A maximum entropy-based sentence simplifier for machine translation

We present a method for removing unnecessary words from sentences to expedite automatic machine translation. Our hypothesis is that the resulting simplified sentences are easier to automatically translate, giving improved translation performance. We evaluate the sentence simplifier in two ways. Firstly the system is tested directly against humans in the word deletion task. The output of our system is evaluated against a set of reference sentences and its performance compared to a test set of human-shortened sentences. We show the system is able to perform at close to human performance on this task. Secondly we evaluate the system when used as a preprocessor to two different machine translation systems. We show that we are able to significantly improve the performance of a machine translation (MT) system based on the publicly available GIZA++ software by pre-processing the input, and make a small improvement to the performance of the more capable ATR translation system.