There are many things to improve in PBMT. For example, there is a trade-off between the coverage of input sentences and the translation quality in PBMT. When we made the ``high probability phrase table'', we set the threshold to 0.1. This was a completely heuristic value. If this value sets low, we obtained many word pairs and many patters. However the reliability of these value was decrease. So we must cut and try this value.
Moreover, there were many bugs in our system. There were 10,000 test sentences in this experiment. Of these 10,000 sentences, 1,143 sentences matched the Japanese-English patterns. We think this number is small for our experience. One possible cause is that we might not have obtained all the possible Japanese-English patterns. We will work on improving the performance of our pattern-based MT system.