UW-BHI at MEDIQA 2019: An Analysis of Representation Methods for Medical Natural Language Inference

  • 2019-07-09 16:47:50
  • William R. Kearns, Wilson Lau, Jason A. Thomas
  • 3

Abstract

Recent advances in distributed language modeling have led to largeperformance increases on a variety of natural language processing (NLP) tasks.However, it is not well understood how these methods may be augmented byknowledge-based approaches. This paper compares the performance and internalrepresentation of an Enhanced Sequential Inference Model (ESIM) between threeexperimental conditions based on the representation method: BidirectionalEncoder Representations from Transformers (BERT), Embeddings of SemanticPredications (ESP), or Cui2Vec. The methods were evaluated on the MedicalNatural Language Inference (MedNLI) subtask of the MEDIQA 2019 shared task.This task relied heavily on semantic understanding and thus served as asuitable evaluation set for the comparison of these representation methods.

 

Quick Read (beta)

loading the full paper ...