daniel bigham.ca

Story 3: My wife's name is Meredith
July 15, 2008

Story: My wife's name is Meredith
Question: What is my wife's name?
Answer: Meredith

Transformation 1

{noun}'s {noun} -> $1.$2

Application:

wife's name -> wife.first_name

Note: This uses the following word/entity mapping:

"name" -> first_name

Transformation 2

From story 1:

my {noun} -> speaker.$1

Application:

my [wife.name] -> speaker.wife.name

Required knowledge about the world:

Before we can apply this transformation, we need to verify that speaker.wife makes sense. This requires the following knowledge about the world:

speaker is_a person
0.5: person is_a man
0.5: person is_a woman
0.4: man has:0 wife
0.6: man has:1 wife

Transformation 3

From story 1:

{noun} is {word}

Application:

speaker.wife.name = "Meredith"

Story 2: I am 27
July 15, 2008

Story: I am 27
Question: How old am I?
Answer: 27

Transformation 1

I am {#}. -> speaker.age = $1

We introduce a special notation for referring to numbers: {#}

We also introduce the use of a period in the input specification to imply that it represents a whole statement. This prevents it from being used to parse the following: "I am one step away from grounding you for the whole week!".

Transformation 2

How old am I -> speaker.age

For the time being we are using a very simplistic parsing strategy for a statement like this. But this highlights how we can progress in a very iterative manner, starting with simplistic transformations and revisiting them later as needed.

Alternative transformations

The following represents another, more generic, possibility:

How old {noun} -> $1.age

am I -> speaker

Stories
July 15, 2008

The basis for interacting with a basic AI is to tell it stories and then to ask questions to evaluate its understanding. I say stories, but in the most trivial sense, a story is just a simple statement about the world. Our first story is as follows:

My name is Daniel

Our question that follows is:

What is my name?

And the expected answer is:

Daniel

More advanced stories might consist of several statements and several questions, each question having an "expected answer".

In this way, stories act as unit tests and provide a method of regression testing. Progress is made by giving an AI the ability to understand new stories without breaking its understanding of previous stories.

older >>