Semantic Term weighting representation for Kannada Document Classification

Abstract

Abstract In natural language processing, sequence order of terms plays a vital role. This positional sequence information helps in the semantic analysis of the natural language. The absence of semantic information in term weighting methods motivated us to propose the semantic term weighting representation. On the other hand, to address the demand for Indian regional language resources, especially for the Kannada language we have created an 11,045 Kannada documents dataset. This dataset is multilabel and unbalanced. The proposed semantic term weighting representation methods (Term Frequency-Positional encoding (TF-PE) and Term Frequency-Inverse document frequency-Positional encoding (TF-IDF-PE)) are applied to the proposed dataset. Further, the K-Fold and normal train-test split experimentations are carried out on the proposed dataset. Among the proposed representation methods Unicode encoded Term Frequency-Inverse document frequency-Positional encoding (TF-IDF-PE) representation performed better than Term frequency-Positional encoding (TF-PE) representation. The Unicode encoded TF-IDF-PE representation with the SVM classifier yields better average accuracy of 68.62% in K-10 Fold experimentations.