Using World Knowledge

Reference: Chapter 15 of Allen

Aim:

This section considers the problem of using world knowledge in interpreting logical forms.

Keywords: coherence, reference, disambiguation, implicature, expectations, matching, scripts

Plan:

establishing coherence
matching against expectations
using scripts

Coherence

Coherence is about making sense of sequences of sentences (or discourse).
Frequently that involves using knowledge of the world to work out the connection between the sentences.
Example:

1a. Jack took out a match.
1b. He lit a candle.
What are the connections between 1a and 1b?

Image from www.nelms.org.uk
via Google images.

Coherence 2

1a. Jack took out a match.
1b. He lit a candle.
The connections come under several headings:

reference — He refers to Jack1

disambiguation — lit refers to igniting¹ rather than illuminating² the candle.

implicature — The instrument for lighting the candle is the match in sentence 1a.

Footnotes:
[1] ignite comes from the Latin word ignis = fire
[2] illuminate comes from the Latin word lumen = light

Coherence 3

1a. Jack took out a match.
1b. He lit a candle.
To handle the implicature part, you need to know that a typical way to light a candle involves using a match.
Not all pairs of sentences are related in this way. For example, with

2a. John took out a match.
2b. The sun set.

... there is no implicature-type relation. The only likely relation is temporal - that is, that 2a → 2b is a time sequence.

Matching Against Expectations

The output of processing a sequence of sentences, as well as parses, logical forms, and reference information, will include any inferences made in order to establish coherence.
The inferences are used to generate a set of expectations about what may happen next.
In some cases there may be alternative inferences, one of which will turn out to match the material in later sentences, and the other(s) will be discarded.

Matching Against Expectations 2

An expectation from 1a. Jack took out a match would be:

Expectation 1:
Ignite(E1) & Agent(E1) = Jack1 & Instrument(E1) = Match33

That is, there is a Ignite-event E1 with agent Jack1 and instrument Match33
or
```
event(e1).
isa(e1, ignite).
agent(e1, jack1).
instrument(e1, match33).
   
```

Matching Against Expectations 3

The word lit in 1b is ambiguous between Ignite1 and Illuminate1, as in The lamp lit the room.
So, in isolation, 1b has two interpretations:

Interpretation 1:
Ignite(E2) & Agent(E2) = Jack1 & Theme(E2) = Candle1

Interpretation 2:
Illuminate(E3) & Agent(E3) = Jack1 & Theme(E3) = Candle1

Matching Against Expectations 4

We have to somehow connect Expectation 1 with one of the Interpretations.
Could it be that Expectation 1 ⇒ Interpretation j (or the converse)?
In fact Expectation 1 implies neither interpretation.
Similarly, neither interpretation implies Expectation 1.
What we need is a Prolog-style matching/unification process.

Matching Against Expectations 5

semantic net diagrams

Figure 15.1 from Allen
The expectation and two possible interpretations

Matching Against Expectations 6

result of matching semantic nets

Figure 15.2 from Allen
The result of matching/unifying E1 and E2

Matching Against Expectations 7

Or, in Prolog terms:

event(e1_2).
isa(e1_2, ignite).
agent(e1_2, jack1).
instrument(e1_2, match33).
theme(e1_2, candle1).

Scripts: Understanding Stereotypical Situations

Simple unification like what we have just seen handles simple forms of world knowledge. For more complex clusters of world knowledge, scripts can help.
A script describes a common situation or scenario:
restaurant_visit
terrorist_attack
travel_by_train (other scripts for bus, plane, taxi,...)
attend_lecture
apply_for_job
birthday_party
...
A script is rather like a frame, that describes a sequence of related sub-events.

Scripts: Understanding Stereotypical Situations 2

Example of (generic) script:

Roles	Actor, Clerk, SourceStation, DestStation, Train, Booth, Ticket, Money
Constraints	Person(Actor), Person(Clerk), Station(SourceStation), Station(DestStation), TicketBooth(Booth), In(Booth, SourceStation), At(Clerk, Booth), TicketFor(Ticket, SourceStation, DestStation), Value(Money, PriceOf(Ticket))
Preconditions	Owns(Actor, Money), At(Actor, SourceStation)
Effects	¬ Owns(Actor, Money), ¬ At(Actor, SourceStation), At(Actor, DestStation),
Decomposition	PurchaseTicket(Actor, Clerk, Ticket, SourceStation) Goto(Actor, Location(Train)) GetOn(Actor, Train) Travel(Train, SourceStation, DestStation) Arrive(Train, DestStation) GetOff(Actor, Train)

Scripts 3

Scripts are instantiated in the same way as frames.
In the example script above, you would need to instantiate each of the roles in the script to produce an instance script.
e.g. Actor=Fred1,
Clerk=Harry1,
SourceStation=Central,
DestStation=Burwood,
Train="4.45pm to Emu Plains",
Booth=Booth13,
Ticket=Ticket123456789,
Money=3.00

Scripts 4

This produces an instantiated decomposition, preconditions, effects, and constraints.
E.g. instantiated decomposition:

PurchaseTicket(Fred1, Harry1, Ticket123456789, Central)
Goto(Fred1, Location("4.45pm to Emu Plains"))
GetOn(Fred1, "4.45pm to Emu Plains")
Travel("4.45pm to Emu Plains", Central, Burwood)
Arrive("4.45pm to Emu Plains", Burwood)
GetOff(Fred1, "4.45pm to Emu Plains")

Scripts 5

These instantiated propositions can then be verified/matched against what is coming in from the parsing/interpretation/reference process.
How do you choose the (right) script?
How do you know when you move to another script?
What happens when scripts are overlapping or embedded in one another? - e.g. in a discourse about someone visiting a restaurant while at an airport.

Summary

Using World Knowledge
We have looked at the problem of establishing coherence in a discourse, and the fact that doing so may involve eliminating unlikely readings of the sentences in the discourse. We saw that unification/matching can be helpful in this process. We have also looked at scripts, that is, frames that describe sequences of events in stereotyped/common situations, as a mechanism for dealing with longer discourses.

Using World Knowledge

We have looked at the problem of establishing coherence in a discourse, and the fact that doing so may involve eliminating unlikely readings of the sentences in the discourse. We saw that unification/matching can be helpful in this process.

We have also looked at scripts, that is, frames that describe sequences of events in stereotyped/common situations, as a mechanism for dealing with longer discourses.

© Copyright Bill Wilson, 2007, except where another source is acknowledged. The material in this section is adapted from James Allen's book "Natural Language Processing".
Bill Wilson's contact info
Last updated: