Description of team member contribution

Group

Azfar Imtiaz

Help out with data reading tasks, write data to pickle files (obsolete now)
Write function to get vectors for words from the Word2Vec model (with support for padding and unknown tokens)
Write function to generate sentence vectors
Write function to generate batch vectors at train/test time
Write script to fine-tune the BERT model (using the pytorch_pretrained_bert module) on the training data
Fix bug of mismatch between maximum sentence length and number of epochs used for training BERT
Save BERT model and tokenizer to disk, using the rather convoluted method of saving the model as .bin file, saving the config file, and packaging them into a .tar.gz file for the former
Write script for evaluating the BERT model on testing data
Correct and improve structure for RNN network
Combine RNN hidden states and feed them into classifier network
Write the updated script for training the RNN models and classifier model on training data
Create RNN network, attention network and classifier network for bonus part
Incorporate support for attention model in the updated training script for RNN
Fix bugs in the prediction part of the RNN and classifier models
Add support for equalizing class counts in training data
Add support for using an additional amount of records for the SAME class in training data
Update config file to contain all sorts of parameters for model training and testing, such as whether to use attention or not, whether to use equalized class counts in training or not, batch size, hidden size etc for all three model types.
Train all models on 100,000 records of training data, evaluate them on the testing data, and report results to be put into the README
Add details into README about the network structure, motivation behind using attention mechanism, and information about the function of each script
Bug fix for using command line arguments, add check to ensure that the test size ratio provided is between 0.0 and 1.0

Elin Hagman

Reading and parsing XML file, determine speaker names and boundaries
Replacing British words and compound words to their American equivalent
Removing weird tokens
Writing sentences and boundaries to .csv file
Optimized training script and RNN models for GPU
Created testing script for RNN models and classifier
Created testing script for RNN models and models with attention
Documented all changes and decisions in report
Documented results of trained models

Sandra Derbring

Creating a dataloader for the dataset
First version of train test splitter
Post processing script, check tokenization and concatenate lines
Re-create sentences and boundaries from the post processing
Supervising that the various post processing functionalities didn't affect the data in a bad way
Built structure of first RNN network model
First training script for RNN network model
Creating variables for the RNN network model in the config
Enabling command line options for the train/test splitting
Contributed to the README report