Extract what foods were eaten from a transcript of a mealtime conversation.
Python 3
The spacy python package, which can be installed with: pip install spacy
Spacy english data, which can be downloaded with: python -m spacy download en
pytorch is required, installation instructions are here Note that every neural net is set to use cuda, this could be adapted with little effort.
The transformers package, which can be installed with: pip install transformers
Run clean_hslld.py to clean all the transcripts (create the folder 'transcripts' for them to be saved to)
Then run parse_json_labels.py to generate evidence data (create the destination 'chunks')
This uses the labels and the clean transcripts to create a set of (text_chunk, food) pairs
Run augment_data.py to augment the data by swapping in different foods to increase the number of true positives (create the destination folder 'chunks_augmented')
Then run generate_evidence_classifier_data.py to convert the evidence data into labeled bert data (create the folder 'bert_evidence'). A true label indicates that the chunk contains evidence of a food
If using augmented data, this should also be run on the augmented chunks (create the folder 'bert_evidence_augmented'). The input and ouput path are declared after the imports so these can be set easily.
Run train_test_split.py to create training and desting csv files. Some parameters can be tweaked in this file, like which files end up in the test set, and whether or not the training set is balanced and augmented
At this point you can run train_classifier.py or nn_evidence.py to train a classifier or neural network