Cross-component Clustering for Template Induction

We suggest an unsupervised approach to template induction for information extraction, through detecting sub-topics and themes that cut across the documents of a topical corpus. We introduce a new method – cross component clustering – that simultaneously clusters the components forming our setting, each of which consists of the words of a single article. Our algorithm is derived from the Information Bottleneck clustering algorithm. The resulting clusters are found to be in systematic correspondence with sets of terms that are used in filling the slots of the MUC3/4 ready-made template, which was used for evaluation.