Automatic Evaluation of Alignments without using a Gold-Corpus
Abstract
This paper presents a test designed to automatically evaluate the alignment quality of a bilingual aligned corpus without comparing it to a reference/gold corpus. We justify bypassing a reference corpus and present the advantages of the test, particularly for language pairs where few bilingual aligned corpora exist. The test, which is based on the observation of single-translation words, is demonstrated using three bilingual French-Japanese corpora : a manually translated and aligned corpus, a freely translated and semi-manually aligned corpus, and a freely translated and automatically aligned corpus. This paper shows that the results achieved validate the proposed evaluation technique.
Origin : Files produced by the author(s)
Loading...