Feiyan Wang
Professional terminology is not translated with enough credibility in specialized situations because there is little information available to the domain and very ambiguous semantics. To overcome it, in the given research first a lightweight knowledge graph with respect to a particular domain is built that allows combining term definitions, hierarchical relations, and cross-linguistic alignment data as triples. A knowledge-aware context encoding process is then created which dynamically constructs a knowledge context vector which combines the graph semantics of each input term via graph neural network. Also, it is suggested that a multi-source information fusion decoding strategy should be used, and learnable gating should be used to synchronize the attention of text and knowledge-constrained attention at the decoding stage. Lastly, the concept of adversarial consistency training is presented to increase the pronunciation of the special and non-special words in the model. Experimental findings indicate that the framework proposed translates terminologies with an accuracy of 89.3% which is an improvement of 15.1 and 7.7 percentage points over the base model, respectively. The accuracy of the method is 75.3 and 78.6 in low-frequency term (occurs once) and highly ambiguous term (>=4 concepts) tests, respectively which proves the great benefits of this technique in enhancing the accuracy and strength of translating the terms.