The Evaluation of Thai Poem's Content Consistency using Siamese Network

Abstract

Many research describes Textual Entailment model for compare pair of the sentence but two sentences in term of the poem content consistency are not the same. The content consistency is very important for storytelling in Thai poem composing. In this article, we propose the model and result of The evaluation of Thai poem's content consistency using The Siamese Network 3 models comprise 1) Merge Vector Model 2) Siamese Absolute Different Model and 3) Siamese Dot Vector Model compare with the Basic CNN model. The training data is Thai poem 14,173 pair (batt) and validation data is Thai poem 3,544 pair. All models learn by apply one shot learning technic. The accuracy of Siamese Absolute Different Model near 100%. The macro average of F1-score shows 99.27%. The Area Under Curve shows 0.997 near the perfect value.