Summary

International Symposium on Nonlinear Theory and Its Applications

2022

Session Number:B1L-E

Session:

Number:B1L-E-03

Investigation of the Influence of Datasets on Image Generation Using Sentence-BERT

Masato Izumi ,   Kenya Jin'no,  

pp.252-255

Publication Date:12/12/2022

Online ISSN:2188-5079

DOI:10.34385/proc.71.B1L-E-03

PDF download (985.8KB)

Summary:
We verified the degree to which sentence vectors, which are distributed representations of sentences generated by Sentence-BERT, capture the meaning of sentences using k-means and UMAP, and have confirmed that the sentence vectors generated by Sentence-BERT capture the meaning of sentences extremely well. In this article, we examine image generation from sentence vectors generated by Sentence-BERT to see if it is possible to generate images that match the meaning of the sentences. We conduct some experiments using various datasets to investigate the differences in training results among datasets in the Sentence-BERT sentence vector-based image generation model. We then discuss the optimal dataset for the Sentence-BERT sentence vector-based image generation model.