Exploring the Value of Personalized Word Embeddings

  • 2020-11-11 20:23:09
  • Charles Welch, Jonathan K. Kummerfeld, Verónica Pérez-Rosas, Rada Mihalcea
  • 16


In this paper, we introduce personalized word embeddings, and examine theirvalue for language modeling. We compare the performance of our proposedprediction model when using personalized versus generic word representations,and study how these representations can be leveraged for improved performance.We provide insight into what types of words can be more accurately predictedwhen building personalized models. Our results show that a subset of wordsbelonging to specific psycholinguistic categories tend to vary more in theirrepresentations across users and that combining generic and personalized wordembeddings yields the best performance, with a 4.7% relative reduction inperplexity. Additionally, we show that a language model using personalized wordembeddings can be effectively used for authorship attribution.


Quick Read (beta)

loading the full paper ...