Role of examples in translation

Abstract

This paper highlights the role of examples in machine translation. After limited success using rule-based approaches, the process of translation has been examined from an entirely different angle. A person without the knowledge of the complete grammar of a language, attempts to learn a second language by storing some example pair sentences (pair of source and target language sentences) in his/her memory. He/she tries to make use of his/her past experiences, whenever a new event occurs. The same theory has been extended to the machine translation. The system performs the task of translation using its bank of example pair sentences from source to target language. Like a person, the system also acquires the necessary knowledge required for translation in implicit form from example sentences. However, some of the anticipated problems with the example-based system are: the requirement of large example-base due to raw examples (examples in their original form), large matching time which is directly proportional to the size of the example-base, and the complexity in design of a suitable distance function to get the best match. A new approach named HEBMT (Hybrid Example Based Machine Translation) has been proposed by the authors to take care of the above problems. The HEBMT is an attempt to combine the strategies used in the rule-based approach and example-based approach. Examples are stored in their abstracted form instead of raw form. This drastically reduces the size of the example-base. The abstraction is done by identifying the surface level syntactic tokens in the source sentence. Matching is done with similar type of examples only, which makes the matching time directly proportional to the number of partitions and the number of examples of that type.