// the find
nusnlp/nea
Neural Essay Assessor: An Automated Essay Scoring System Based on Deep Neural Networks
A Keras implementation of the neural essay scoring model from Taghipour and Ng's 2016 EMNLP paper, trained and evaluated on the ASAP automated essay scoring dataset. It's for researchers and students who want a reference baseline for a CNN or recurrent scorer, not for anyone who needs a working scoring service.
- The 5-fold train, dev, and test splits are committed as ID lists under data/fold_0 through fold_4, so the evaluation setup can be reproduced without re-running a random split.
- The CNN and GRU/LSTM variants share one training script with command-line switches, which makes it a usable side-by-side baseline.
- It includes a quadratic weighted kappa implementation, the metric ASAP uses, so the numbers can be compared against the published results.
- Preprocessing is a separate script from training, and embeddings are read from word2vec-format text files, so swapping in different embeddings is mostly a file change.
- It depends on Keras with the Theano backend. Theano stopped development in 2017, so running this on a current Python and Keras stack means porting it first. The README's THEANO_FLAGS example is also GPU-specific.
- There is no requirements file or pinned dependency versions. The setup section is three bullet points that don't name a Keras version or a Python version.
- The last push was January 2018, and the repo shows no sign of recent maintenance. Treat it as a reference implementation for the paper, not as code to build on.
- It is GPLv3. Shipping it inside a commercial product means either complying with the GPL or buying the separate commercial license the README mentions.