Spelling reveals a lot about humanity: it's messy, inconsistent, and self-contradictory. Despite all of this, people have mental heuristics for spelling and pronouncing new words. I want to see how well neural networks can learn those same heuristics.
The goal of this project is to train recurrent neural networks on two tasks:
- Spelling: converting phonetic transcriptions to an English spelling.
- Pronunciation: converting English spellings to phonetic transcriptions.
After training for one epoch, here are some results:
|Word||Pronunciation||Network Spelling||Network Pronunciation|
First, you should install and configure Go. Make sure your GOPATH is setup.
If you don't want to setup Go yourself, you can use the Docker image for Go. It has everything that you will need:
$ docker run -it golang:1.7 /bin/bash
Downloading the code
Next, download the code as follows:
$ go get -d -u github.com/unixpickle/neuralspell/...
The repository includes sub-directories for various commands. The train command is the first thing you will want to use. After that, you can use the eval command to run the network on a new word or phonetic transcription.
$ cd $GOPATH/src/github.com/unixpickle/neuralspell/train $ go run *.go
This will train a new network on the spelling task. Pressing Ctrl+C will gracefully stop training and save the result to a file called
out_net. Make sure not to press Ctrl+C more than once. To train on the pronunciation task, add
$ go run *.go -task pronounce
I recommend training for at least an hour per task. See
-help for more information on training.
To evaluate the network on new inputs, you can use the
eval command. Note that you must already have trained the network on the task at hand.
$ cd $GOPATH/src/github.com/unixpickle/neuralspell/eval $ go run *.go -phonetics dɔg Spelling: dog $ go run *.go -spelling dog -task spell Pronunciation: dag
If you have not trained the network, the command will probably take a long time to run. This is because, to decode the output of the network,
eval uses a technique called prefix search decoding. If the network was not trained very much, prefix search decoding has to search a vast array of possible decodings.
How it works
The technical side of the project is fairly unoriginal. I use an architecture similar to one that might be used for neural speech recognition. In this architecture, a bidirectional-RNN "reads" the input and produces a labeling using Connectionist Temporal Classification.
There were other implementation routes I could have taken. I think it would be particularly interesting to have two "encoder" networks that convert spelling/phonetics to a universal vector representation, and then have two "decoder" networks turn said representation into spelling/phonetics.