License: Creative Commons Attribution 3.0 Unported license (CC BY 3.0)
When quoting this document, please refer to the following
DOI: 10.4230/OASIcs.SLATE.2020.9
URN: urn:nbn:de:0030-drops-130229
URL: https://drops.dagstuhl.de/opus/volltexte/2020/13022/
Go to the corresponding OASIcs Volume Portal


Sousa, Tiago ; Gonçalo Oliveira, Hugo ; Alves, Ana

Exploring Different Methods for Solving Analogies with Portuguese Word Embeddings

pdf-format:
OASIcs-SLATE-2020-9.pdf (0.4 MB)


Abstract

A common way of assessing static word embeddings is to use them for solving analogies of the kind "what is to king as man is to woman?". For this purpose, the vector offset method (king - man + woman = queen), also known as 3CosAdd, has been effectively used for solving analogies and assessing different models of word embeddings in different languages. However, some researchers pointed out that this method is not the most effective for this purpose. Following this, we tested alternative analogy solving methods (3CosMul, 3CosAvg, LRCos) in Portuguese word embeddings and confirmed the previous statement. Specifically, those methods are used to answer the Portuguese version of the Google Analogy Test, dubbed LX-4WAnalogies, which covers syntactic and semantic analogies of different kinds. We discuss the accuracy of different methods applied to different models of embeddings and take some conclusions. Indeed, all methods outperform 3CosAdd, and the best performance is consistently achieved with LRCos, in GloVe.

BibTeX - Entry

@InProceedings{sousa_et_al:OASIcs:2020:13022,
  author =	{Tiago Sousa and Hugo Gon{\c{c}}alo Oliveira and Ana Alves},
  title =	{{Exploring Different Methods for Solving Analogies with Portuguese Word Embeddings}},
  booktitle =	{9th Symposium on Languages, Applications and Technologies (SLATE 2020)},
  pages =	{9:1--9:14},
  series =	{OpenAccess Series in Informatics (OASIcs)},
  ISBN =	{978-3-95977-165-8},
  ISSN =	{2190-6807},
  year =	{2020},
  volume =	{83},
  editor =	{Alberto Sim{\~o}es and Pedro Rangel Henriques and Ricardo Queir{\'o}s},
  publisher =	{Schloss Dagstuhl--Leibniz-Zentrum f{\"u}r Informatik},
  address =	{Dagstuhl, Germany},
  URL =		{https://drops.dagstuhl.de/opus/volltexte/2020/13022},
  URN =		{urn:nbn:de:0030-drops-130229},
  doi =		{10.4230/OASIcs.SLATE.2020.9},
  annote =	{Keywords: analogies, word embeddings, semantic relations, syntactic relations, Portuguese}
}

Keywords: analogies, word embeddings, semantic relations, syntactic relations, Portuguese
Collection: 9th Symposium on Languages, Applications and Technologies (SLATE 2020)
Issue Date: 2020
Date of publication: 16.09.2020


DROPS-Home | Fulltext Search | Imprint | Privacy Published by LZI