Interpretable Semantic Textual Similarity for Indonesian Sentence

Abstract

We develop iSTS (Interpretable Semantic Textual Similarity) model to Indonesian corpus. System of iSTS is not only to represent the STS (Semantic Textual Similarity) score but also to give an explanation of the semantic similarity of the pair of sentence. The term of explanation refers to a pair of chunks with type such as EQUI, OPPO, SPE1, SPE2, REL, SIMI, NOALI and score ranged 0 to 5. Nowadays, iSTS corpus has not existed in the Indonesian version yet, by that mean we build that corpus. We adapt two best iSTS techniques for English corpus: VRep and UWB. VRep uses WordNet to representing word semantic, while UWB uses word embedding. Both of the techniques use similar process, such as preprocess, feature extraction, and classification. The adaptation of VRep and UWB on this research is performed by changing English resources in Indonesia such as WordNet, word embedding, etc. We also use four classifier as well as decision tree, SVM, random forest, and multilayer perceptron. VRep becomes the best model on type aspect and score aspect, while UWB becomes the best model on type + score aspect.