Recognizing Japanese Historical Cursive with Pseudo-Labeling-aided CRNN as an Application of Semi-Supervised Learning to Sequence Labeling

摘要

Pseudo-labeling is semi-supervised learning to learn unlabeled data as well as labeled data by predicting labels of unlabeled data. As far as we know, pseudo-labeling is applied to the task to predict a category label. In this paper, we apply pseudo-labeling to sequence labeling which is a task to predict a sequence of labels for sequential data such as texts. To predict the pseudo-label of unlabeled data, we first initialize a representative sequence with one of the sequences inferred by multiple instances of a deep neural network by measuring edit distances among them. Then, to make the sequence more natural, the representative is refined by a local search that aims to minimize a metric defined as a linear combination of perplexity and average edit distance. To show the effectiveness of our method, we focus on recognizing Japanese historical cursive. Experimental results on ALCON2017 and Kaggle competition show that our method outperformed most of the prior works. Our method reduced the character error rate (CER) at most around 10%.