Improve User Experience on Web for Machine Translation System Using Storm

Abstract

A transfer based Machine Translation (MT) system is a large complex functional application where the job completion time is proportional to job size. When these applications are deployed on web with increasing translation load web user experience degrades. The end user has to wait unusually longer to get his first visible response. Generic layer 3 load balancing techniques does not help to improve the response time as each job is assigned similar compute resources irrespective of job size. This paper presents an engineering approach to deploy MT system on cloud platform using Storm, a distributed computing framework. This scheme, by utilizing the inherent parallelism of a functional application, not only reduces the job completion time considerably but it, also as a web application, gives very good user experience, viz., the first visible response time within an acceptable time limit, and the subsequent responses well before the user finishes perusing the preceding response. Using Storm framework a translation job is split into multiple job partitions and is streamed into the storm cluster such that first job partitions of all jobs are processed before the second partitions, i.e., All nth partitions of all jobs are processed before (n+1) th partitions. Thus machine translation system is able to produce translation output as a continuous stream, sentence by sentence, as soon as each sentence gets translated. The system maintains flow rate of translated sentences stream high enough so that the next translated sentence is produced well before the end user finishes reading the previous sentence, thereby providing very good user experience. There is a class of natural language processing (NLP) applications, viz., machine translation systems, text to speech systems, speech recognition systems, etc., that are functional in nature, and this engineering approach would be equally applicable to them as well.