Near-Real Time Arabic/English Machine Translation by Integrated Statistical and Linguistic Learning Methods

StreamSage proposes an approach to the automatic translation of Arabic and Arabic dialect texts to and from English that significantly extends the state-of-the-art in regards to the integration of statistical and traditional machine translation techniques. This research will greatly increase translation accuracy while decreasing the need for domain-specific training. The proposed near real time translation system will use automatically induced transfer rules between English and Arabic syntactic structures that have been statistically trained on a feature set that is of unprecedented sophistication. This feature set will be automatically generated through the use of tools that have not before been applied to Arabic machine translation, such as language-wide noun and verb sense disambiguation, a TAG-Based Stochastic Parser, and a hierarchical representation of Arabic dialect morphology, lexical features, and syntactic structures. Additional innovations include the application of state-of-the-art Arabic morphological analysis throughout the translation process, from word sense disambiguation to transfer rule induction to generation, and the automatic induction of syntactic-structure to target language generation rules. This research will make use of past work in machine tranlation, Arabic parsing, Arabic dialect analysis, and word sense disambiguation by StreamSage, Columbia University, and CoGenTex.