Semantic Analysis of Math Word Problems

Rad opisuje simbolički pristup problemu automatiziranog rješavanja tekstnih matematičkih zadataka koji u sebi sadrže gotove formule matematičkih izraza i tekst koji pobliže opisuje način na koji ih treba rješavati. Razvijeni sustav svaku rečenicu zadatka prevodi u semantičku reprezentaciju stabla te njegovim obilaskom generira naredbe za izvršavanje od strane rješavača Photomath. Za semantičku analizu zadataka razvijena je arhitektura domensko-specifičnog jezika SML (engl. Semantic Math Language) definiranog tipovima, kontekstno-neovisnom gramatikom i transformacijskim pravilima za generiranje izlaznih naredbi. Funkcionalnost modela isprobana je na nekolicini primjera, no za valjanu evaluaciju jezik SML potrebno je nadopuniti dodatnim znanjem.This thesis describes a symbolic approach to the problem of automatically solving math word problems containing formulae of mathematical expressions and supplement text describing the desired procedure of solving them. The system translates each sentence of the math word problem into semantical representation using trees. By traversing the trees in postorder fashion the system generates instructions for Photomath solver. For the semantic analysis of word problems an architecture of domain-specific SML language (abbr. of Semantic Math Language) is developed. The language consists of user-defined types, context-free grammar, and transformation rules for the generation of output instructions. The functionality of the model is tested on a few examples, but for its proper evaluation further work on the SML language is needed