Automatic Method to Build a Dictionary for Class-Based Translation Systems

Kohichi Takai, Gen Hattori, Keiji Yasuda, Panikos Heracleous, Akio Ishikawa, Kazunori Matsumoto, Fumiaki Sugaya

Abstract: Mis-translation or dropping of proper nouns reduces the quality of machine translation or speech translation output. In this paper, we propose a method to build a proper noun dictionary for the systems which use class-based language models. The method consists of two parts: training data building part and word classifier training part. The first part uses bilingual corpus which contain proper nouns. For each proper noun, the first part finds out the class which gives the highest sentence-level automatic evaluation score. The second part trains CNN-based word class classifier by using the training data yielded by the first step. The training data consists of source language sentences with proper nouns and the proper nouns’ classes which give the highest scores. The CNN is trained to predict the proper noun class given the source side sentence. Although, the proposed method does not require the manually annotated training data at all, the experimental results on a statistical machine translation system show that the dictionary made by the proposed method achieves comparable performance to the manually annotated dictionary.

External IDs:dblp:conf/cicling/TakaiHYHIMS18