Building an Italian-Chinese Parallel Corpus for Machine Translation from the Web

In an increasingly globalized world, being able to understand texts in different languages (even more so in different alphabets and charsets) has become a necessity. This can be strategic even while moving and travelling across different countries, characterized by different languages. With this in mind, bilingual corpora become critical resources since they are the basis of every state-of-the-art automatic translation system; moreover, building a parallel corpus is usually a complex and very expensive operation. This paper describes an innovative approach we have defined and adopted to automatically build an Italian-Chinese parallel corpus, with the aim of using it for training an Italian-Chinese Neural Machine Translation. Our main idea is to scrape parallel texts from the Web: we defined a general pipeline, describing each specific step from the selection of the appropriate data sources to the sentence alignment method. A final evaluation was conducted to evaluate the goodness of our approach and its results show that 90% of the sentences were correctly aligned. The corpus we have obtained consists of more than 6,000 sentence pairs (Italian and Chinese), which are the basis for building a Machine Translation system.