| 摘要 |
A word translation acquisition method comprises: a constraint extraction step that extracts monolingual must-link, monolingual cannot-link and cross-lingual must-link constraints from a bilingual dictionary; a clustering step that clusters words in each language using a word's feature vector such that the monolingual must-link and monolingual cannot-link constraints are obeyed; a cluster representative calculation step that combines feature vectors of the words assigned to the same cluster to form a cluster representative, the feature vectors that are combined to the cluster representative being such that the feature vectors can be compared across languages; an alignment step that aligns the monolingual clusters across languages using the cluster representatives and the cross-lingual must-link constraints given by the bilingual dictionary; and a scoring step that scores each target language cluster by using a probability that the query word is assigned to a source language cluster, and a probability that the source language cluster is aligned to the target language cluster. |
