Name Translation by Enhancing Graph Symmetry
iv, 30 p.
- 원문 URL
This paper studies the problem of name translation. Existing approaches represent name co-occurrences in comparable corpora, using two graphs representing names by nodes and co-ocurrences by edges. Then name nodes across the two graphs are aligned by first identifying seed mappings based on phonetic and contextual similarity, then iteratively propagating translation confidence of seeds to co-occurring nodes. However, this propagation accumulates errors in the presence of asymmetry between graphs. This paper resolves asymmetry by handling coreferences and using temporality. To handle coreferences, we abstract translation as the mapping in the entity layer, between coreference clusters in two languages. Also, we found that graphs in particular period are more symmetric because they consists of temporal names. Our empirical study shows that the F1-score of name translation increases by 15% and 6.1% for each approach compared to baseline.