Overview
For another comparison, I’ve been looking for a way to replace the nGrams with another way of turning a document into a vector. Based on word frequency instead of nGrams, I’ve run a number of tests to see how the accuracy and speed of the algorithm compares for the two.
nGrams
I still intend to look into why the Tragedy of Macbeth does not stay with the rest of Shakespeare’s plays. I still believe that it is because portions of it were possible written by another author.
The old GUI framework just wasn’t working out (so far as adding new features went). So, long story short, I’ve switched GUI layout.
As expected, I’ve decided to change libraries from the** **poor results with the original tests may have been a direct results of a misunderstanding with the code base. I think that the layers were not being hooked up correctly, resulting in low/random values.
After my meeting yesterday with my thesis advisers, I have a number of new ideas to try to improve the efficiency of the neural networks. The most promising of those are described below.
Sliding window
The first idea was to replace the idea of applying the most common frequencies directly with a sliding window (almost a directly analogue to the nGrams themselves). The best way that we could come up to implent this would be to give the neural networks some sort of memory which brought up recurring networks (see below).
Overview
For now, I’ve chosen to work with C# Neural network library. It was the easiest to get off the ground and running, so it seemed like a good place to start.
I’ve been looking for a good Neural Network library to use with the AnnGram project and so far I’ve come across a couple of possibilities:
The top link on Google was an aptly named C# Neural network library. Overall, it looks clean and easy to use and is licensed under the GPL, so should work well for my needs. The framework has two types of training methods: genetic algorithms and backward propagation. In addition, there are at least three different activation functions included: linear, signmoid, and heaviside functions. The main problem with this framework is the spare documentation. The only that I’ve been able to find so far is a generated API reference and a few examples (using their included GUI framework).
Overview
Basically, I got tired of modifying the command line every time I wanted to test new values. To that end, I spent a small bit of time coding up a GUI to make further experiments easier.
Overview
The first algorithm that I’ve chosen to implement is a simple cosine difference between the n-gram vectors. This was the first method used in multiple of the papers that I’ve read and it seems like a good benchmark.
Essentially, this method gives the similarity of two n-gram documents (either Documents or Authors) as an angle ranging from 0 (identical documents) to \pi/2 (completely different documents). Documents written by the same author should have the lowest values.
