Constrained Text Clustering Using Word Trigrams

In recent years there has emerged the field of Constrained Clustering, which proposes clustering algorithms which are able to accommodate domain information to obtain a better final grouping. This information is usually provided as pairwise constraints, whose acquisition from humans can be costly. In this paper we propose a novel method based on word n-grams to automatically extract positive constraints from text collections. Clustering experiments in text collections composed by different types of documents show that the constraints created with our method attain statistically significant improvements over the results obtained with constraints created using named entities and over the results of a high-performing non-constrained algorithm.