Improving machine translation evaluation using language learner evaluation techniques.

摘要

This thesis shows that machine translation evaluation (MTE) benefits from research performed for computer-assisted language learning and second language acquisition, particularly from the area of language learner evaluation. MTE has lacked organization and principled approaches, a situation remedied in part by this dissertation. The process of organizing MTE shows commonalities with second language learner evaluation (LLE), particularly in processing ill-formed input and first language interference. These commonalities facilitate the successful application of LLE testing techniques to MTE problems. To demonstrate this, a series of experiments has been developed which use language testing research results and techniques in MTE. The first experiment shows the feasibility of adapting techniques from LLE to MTE. The second demonstrates a method for assessing the intelligibility of MT output. The third provides a new metric for measuring the adequacy of a translation's semantic content. The outcomes show the efficacy of using LLE testing, appropriately modified, in MTE. They also provide a basis for recommending a novel, effective, more automated approach to MTE.