Work flow includes -
- Environment setup and run needed servers
- Feature generation and saving
- Running specific combinations from report
============================================================================
-
All needed packages are in cs666.yml and also requirements.txt
- conda env create -f cs666.yml
- Download stanfordNLP package from https://stanfordnlp.github.io/CoreNLP/, and unzip the stanford-corenlp-full-2018-10-05 into root path
- Activate stanfordNLP server with below commands or use run.sh
- cd stanford-corenlp-full-2018-10-05;
- java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 5000 -timeout 10000
-
Feature generation and saving For ease of experimenting with features. we generate the features and save them to compressed jsons individually. They are used later to build combination features.
From top level project directory -
2.1. Genetrate baseline features
python -m baseline.features
This should create 10 features.json files for each of A and B's training and testing data.
Move them to obtained_features/baselines.
NOTE: Since you are pulling the project from git they might already be there as they are small files and we are able to push them
2.2. Generate behaviour features
python -m behavior_model.features
This should create 5 features.json files for each of A and B's training and testing data.
Move them to obtained_features/behavior_model.
NOTE: Since you are pulling the project from git they might already be there as they are small files and we are able to push them
2.3. Generate word embedding features
BERT: python -m word_representations.bert
ELMo: python -m word_representations.elmo
Skip-gram: python -m word_representations.skipgram
CBOW: python -m word_representations.cbow
These are large files and couldn't be pushed to git. But inlucded in zipped code.
These json.gz files should be in project top level directory
2.4 Generating BERT Fine Tuned
- Running classifiers and print metrics
From top level project directory -
3.1 Baseline + Logistic regression
3.2 Baseline + Behavior + Logistic regression
3.3 Baseline + Behavior + PCA + logistic Regression
3.4 Word Embeddings + Logistic Regression
pyhton -m hybrid_combinations.combo1_lr
3.5 Word Embeddings + MLP
pyhton -m hybrid_combinations.combo1
For fine-tuning BERT, follow the instruction here
to get the updated language model, and replace the directory in the word_representations/bert.py.
You could run python -m word_representations.convert_tweet to convert the given dataset into the format desired for fine-tuning BERT
3.5 BERT Fine Tuned + LR
python -m word_representations.bert
3.6 All features + LR
3.7 All features + PCA + LR