Conditional checkpoint selection strategy based on sentence structures for text to triple translation using BiLSTM encoder–decoder model

Understanding natural languages is one of the primary goals of artificial intelligence. Natural languages contain multiple clauses and long-term dependencies making them difficult for machines to understand, also sentences can have different types of dependency structures such as simple, compound, or complex which makes interpretation further difficult. One alternative way is to represent language using predicate logic which is easier for machines to understand. The task of manually converting language to predicate logic or ontologies can be cumbersome, but it can be automated using machine translation. For these ontologies to be effective the quality of translation should be good. In this research, we focus on analyzing the effect of sentence structure on machine translation quality, i.e., how the model performance is affected if the dataset is structurally skewed meaning it has sentences of similar structure more as compared to other structures. We further investigate the model learning behavior by performing statistical analysis on features learned by these models and understand the effects of sentence structure on these learned features. The statistical analysis helps us understand the distribution followed by the features and based on the insights gained we proposed a conditional checkpoint selection strategy centered on sentence structure along with utilizing Modified J-Divergence as a loss function for optimizing model performance for different sentence structures thus achieving better translation quality.