Replicating SelfExplain

A coursework project for CS 6957 (NLP w/ Deep Learning) that replicates a part of Rajagopal et. al's SelfExplain architecture. This text classifier explains what constituent concepts from the input contributed to a prediction. Parse trees were created to obtain concepts and a layer for classifying representations of sentences without certain concepts. A large difference from the actual prediction scores meant a concept was relevant. With the RoBERTa model, the SelfExplain model achieved 94.8% validation accuracy on the SST2 dataset with scores for relevant concepts.