This originally began as a (slightly altered) run-through of the pytorch recurrent neural network tutorial that can be found at https://pytorch.org/tutorials/intermediate/char_rnn_classification_tutorial.html.
However, upon further reflection, I think that this will serve as a good place to dump a variety of initial pilot scripts for the project. We can add a more formal development repo to the GDD org at a later date.
The first phase was simply to recreate the basic tutorial as is and get it up and running. I made some slight changes to the namespace to make it more case-specific/logical, but otherwise left it mostly unaltered. The test was a success and the basic script can be found in tutorial.py.
This script takes multiple .txt files of different surname and language tuples, stored in the languages subdirectory. It then trains a neural network to "guess" what language a name is.
This tutorial deviates from a proper ML exercise in that there is no separation of the data into test and training subsets, so there is no real model validation.
This phase was successfully completed in the initial commit 47f7304aa4d9f4154e16a62c525239a78e073991.
I have compiled an excel dictionary of name and categories tuples. The categories are
- surname (i.e., last names of people in various languages)
- locality
- county
- countries
- continents
- units (i.e., names of geologic units/rock bodies)
I want to see if we can use the same basic model to make good guesses as name-type. Obviously, many places, people, and units share the same name - i.e., a place named after a person or a unit named after a nearby town. I suspect that it won't work, but let's give it a try. It's a pretty hefty training set!
This was successfully completed in commit 4fe675c78d5841444b6ec4cfadc96ab3cc1dcf02 and includes a Jupyter notebook implementation. The concept was found to be solid, but the model was poorly tunable. This is likely because the model was poorly motivated.
We are now exploring different doc2vec implementations. There is already a working doc2vec implementation as part of the xDD infrastructure that we would like to see made accessible through the xDD API.
However, we would like a simple doc2vec implementation of our own that can take a set of documents and assess similarities. In particular, we want to potentially identify duplicates based on "high" similarity.
Move on to neural co-ref