A transputer-based parallel machine translation system for Indian languages

This paper presents the performance studies carried out in the design of a parallel machine translation system (MTS). It has been implemented in the language Occam-2 on a PC AT hardware platform augmented with a 3-transputer add-on card. Exploitation of parallelism and a good dictionary organization are identified as the two ways of providing an efficient solution. Parallelism in the algorithm for the morphological and phrase-level analyses and in the generation phases of translation have been fully utilized. The performance of the MTS is discussed for various methods of task distribution and the way the dictionary is accessed. The performance studies have also been conducted on the MTS with and without load sharing among the transputers. The results show that the system performance is nearly doubled with the addition of the transputer card and with the transputers working under no-load sharing condition. © 1997 Elsevier Science B.V. All rights reserved.