Publication

Abstract : Text kernels compute similarity using mathematical functions directly from text, allowing for interpretable and structured analysis of textual data. Traditionally, text kernels rely solely on syntactic features, limiting their capacity to capture deeper semantic relationships. This work introduces a novel framework that integrates advanced artificial intelligence techniques, specifically semantic word embeddings, with traditional text kernels to improve the accuracy of text classification tasks. Kernels are created from deep word embeddings using the cosine similarity measure. The amalgamation of Word Embedding Kernels thus created with other text kernels derived twenty-two Semantic Text Kernels. Additionally, a topic-space-based technique is employed to reduce the size of text datasets, enhancing both scalability and computational efficiency. The resulting semantic kernels are applied to the artificial intelligence-driven task of text classification, leading to performance improvements across benchmark datasets. As illustrative results, we observe absolute performance gains of 4.19% on the Movie Review dataset, 0.96% on the Stanford Sentiment Treebank binary dataset, 1.20% on the Reuters eight-class dataset, and 5.50% on the 20 Newsgroups dataset, respectively. This work represents a significant advancement in interpretable text analysis and natural language comprehension, effectively bridging the gap between traditional text kernels and modern deep learning-based word embeddings.