Conceptualized phrase clustering with distributed k -means

Abstract

A vast majority of text mining and machine learning algorithms such as topic models, classification, clustering are based on statistical methods thus the semantics or meaning of the words or phrases are not considered. Interpretation of outputs generated by such algorithms are difficult for humans because of the absence of sufficient contextual information. Distributional semantics is a relatively new but active research area in natural language processing that quantifies semantic similarities between linguistic elements considering the context in which they occur. Conceptualization algorithms on the other hand enriches short text such as words and phrases. This paper proposes an approach that uses a map-reduce framework for combining these two techniques to generate conceptualized semantic clusters of phrases using distributional representation. Rigorous and systematic experiments on unstructured text datasets show that this approach can generate semantically rich and human interpretable concept clusters from large datasets. Further, the approach is scalable when dealing with high dimensional data since this method uses a map-reduce based framework for clustering.