Multi-Hop Question Answering using GCNs

Summary

  • Formulated a graphical model allowing multi-hop reasoning wherein context across several documents was incorporated
  • Modeled graph connecting coreferences and nouns within documents and named entities and co-occurrences across documents
  • Developed a model with hierarchical attention that gave an accuracy score of 60% on Wikihop dev-set, in top 5 positions at the time

Related