The goal of this project is to use modern NLP deep learning techniques to create a knowledge embedding that has properties of graphs while still being easily integratable with NLP applications. The resulting knowledge graph theoretically can improve the performance on NLP reasoning tasks when used in conjunction with graph neural networks that has MAC (Memory, Attention, and Composition) capability.
- Use encoder-decoder Transformer to perform sentence reconstruction. (In progress)
- Insert and train a module in between the encoder and decoder that will embed the knowledge graph.
- Use the knowledge graph on reasoning tasks.