Guiding ontology translation with hubness-aware translation memory

Mingjie Tian, Fausto Giunchiglia, Rui Song, Hao Xu

Abstract: Ontology, as the foundational architecture for knowledge representation, necessitates multilingualization to facilitate cross-lingual knowledge sharing, posing challenges that require a domain-specific translation system capable of producing credible translations due to its specialized vocabulary and limited context. This work presents an approach to model and inject reference knowledge to enhance ontology translation by integrating translation memory-augmented neural machine translation, aiming to solve the problem of insufficient coverage that traditional neural machine translation cannot solve. Firstly, this work enhances TM retrieval from monolingual to cross-lingual, utilizing meaning equivalents in parallel data to enrich bidirectional context. Secondly, this work proposes a novel retrieval measurement to mitigate the hubness issue that occurs in similarity-based greedy search retrieval. Furthermore, this work introduces a cross-lingual agreement matching task and an adversarial learning task to enhance the retrieval and translation models in translation memory-augmented neural machine translation. Experimental results and analysis demonstrate the effectiveness of the proposed approach, outperforming strong baselines across four domains’ ontology datasets in two language pair directions.