[Talks@IDSIA] This Wednesday at 14:00, meeting room, "Dynamic Thresholds for Author Clustering"

We will have a talk tomorrow (Wednesday 18 April), in the meeting room. ::Guest: Mirco Kocher Computational Linguistics group Computer science department of the Université de Neuchâtel. ::Title: Dynamic Thresholds for Author Clustering ::Abstract: In author clustering, we have a corpus with texts and we want to regroup them such that each of the clusters corresponds to a distinct author. We propose to use a dynamic threshold for single link hierarchical clustering. First, we determine the probability for each text pair to be written by the same person. To do so, we present an effective feature selection strategy and compute the pairwise distances of documents in the collection. Assuming a Gaussian distribution of the distances for each text, we can select the relative shortest distances and assign them higher probabilities. Comparisons with other systems from PAN CLEF have shown that this approach achieves some of the highest clustering performances and also completes the task with one of the shortest runtime.