Learning Emotion Representations on a Manifold
Qingfeng Zou · 2022 IEEE Conference on Telecommunications, Optics and Computer Science (TOCS) · 2022
Previous works on emotion embedding construct models for emotion categories to analyze the emotion relations, but they neglect the point that the emotion embedding result is actually in the form of a high dimension array. The result being as a high dimension array means that it could also be a manifold of lower dimension existing in such high dimension. Ignoring the assumption of the manifold will distract the actual relation information among emotion categories, and undermine the application effect of emotion embedding on downstream tasks. To solve this problem, we utilize LLE (Locally Linear Embedding), an unsupervised learning algorithm, to capture the manifold characteristics of the emotion embedding. We visualize and analyze the result in 2 dimensions, making it more effective for downstream applications.