Sparse Variational Autoencoder-Based Interpretable Bimodal Word Embeddings
Jingyao Tang, Weiyu Zhong, Qianhua Cai, Guojun Lu, Zehao Yan, Yun Xue, Xinguang Li · 2021
Word embedding is a basic task in the field of natural language processing, which is widely applied to a variety of tasks. In spite of delivering the semantic information, there is no meaningful explanation for specific dimension of the word embedding. As such, research is ongoing to explore the interpretability of word embeddings and thus improve their performance in downstream tasks. Current interpretable word embedding models, however, merely focus on the textual information of the words instead of the multimodalities. In line with the cognition principle, we establish a sparse variational autoencoder, which exploits both the textual and visual information to generate interpretable word embedding. Experiments are conducted to verify the interpretability of word embeddings together with their performance in downstream tasks. Comparing to the state-of-arts, experimental results indicate that the proposed interpretable word embeddings not only effectively increase the interpretability, but also obtain better result in most downstream tasks.