Effective Grammatical Error Correction with Neural Machine Translation Techniques

Shubha Guha, M.Sc. Artificial Intelligence

University of Edinburgh, 2016-2017

Formal Writing

• Informatics Research Proposal
• Mid-term Progress Report
• Final Master's Thesis

Informal Notes & Documentation

• Development Notes
• Meeting Notes

Submission Summary

All original code can be found under scripts/. Most important to note:

• modified_nmt.py

  This is my implementation of training for my advanced models and should be compared to nematus/nmt.py under the Nematus project (https://github.com/EdinburghNLP/nematus/) at commit 73037e94884fd2d1c1d18d81686cd1f6ea32d073.

• generate_edit_vectors.py

  This is my implementation of using the MaxMatch aligner to extract edit vectors. The file created by this script is passed in as the value of command line argument --edit_vectors in modified_nmt.py.

Also relevant:

• preprocess_data.sh

  This is how I preprocess training, validation, and test data.

• evaluate_model.sh

  This is how I evaluate a trained model.

• modified_data_iterator.py

  This is a modification of nematus/data_iterator.py to handle the additional input of edit vectors.

Not original code:

• levenshtein.py

  This is directly from the CoNLL 2013 shared task scripts for the M2 scorer. It is necessary to run generate_edit_vectors.py.

All other code files are to set up the environment for training remotely and for convenience during development and testing.