Independent Study

This is a repository containing the work for my Independent Study under professor Andrew McCallum in IESL.

Task

At a high level I was involved in the task of Shallow Semantic Parsing for Scientific Procedural Texts. Within this I was responsible for implementing and running a supervised learning model using PyTorch. Apart from a function or two, majority of the code has been written from scratch.

Data

TO DO

Formatting

TO DO

Requirements

• Python >= 3
• PyTorch >= 1.0.1
• Scikit-learn >= 0.18.1

Configuration

The configuration file must be defined as a single json dict. The following are the fields for the configuration file along with their requirement statuses.

Field	Description	Type	Restriction	Required
TRAINING_FILE_PATH	Path to the file where the training data is stored	String	File Must Exist	Always
DEV_FILE_PATH	Path to the file where the development data is stored	String	File Must Exist	Always
TEST_FILE_PATH	Path to the file where the testing data is stored	String	File Must Exist	Always
MODEL_FILE	Path where the model is supposed to be dumped/loaded from depending on the switches	String	If in test mode File Must Exist	Always
TRAIN_MODE	A switch specifying whether the model should train on data or not	Boolean	None	Always
NUM_EPOCHS	Number of passes to make over training data while training	Integer	>= 10	Only if in train mode
DROPOUT_P	Dropout probability while training	Float	In range [0,1]	Only if in train mode
LEARNING_RATE	Learning rate for training the model (step size)	Float	In range (0,1)	Only if in train mode
EMBED_SIZE	Embedding dimension for Predicates and Arguments (same size used for both)	Integer	>= 20	Only if in train mode
LINEARITY_SIZE	Linear mapping dimension for the concatenated predicate argument embedding	Integer	< 2 * EMBED_SIZE	Only if in train mode
BATCH_SIZE	Number of passes to make over training data while training	Integer	>= 5	Only if in train mode
CHECK_EVERY	Number specifying the frequency of running model on dev set while training. Model is run on dev set every CHECK_EVERY iterations	Integer	>= 3	Only if in train mode
TEST_MODE	A switch specifying whether the model should predict on data or not	Boolean	None	Always
DUMP_TEST_PREDS	Switch to specify whether write the predictions made on test set to a file (predictions.txt)	Boolean	None	Only if in test mode
RUN_BASELINES	Switch to specify whether to run baseline models or not	Boolean	None	Only if in test mode

Notes:

1. One out of TRAIN_MODE or TEST_MODE must be switched on in order for the code to do something. If this violated an error will be thrown.
2. If the paths are relative, they must be so with respect to driver.py's directory.

For an exact example for formatting this config as a json file have a look at src/config.json

Usage

Once the configuration has been properly defined, (for ease) place it in the same directory as driver.py. Finally, to run the program, simply navigate to the location of "driver.py'' and run the following command:

python driver.py config.json