Process Knowledge Extraction

This project presents two novel techniques to improve existing semantic role representations to enable better understanding of the language. Firstly, We have tried to retrofit word vectors generated from LSTM model with scientific processes corpus to generate better word embeddings. Second technique uses a semi-supervised model which learns word embeddings using role as context. On testing, We found that first model outperforms existing role labeling models for scientific processes. The second model also performs well even for small annotated datasets. We have concluded by suggesting few ideas for further optimizing this model.