Resumo

PAK, Alexandr et al. Word Embeddings: A Comprehensive Survey. Comp. y Sist. [online]. 2024, vol.28, n.4, pp.2005-2029.  Epub 25-Mar-2025. ISSN 2007-9737.  https://doi.org/10.13053/cys-28-4-5225.

This article is a systematic review of available studies in the area of word embeddings with an emphasis on classical matrix factorization techniques and contemporary neural word embedding algorithms such as Word2Vec, GloVe, and Bert. The efficiency and effectiveness of these methods for mapping semantic and lexical relationships are evaluated in greater detail providing analysis of the topology of these techniques. In addition, this approach demonstrates a model accuracy of 77%, which is 3% below the best human performance. At the same time the study has also shown the weaknesses of some models such as BERT, which lead to unrealistic high accuracy due to spurious correlations in the datasets. We see that there are three bottlenecks for the subsequent development of NLP algorithms: assimilation of inductive bias, common sense embedding, and generalization problem. The outcomes from this research help in enhancing the strength and applicability of word embeddings in natural language processing tasks.

Palavras-chave : anguage models; distributive semantics; word embeddings; natural language processing; deep learning.

        · texto em Inglês     · Inglês ( pdf )