Ampy: A Python AM implementation

Paul Huff

Linguistics 615

Motivation

• Implementing the algorithm for myself gives me a clearer understanding of how it works.
• Python is a clean, object oriented language which allows for some interesting re-tooling of the basic algorithm.
• Programming is fun!

Context

• Statistical classifiers in general try and decide, based on a set of examples, how a new item best fits in the mix.
• AM does this in a rather novel way (e.g. compare k-NN).

Design principles

• Build a clean, easy to understand implementation.
• Use a good object model so that overloading and manipulation are easy.
• Make it hackable.

The basic algorithm

1. Snarf the examples, putting each one in several big containers.
2. For each test set member:
   1. Find the analogical set (the list of containers we need to look in).
   2. Tally up the votes.
   3. Announce the winner

Considerations

• This is the stupid version of AM (where stupid = slow).
• However, it allows for improvements to the process.
  • Weighted Analogical Modeling (WAM) (explain on board)
  • Map-Reduce: perhaps easier with this approach?
  • Abstract Analogical Modeling, Weighted Abstract Analogical Modeling (future direction)
• Can get rid of the niggling bits (command line options, gui, etc.).
• Open source project, ampy.sf.net (this presentation), sf.net/projects/ampy for downloads, etc.