Decoding Word Sense with Graphical Embeddings

Word Sense Disambiguation

• This project was sponsored by the Discover Grant of Summer 2019 at the University of Rochester.

• I was supervised by Prof. Aaron Steven White from the Formal And CompuTational Semantics lab (FACTS.lab)

• Implemented three deep graph encoders over WordNet for word sense embeddings bounded by hypernyms-hyponyms and meronyms-holonyms: a child-sum graph bi-LSTM, a matrix decomposition method with graph kernels, and a GCN.

• Trained a sequence-to-sequence model with Von Mises-Fisher loss for word sense disambiguation (WSD); improved the model to work on unknown words by incorporating the Supersense relations from WordNet. The best model achieved a 75% accuracy on the Universal Decompositional Semantics Word Sense dataset.

• Result Visualization

  • Graph bi-LSTM over the WordNet (depth = 3)
    
    
    
    
    

The code can be found here.