Comp 551 Final Project

Yuchen Chen, Yun-Fei Cheng, Jiayao Zhang

Abstract: It has become common knowledge for convolutional neural networks (CNN) to perform well on image data, so, intuitively, CNNs should be used for word vectorization of logographic characters such as those from Chinese. The paper introducesthreeCNN-basedinnovationstoobtainbetterChinesewordembeddings for use in natural language processing (NLP) tasks. Our ablation studies will focus on the first two innovations: “Using historical scripts to enrich the pictographic evidence in characters” and using “CNN structures tailored to Chinese character image processing”. We will verify that it does indeed produce better results. We then generalize and experiment whether the findings in this paper apply to another logographic but subtly different language- Japanese.

Track: Ablation

NeurIPS Paper Id: https://openreview.net/forum?id=Hye3LNSxLSHye3LNSxLS