Deriving labeled training data for topic link detection by alternating words

Although classifiers can be trained to estimate whether two short text segments relate to a common topic, obtaining training data for supervised learning presents a hurdle. The natural approach would be to train with topic-aligned pairs of text segments from a large corpus, but nothing is available to locate such alignments. We offer that simply partitioning the words of a large document according to their odd and even positions will yield training data suitable for certain applications and sets of features. The reason is that the partitioned texts are topic-aligned along their respective lengths despite sharing no original word instances. We further show that parametrically introducing a small amount of overlap into the partitioned texts can greatly improve the precision of a classifier.