Organizing Flickr30k Using Text Clustering

Abstract

Text clustering is the process of clustering similar documents together based on the textual information within a document. The captions provided with the Flickr30k dataset will be used to organize the images. The dataset consists of captioned images of everyday life. The two approaches to clustering (hierarchical and partitional) will be implemented to assess the formed clusters. K-means and agglomerative clustering will be used to experiment with. The performance of the two algorithms will be assessed using internal validity measurements. The difference between the two algorithms was too small to judge which one performed better. However the clusters that are formed did differ. K-means made a distinction between ‘adult people’ vs. ‘young people’, agglomerative clustering made a distinction between ‘people’ vs. ‘bullfighting’.