The Role of Local and Global Weighting in Assessing the Semantic Similarity of Texts Using Latent Semantic Analysis

In this paper, we investigate the impact of several local and global weighting schemes on Latent Semantic Analysis’ (LSA) ability to capture semantic similarity between two texts. We worked with texts varying in size from sentences to paragraphs. We present a comparison of 3 local and 3 global weighting schemes across 3 different standardized data sets related to semantic similarity tasks. For local weighting, we used binary weighting, term-frequency, and log-type. For global weighting, we relied on binary, inverted document frequencies (IDF) collected from the EnglishWikipedia, and entropy, which is the standard weighting scheme used by most LSA-based applications. We studied all possible combinations of these weighting schemes on the following three tasks and corresponding data sets: paraphrase identification at sentence level using the Microsoft Research Paraphrase Corpus, paraphrase identification at sentence level using data from the intelligent tutoring system iSTART, and mental model detection based on student-articulated paragraphs in MetaTutor, another intelligent tutoring system. Our experiments revealed that for sentence-level texts a combination of type frequency local weighting in combination with either IDF or binary global weighting works best. For paragraph-level texts, a log-type local weighting in combination with binary global weighting works best. We also found that global weights have a greater impact for sententence-level similarity as the local weight is undermined by the small size of such texts.