8 - Pragmatics and World Knowledge

Storing WK – Semantic Nets

conceptual relations or semantic relations:

is_type_of indicates the hypernym relation

EXEMPLIFICATION RELATION = concept + its exemplification (example(s) of the concept)

PARTITION RELATION = concept + its partition (what the parts of the concept are)

CAUSATION RELATION = concept + its cause (what caused the concept usually an action or a state)

NOMINATION RELATION = concept + its name (what it is called) (Also called APELLATION)

dog includes sheepdog (Concept=dog, hyponym=sheepdog)

sheepdog has_an_instance Shep (Concept=sheepdog, instance=Shep)

mainframe3 is_type_of computer1

PC4 is_type_of computer1

frames and slots method - all the data about a concept is held in the node (“frame”) as a table of rows (“slots”)

Where a semantic net contains hypernym relations, sets of characteristics (stored as slots in a frame) may be inherited from parent class to daughter class – OO methodology

Sowa’s Conceptual Graphs

• conceptual relation nodes.

Some example conceptual graphs of various “arities”, taken from Luger and Stubblefield:

<see your notes>

In notes usually use the word theme instead of object.

Easier Ways to Create KBs ?

We might actually be able to automate the process of building semantic

Pragmatic Knowledge

I purple ate cheese some.

I ate some colourless purple cheese.

Knowledge about Causation

The notion of causation –

actions have effects

• the use of a causes arrow in a semantic net.

But note that the source of such an arrow has to be a process rather than a static thing.

Applying WK in NLP programs – Text Understanding

…if we can find an assumption regarding the already-read text that is consistent with it (i.e. a possible expectation), and we can also show that that assumption entails (allows, causes, makes possible) one of the possible interpretations (meanings) of the latest sentence read, then we can add that particular interpretation of the latest sentence to the understanding structure of the text.

For example, in the “heart attack at the wheel” story above, the first sentence gives us the expectation (assumption) that the driver lost control of the car. We know that losing control of a car causes that car to crash (from our WK), and this would be perfectly consistent with the story so far. The second sentence tells us that the car did indeed crash (that’s the only possible interpretation, because it is an unambiguous stated fact). Thus this interpretation is entailed by a consistent expectation/assumption from the existing text, and we join the second sentence into our understanding construct formed by the text so far (sentence 1).

(Agent(E1)=Jack) &

(Instrument(E1)=Match)

Task is to examine the two interpretations of the second sentence and see if we can “match” one of them to the expectation arising from the preceding text.

EXPECTATION INTERPRETATION1 INTERPRETATION2

Jack Match Jack Candle Jack Candle

We shall say that an expectation matches an interpretation if there is a consistent set of equality assumptions that can be made which make the expectation exactly match the interpretation (i.e. the two semantic net fragments are then identical), or one a subset of the other (i.e. the same semantic net but with part missing).

SET_ALIGHT

is_an_instance_of

• We are allowed to assume a theme (=Candle) for the SET_ALIGHT event type, because such an event type allows a theme (in fact it actually requires one).

• What of the other interpretation – the third diagram? Could we have found a set of equality assumptions for a match here? No.

Building Logical Forms from Parse Trees, using WK

S

NP VP

• We also need some WK stored in a KB, and for this we shall use knowledge linked to verbs.

- use knowledge of relationships, constraints and defaults for this verb - a case frame for each verb:

A recursive top-down algorithm:

- the semantic representation of a node in the parse tree is the combination of the semantic representations of the node’s daughter nodes…

object unspecified man

instrument dog#1’s teeth

“The” some dog

COMBINE

NOTES:

• we did indeed find a consistent meaning for the sentence

The Usefulness of Class Hierarchies

• It is not surprising that it also turns out to be an extremely useful technique in the modelling of knowledge.