This paper describes a new learning method for acquisition of word translations from small parallel corpora. Our proposed method, Local Focus-based Learning (LFL), efficiently acquires word translations and collocation templates by focusing on parts of sentences, not on entire sentences. Collocation templates have collocation information to acquire word translations from each sentence pair. This method is useful even when frequency of appearances of word translations is very low in sentence pairs. The LFL system described in this paper extracts Ainu-Japanese word translations from small Ainu-Japanese parallel corpora. The Ainu language is spoken by the Ainu ethnic group residing in northern Japan and Sakhalin. An evaluation experiment indicated that the recall was 57.4% and the precision was 72.0% to 546 kinds of nouns and verbs in 287 Ainu-Japanese sentence pairs even though the average frequency of appearances of the 546 kinds of nouns and verbs was 1.98.
[1]
Hiroaki Saito,et al.
Extracting Word Sequence Correspondences with Support Vector Machines
,
2002,
COLING.
[2]
I. Dan Melamed.
A word-to-word model of translational equivalence
,
1997
.
[3]
Hiroshi Echizen-ya,et al.
Study of Practical Effectiveness for Machine Translation Using Recursive Chain-link-type Learning
,
2002,
COLING.
[4]
Vasileios Hatzivassiloglou,et al.
Translating Collocations for Bilingual Lexicons: A Statistical Approach
,
1996,
CL.
[5]
Kevin McTait.
Linguistic Knowledge and Complexity in an EBMT System Based on Translation Patterns
,
2001
.
[6]
Satoru Ikehara,et al.
Learning Bilingual Collocations by Word-Level Sorting
,
1996,
COLING.