Fast and Efficient Parallel Alignment Model for Aligning both Long and Short Sentences

Abstract

Recently, demand for fast and efficient translation system been widely seen. However, translation model are dependent parallel corpora. However, it is challenging to obtain large parallel corpora for resource starved language such as Kannada-Telugu pair. The existing Giza++ based word alignment and Moses phrase based alignment model are efficient for aligning only short sentences. However, for longer sentence the accuracy of model degrades. For performing alignment for longer sentences, neural based alignment has been presented in recent times. However, these models are trained using fixed vector length. Thus, induces memory and training overhead. For overcoming research challenges, this work presents a parallel alignment model using recurrent neural network (RNN). Further, to utilize memory efficiently and minimize training time parallel execution of RNN under GPU is considered. For improving alignment accuracy presented a cost function by combing statistical and neural based alignment method. Experiment are conducted to evaluate the proposed alignment model in terms of accuracy, Word alignment error (WAE), memory utilization, and computation time.