Mapping Distributional Semantics to Formal Concept Lattice-based Property Norms

EasyChair Preprint 1157

Abstract

Distributional models characterize the meaning of a word by its observed contexts. They have shown great success in many natural language processing tasks, however they are unable to differentiate clearly between different semantic relations. In cognitive psychology, a word is represented by its relations with properties. In this work, we propose that the mathematical structure of formal concept lattice (FCL) can be attached to property-based concepts in the property norm space to model the conceptual hierarchies. The k-nearest neighbors (KNN) method is then used to build a mapping from a distributional semantic space onto a FCL-based property space automatically for predicting property norms of unknown concepts. We evaluate our method on word embeddings learned with different types of contexts and demonstrate the potential of learning large-scale property-based concept representations from a modest-sized human-annotated perceptual data.

Keyphrases: Formal Concept Lattice, Property Norms, distributional semantics

@booklet{EasyChair:1157,
  author    = {Dandan Li and Douglas Summers-Stay},
  title     = {Mapping Distributional Semantics to Formal Concept Lattice-based  Property Norms},
  howpublished = {EasyChair Preprint 1157},
  year      = {EasyChair, 2019}}