Presenting A New Method for Cross-lingual Processing

Abstract

Machine Learning (ML) was most common approach for data classification. In this method, data classified by algorithm and make prediction, But when classifier learn on one domain data and try to classify on different domain of data, algorithm face domain shift problem. In order to solve domain shift problem, Transfer Learning (TL) and Domain Adaptation (DA) commanded. In NLP which is one of Artificial Intelligence (AI) branch, researcher always faced the shortage of labeled data for language processing. English language has great resource of labeled data but other languages don&rsquot. In order to classify a low resource language by using English language data. We propose an approach AHCL base on TL, which learn knowledge in domain A and transfer that knowledge to domain B and classify B domain unlabeled data. AHCL has three-phases first extract features, second classify source data and third try to discriminate two language. At first extracted source data send to classifier to create a model to use in second phases, then discriminator, discriminate two languages. If it discriminate well use penalty parameter in order to discipline the model.At third phases, by using knowledge of the first and second phases modify feature extractor model. Average accuracy of AHCL on 5 run is 45.91% where it out perform other state-of-art method with 3.42% improvement.

Key Words : Cross Lingual, Adversarial Learning, Cross Lingual Processing, Sentimental Analysis.