(Vincent)
    s3-r0: bark(Fido)
<BLANKLINE>

In order to play around with your own version of background knowledge, you might want to start off with a local copy of background.fol:

>>> nltk.data.retrieve('grammars/book_grammars/background.fol')
Retrieving 'nltk:grammars/book_grammars/background.fol', saving to 'background.fol'

After you have modified the file, the load_fol() function will parse the strings in the file into expressions of nltk.sem.logic.

>>> from nltk.inference.discourse import load_fol
>>> mybg = load_fol(open('background.fol').read())

The result can be loaded as an argument of add_background() in the manner shown earlier.

Regression Testing from book

>>> logic._counter._value = 0
>>> from nltk.tag import RegexpTagger
>>> tagger = RegexpTagger(
...     [('^(chases|runs)$', 'VB'),
...      ('^(a)$', 'ex_quant'),
...      ('^(every)$', 'univ_quant'),
...      ('^(dog|boy)$', 'NN'),
...      ('^(He)$', 'PRP')
... ])
>>> rc = DrtGlueReadingCommand(depparser=MaltParser(tagger=tagger))
>>> dt = DiscourseTester(map(str.split, ['Every dog chases a boy', 'He runs']), rc)
>>> dt.readings()
<BLANKLINE>
s0 readings:
<BLANKLINE>
s0-r0: ([z2],[boy(z2), (([z5],[dog(z5)]) -> ([],[chases(z5,z2)]))])
s0-r1: ([],[(([z1],[dog(z1)]) -> ([z2],[boy(z2), chases(z1,z2)]))])
<BLANKLINE>
s1 readings:
<BLANKLINE>
s1-r0: ([z1],[PRO(z1), runs(z1)])
>>> dt.readings(show_thread_readings=True)
d0: ['s0-r0', 's1-r0'] : ([z1,z2],[boy(z1), (([z3],[dog(z3)]) -> ([],[chases(z3,z1)])), (z2 = z1), runs(z2)])
d1: ['s0-r1', 's1-r0'] : INVALID: AnaphoraResolutionException
>>> dt.readings(filter=True, show_thread_readings=True)
d0: ['s0-r0', 's1-r0'] : ([z1,z3],[boy(z1), (([z2],[dog(z2)]) -> ([],[chases(z2,z1)])), (z3 = z1), runs(z3)])
>>> logic._counter._value = 0
>>> from nltk.parse import FeatureEarleyChartParser
>>> from nltk.sem.drt import DrtParser
>>> grammar = nltk.data.load('grammars/book_grammars/drt.fcfg', logic_parser=DrtParser())
>>> parser = FeatureEarleyChartParser(grammar, trace=0)
>>> trees = parser.parse('Angus owns a dog'.split())
>>> print(list(trees)[0].label()['SEM'].simplify().normalize())
([z1,z2],[Angus(z1), dog(z2), own(z1,z2)])