Rime is a neural network for categorising a pair of words e.g. ["CAT", "HAT"] as RHYMING or NOT RHYMING.
The Rime network is a simple three-layer perceptron.
- 781 input neurons (15 letters x 2 words x 26 letters + 1 bias)
- 20 hidden neurons in 1 layer
- 2 output neurons (rhyming/not rhyming classification)
As such, the network has 15,660 connections.
I wish to experiment using more hidden layers as this may aid the pattern recognition.
The accuracy on the training data was around 98% on the last run.
This training and testing was done using thousands of rhyming and non-rhyming word pairs created from the dataset 'rhymelist.txt'.
It's hard to find a set of rhyming words for input. I ended up creating a tiny one myself, by typing up some rhyming words.
Use of the 'dotheyrhyme' script looks like this:
PS D:\code\rime> node dotheyrhyme people steeple
YES - I believe that 'PEOPLE' rhymes with 'STEEPLE'.
Some of these pairs existed in the training data, and some did not.
PS D:\code\rime> node dotheyrhyme rhyme time
YES - I believe that 'RHYME' rhymes with 'TIME'.
PS D:\code\rime> node dotheyrhyme dinner thinner
YES - I believe that 'DINNER' rhymes with 'THINNER'.
PS D:\code\rime> node dotheyrhyme sight site
YES - I believe that 'SIGHT' rhymes with 'SITE'.
PS D:\code\rime> node dotheyrhyme might bite
YES - I believe that 'MIGHT' rhymes with 'BITE'.
PS D:\code\rime> node dotheyrhyme bite boat
NOPE - I think that 'BITE' does not rhyme with 'BOAT'.
PS D:\code\rime> node dotheyrhyme trout throughout
YES - I believe that 'TROUT' rhymes with 'THROUGHOUT'.
PS D:\code\rime> node dotheyrhyme banana ribcage
NOPE - I think that 'BANANA' does not rhyme with 'RIBCAGE'.
PS D:\code\rime> node dotheyrhyme not rhyming
NOPE - I think that 'NOT' does not rhyme with 'RHYMING'.
PS D:\code\rime> node dotheyrhyme frogs legs
NOPE - I think that 'FROGS' does not rhyme with 'LEGS'.
PS D:\code\rime> node dotheyrhyme chinese lanterns
NOPE - I think that 'CHINESE' does not rhyme with 'LANTERNS'.
Rime was good at memorising pairs from the training set, but didn't appear to very well understand the concept of rhyming.
While there were some pleasant surprises when trying non-training data, there were also plenty of unsuccessful classifications. Among the most embarrassing of these are classifications of clearly non-rhyming word pairs as rhyming word pairs.
PS D:\code\rime> node dotheyrhyme range strange
NOPE - I think that 'RANGE' does not rhyme with 'STRANGE'.
PS D:\code\rime> node dotheyrhyme grass brass
NOPE - I think that 'GRASS' does not rhyme with 'BRASS'.
PS D:\code\rime> node dotheyrhyme banana apricot
YES - I believe that 'BANANA' rhymes with 'APRICOT'.
PS D:\code\rime> node dotheyrhyme quite interesting
YES - I believe that 'QUITE' rhymes with 'INTERESTING'.
What might help with these issues:
- A larger dataset for training (and testing)
- A deeper network to better pick apart the patterns of rhyming words
- Language: JavaScript
- Environment: Node.js
- Libraries: Synaptic
- Graph: Microsoft Excel
Neural network built and trained using Synaptic by Cazala.
I experimented some more with various settings and network constructions. Using a network with 781 input --> 5 hidden --> 5 hidden --> 2 output, I was able to achieve 100% accuracy on the training set with some consistency. There was a noticeable improvement in classifying non-rhyming word pairs. However, this does not really improve the results classifying non-training rhyming word pairs - it is still apparently just memorising the training data. A better training data set could make this a much less noticeable issue.