Metanode Composition Method for Multilingual Parallel-text Having Many-to-many Relationship

Parallel texts are sets of example sentences in one language with their unique translations in other languages. Such corpora can be used for accurate multilingual communication. However, parallel texts often have many-to-many relations because of semantic polysemy, which can complicate way to use parallel texts. This problem cannot be easily solved with only the information of association between example sentences. Here, we propose a metanode composition method for parallel-text graphs having many-to-many combinations. The metanodes relate sentences with the same meaning to resolve the many-to-many relations into one-to-one relations. We compare the proposed method with existing approaches and show that it has several advantages that outweigh the increased computational cost.