Investigation of negation effect for English–Assamese machine translation

Computational linguistics deals with the computational modelling of natural languages, in which machine translation is a popular task. The aim of machine translation is to automatically translate one natural language into another, which minimizes the linguistic barrier of different linguistic backgrounds. The data-driven approach of machine translation, namely, neural machine translation achieves state-of-the-art results on different language pairs, however it needs a sufficient amount of parallel training data to attain reasonable translation performance. In this work, we have explored different machine translation models on a low-resource English–Assamese language pair and investigated different sources of errors, particularly due to negation in English-to-Assamese and Assamese-to-English translation. Negation is a universal, essential feature of human language that has a substantial impact on the semantics of a statement. Moreover, a rule-based approach is proposed in the data preprocessing step which handles modal-verb negation problem that shows significant improvement in translation performance in terms of automatic and manual evaluation scores.