Evaluating Word Embedding Models: Methods and Experimental Results

  • 2019-01-28 16:33:25
  • Bin Wang, Angela Wang, Fenxiao Chen, Yunchen Wang, C. -C. Jay Kuo
  • 7

Abstract

Extensive evaluation on a large number of word embedding models for languageprocessing applications is conducted in this work. First, we introduce popularword embedding models and discuss desired properties of word models andevaluation methods (or evaluators). Then, we categorize evaluators intointrinsic and extrinsic two types. Intrinsic evaluators test the quality of arepresentation independent of specific natural language processing tasks whileextrinsic evaluators use word embeddings as input features to a downstream taskand measure changes in performance metrics specific to that task. We reportexperimental results of intrinsic and extrinsic evaluators on six wordembedding models. It is shown that different evaluators focus on differentaspects of word models, and some are more correlated with natural languageprocessing tasks. Finally, we adopt correlation analysis to study performanceconsistency of extrinsic and intrinsic evalutors.

 

Introduction (beta)

None

 

Conclusion (beta)

None