Meaning Representation

Definition: semantic analysis — [in the "representational" approach taken in JM] mapping semantic representations to linguistic inputs.

Representation schemes

• Propositional logic
• Predicate logic (Wikipedia)
• Prolog examples.
• Semantic Networks

  Defined.

• Conceptual Dependency Diagrams (aka Concept Maps)

  Defined.

• Frames

  Defined.

• Example — representation schemes.

Computational Desiderata of a Representation Scheme [JM, Section 14.1]

[JM] suggest five requirements of a representation scheme:

1. Verifiability
2. Unambiguous representation
3. Canonical form
4. Inferences and their implications in the use of variables
5. Expressiveness

Verifiability

The representation of meaning must provide the ability to determine the relationship between meaning and the world as we know it. [JM] use the follow sample sentence:
14.1 Does Maharani serve vegetarian food?
as an example for discussion.

This can be "glossed" as mapping to Serves(Maharani,VegetarianFood). [What representation scheme is being used?]

It can be further recognized that there is required someway of comparing the representation scheme used against the "world knowledge" stored in the interpretation system's knowledge base.

This correlation of representation schemes is relevant to ontologies.

A project which combines the idea of ontology and world knowledge is Cyc.

Unambiguous Representation

There are two aspects of ambiguity

• semantic ambiguity — inherent in the language
• representational ambiguity — caused by interpretation of the representation scheme

Sample sentence:
14.2 I wanna eat someplace that's close to ICSI.

Example of semantic ambiguity — eat at someplace vs literal eat as Godzilla would.

A related term is vagueness, which is not always distinguishable from ambiguity. A representation needs to be able to express different levels of vagueness.
Sample sentence:
14.3 I want to eat Italian food.

Here, the term Italian food is what the some authors call vague, rather than ambiguous. [JM]

Canonical forms

Contrasting the "one-to-many" problem of one sentence ambiguously mapping to more than one knowledge representation, i.e., having several possible meanings with its converse problem:
The "many-to-one" problem, where multiple sentences should have the same meaning, and hence, map to the same representation.

Example — unique file names.

[JM] models canonical form requirements with example sentences which, perhaps, should map to a single meaning representation.

Sample sentences:
14.4 Does Maharani have vegetarian dishes?
14.5 Do they have vegetarian food at Maharani?
14.6 Are vegetarian dishes served at Maharani?
14.7 Does Maharani serve vegetarian fare?

The concept of word sense disambiguation is described as the process of determining the meaning of each element in the above sentences, facilitating the mapping of all sentences to the same meaning.

Inferences & variables

Sample sentence:
14.10 Can vegetarians eat at Maharani?

[JM] suggests that no approach to canonical form can be used to tie the meaning of this sentence (and, hence, its answer) to the previous sentences (14.4 - 14.7.) Rather, a meaning representation which involves variables may be needed to allow this sentence to map to meanings of the previous examples.

Logic Form is a FOL designed for NLP.

Each noun, verb, adjective, adverb, pronoun, preposition and conjunction generates a predicate. Logic forms can be decorated with word senses to disambiguate the semantics of the word. [Wikipedia].

Example from Wikipedia.

Input:  The Earth provides the food we eat every day.
Output: Earth:n_#1(x1) provide:v_#2(e1, x1, x2)
        food:n_#1(x2) we(x3) eat:v_#1(e2, x3, x2; x4) day:n_#1(x4)

Expressiveness

The requirement of expressing a wide range of meaning and the desire for simplicity and consistency suggests use of the "standard" knowledge representation methodologies.