Squashed Shifted PMI Matrix: Bridging Word Embeddings and Hyperbolic Spaces

Zhenisbek Assylbekov, Alibi Jangeldin

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

We show that removing sigmoid transformation in the skip-gram with negative sampling (SGNS) objective does not harm the quality of word vectors significantly and at the same time is related to factorizing a squashed shifted PMI matrix which, in turn, can be treated as a connection probabilities matrix of a random graph. Empirically, such graph is a complex network, i.e. it has strong clustering and scale-free degree distribution, and is tightly connected with hyperbolic spaces. In short, we show the connection between static word embeddings and hyperbolic spaces through the squashed shifted PMI matrix using analytical and empirical methods.

Original languageEnglish
Title of host publicationAI 2020
Subtitle of host publicationAdvances in Artificial Intelligence - 33rd Australasian Joint Conference, AI 2020, Proceedings
EditorsMarcus Gallagher, Nour Moustafa, Erandi Lakshika
PublisherSpringer Science and Business Media Deutschland GmbH
Pages336-346
Number of pages11
ISBN (Print)9783030649838
DOIs
Publication statusPublished - 2020
Event33rd Australasian Joint Conference on Artificial Intelligence, AI 2020 - Canberra, ACT, Australia
Duration: Nov 29 2020Nov 30 2020

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12576 LNAI
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference33rd Australasian Joint Conference on Artificial Intelligence, AI 2020
Country/TerritoryAustralia
CityCanberra, ACT
Period11/29/2011/30/20

Keywords

  • Complex networks
  • Hyperbolic geometry
  • PMI
  • Word vectors

ASJC Scopus subject areas

  • Theoretical Computer Science
  • General Computer Science

Fingerprint

Dive into the research topics of 'Squashed Shifted PMI Matrix: Bridging Word Embeddings and Hyperbolic Spaces'. Together they form a unique fingerprint.

Cite this