README.md

TechTrack Dataset

1. Directory Structure
   1. Setup 1: Training each property Individually
   2. Setup 2: Training the Combined Model
   3. Setup 3: Training on first n steps of procedure
2. How to use it?

Directory Structure

Setup 1: Training each property Individually

In our first setup, we trained the models for each property individually.

./setup_1/bert/ contains the data for BERT model and ./setup_1/prolocal/ for ProLocal model.

Setup 2: Training the Combined Model

Along with individual models for each property, we also train a combined BERT-based classifier to predict property values for each property type. Here, we train a single model on multiple properties, with each property trained on respective training rows.

./setup_2/bert/ contains the data for this setup.

Setup 3: Training on first n steps of procedure

To give BERT model context of previous steps while training, we extend embeddings by training BERT with first n steps for every n^th step in the procedure. We do this only for state-type properties, since for event-type properties, only that particular step for which the prediction is being made is causing the state change. Model is trained to predict the actual state of entity at every step, individually for each property.

./setup_3/bert/ contains the data for this setup.

How to use it?

All the files for the desired setup has to be placed in data/Inputs. This can be done manually or use the scripts provided in the root. For more info, see the root REAMDE.