Widening the HolSum Search Scope

We investigate different areas of the highdimensional vector space built by the automatic text summarizer HolSum, which evaluates sets of summary candidates using their similarity to the original text. Previously, the search for a good summary was constrained to a very limited area of the summary space. Since an exhaustive search is not reasonable we have sampled new parts of the space using randomly chosen starting points. We also replaced the simple greedy search with simulated annealing. A greedy search from the leading sentences still finds the best summary. Finally, we also evaluated a new word weighting scheme: the standard deviation of word distances, comparing it to the previously used tf · log(idf) weighting. Different weighting schemes perform similarly, though the term frequency contributes more than other factors. 1 Language Independent Automatic