Centroid-Based Clustering Using Sentential Embedding Similarity Measure

Abstract

Text is treated as a bag of words in traditional text clustering methods, with the similarity between two texts measured by word co-occurrence. While this technique is appropriate for clustering document-level text, it underperforms when clustering short-text fragments like sentences. This is due to compared sentences can be semantically related without any similar words or phrases. This study describes a new variation of short-text clustering method based on the idea of sentential semantic embedding vectors. These embedding vectors represent compared text using semantic and lexical information driven from a lexical knowledge database built to emulate common human knowledge about words in natural human language. On two-sentence datasets with varying degrees of word co-occurrence, we compared the technique’s performance to that of a graph-based clustering algorithm. We argued that our centroid-based method’s higher performance on datasets containing a low degree of word co-occurrence was due to its ability to utilize the available semantic information.