Peyman Passban - Transfer Talk - 26th April 2016
Title: Machine Translation for Morphologically Rich Languages Using Deep Neural Networks
Supervisor: Prof. Qun Liu, Prof. Andy Way
In our research we address challenges of translating into/from Morphologically Rich Languages (MRLs). Words in MRLs have more complex structures than other languages. Each word can be viewed as a hierarchical structure with several internal subunits, so word-based models in which words are treated as atomic units are not suitable for this set of languages. As a commonly used and efficient solution, morphological decomposition is applied to break up words into atomic and meaning-preserving units but it raises other types of problems which we address. We mainly use Neural Networks (NNs) to perform Machine Translation (MT) in our research, so we deal with difficulties of conventional MT methods as well as problems of NNs. Firstly we try to model Morphologically Rich Words (MRWs) and provide better word-level representations. Words are symbolic concepts which should be represented numerically in order to be used by NNs. The first research question covers this part. Then in the second research question we focus on Language Modelling (LM) and work at the sentence level. In the third research question we tend to use neural information provided by the first two steps to enhance translation quality in the Statistical Machine Translation (SMT) pipeline. Finally in the fourth and last research question, we try to perform end-to-end translation via NNs. Neural MT (NMT) engines have recently improved a lot and perform as well as state-of-the-art systems. Our overall goal is to find the most suitable architecture for translation of MRLs.