A genetic algorithm for logical topic text segmentation

Topic text segmentation is an important problem in information retrieval and summarization. The segmentation process tries to split a text into thematic clusters (segments) in such a way that every cluster has a high cohesion and the contiguous clusters are connected as little as possible. The originality of this work is twofold. First, we propose new segmentation criteria based on text entailment for interpreting the cohesion and connectivity of segments and second, we use a genetic algorithm which uses a measure based on text entailment for determining the topic boundaries, in order to identify a predefined number of segments. The obtained results are compared with against two manually segmented texts.