Learning to Search for Solving Natural Language Processing Tasks

Strukturno predviđanje i učenje sveprisutno je u problemima obrade prirodnoga jezika. Metode učenja pretraživanja pružaju okvir u kojem je te probleme moguće efikasno rješavati. U ovom radu dan je pregled povijesti metoda učenja pretraživanja i osvrt na područja u strojnom učenju koja su omogućila njihov razvoj. Pokazana je i primjena na razne probleme u obradi prirodnog jezika. Istaknute su i bitne razlike između ostalih pristupa strukturnom učenju koje utvrđuju nadmoć i prilagodivost metoda učenja pretraživanja. Ostvaren je sustav koji zadatke označavanja vrste riječi i ovisnosnog parsanja vrši združeno te su istaknute razlike i prednosti na pristupe koji zadatke ne gledaju združeno. Vrednovanje sustava izvršeno je na podacima za hrvatski jezik.Structured prediction and learning is omnipresent for problems in natural language processing. Learning to search (L2S) methods provide a framework in which these problems can be efficiently solved. This thesis gives an overview of L2S methods and their theoretical basis that allowed their development. Their application to various problems of natural language processing is described. Important differences that show the superiority and flexibility of L2S methods are pointed out in the context of previous solutions for structured and joint prediction. Description of the development of a system that considers part-of-speech tagging and dependency parsing as a joint task is provided and detailed analysis of differences and advantages to various approaches is given. Evaluation of the system is done on Croatian corpus