Semi-Automatic Story Generation for a Geographic Server

Abstract

Most existing servers providing geographic data tend to offer various numeric data. We started to work on a new type of geographic server, motivated by four major issues: (i) How to handle figures when different databases present different values; (ii) How to build up sizeable collections of pictures with detailed descriptions; (iii) How to update rapidly changing information, such as personnel holding important functions, and (iv) how to describe countries not just by using trivial facts, but stories typical of the country involved. We have discussed and partially resolved issues (i) and (ii) in previous papers; we have decided to deal with (iii), regional updates, by tying in an international consortium whose members would either help themselves or find individuals to do so. It is issue (iv), how to generate non-trivial stories typical of a country, that we decided to tackle both manually (the consortium has by now generated around 200 stories), and by developing techniques for semi-automatic story generation, which is the topic of this paper. The basic idea was first to define sets of reasonably reliable servers that may differ from region to region, to extract “interesting facts” from the servers, and combine them in a raw version of a report that would require some manual cleaning-up (hence: semi-automatic). It may sound difficult to extract “interesting facts” from Web pages, but it is quite possible to define heuristics to do so, never exceeding the few lines allowed for quotation purposes. One very simple rule we adopted was this: ‘Look for sentences with superlatives!’ If a sentence contains words like “biggest”, “highest”, “most impressive” etc. it is likely to contain an interesting fact. With a little imagination, we have been able to establish a set of such rules. We will show that the stories can be completely different. For some countries, historical facts may dominate; for others, the beauty of landscapes; for others, cultural or economic achievements, and for yet others, unusual facts concerning Nobel Prize winners, food, entertainment, sports, other activities, national symbols, special laws, and so on. The results can be checked on by clicking on any country in the category “Special Information” under “Surprising Facts”. All examples shown in this paper were chosen fairly arbitrarily from over 190 examples, to show that the system is indeed working. There are two points to mention here: (a) it is a work in progress, yet has reached a very useable size; (b) the basic ideas can be applied to any area. The choice of geography was due to the wealth of data and interest in this area, but if our algorithms overlook some important facts, this is less critical than applied to types of medical treatment, etc.