Automatic generation of textual logic puzzles in Slovenian

Peter Holozan (2011) Automatic generation of textual logic puzzles in Slovenian. MSc thesis.

PDF Download (1411Kb)

Abstract

Creating textual logic puzzles requires quite a lot of work. That is why a number of programs have already been created to simplify this process, and they were described in the introduction. One of the programs for Slovenian logic grid puzzles is Spesni, with the downside that it requires a new template for each new theme, and that is a lot of work. In this thesis, Spesni was extended to derive each theme from a sample sentence. Amebis Presis machine translation software technology was used. The first thing was to check if Presis Interlingua supports all the necessary constructions to create logic puzzles. To achieve this, samples of logic puzzles were collected and analyzed into Interlingua. When needed, upgrades to Interlingua (and consequently to the analyzer and generator) were suggested. The program works as follows: the Presis analyzer translates a sample sentence into Interlingua, any changeable characteristics in the puzzle are indentified, and a pool of possible values is selected for each of them. This is done using the hypernym and hyponym information from the Ases database. The logic grid puzzle is generated and the list of clues is transformed into Interlingua written sentences. In this process, the clues are aggregated into sentences as much as possible, and then the sentences are reordered to ensure referential coherence. Finally, the Presis generator translates sentences from Interlingua to a natural language (Slovenian). At the end, the program was successfully tested for English, both when the sample sentence was in English and when the same puzzle was generated in Slovenian and English at the same time. This showed that if Presis supports new languages in the future, it will be quite simple to use Spesni for those languages. The program was tested on a large number of cases (for example, newspaper headlines) to assess how often a puzzle is generated successfully for a sample sentence. It was shown that it is easy to create puzzles with very different themes, and it works even better with manually created sample sentences instead of random ones. There is room for improvement especially for the analyzer (disambiguation) and semantic network completeness in the Ases database. On the other hand, logic grid puzzles with more complex clues could also be made. This technology could also be used with other types of textual logic puzzles.

Actions (login required)