| Popis: |
Odgovarjanje na vprašanja je pomembna in pogosto naslovljena naloga obdelave naravnega jezika v angleščini. Jezikom z manj viri, kakršna je slovenščina, je namenjeno manj pozornosti. V tem delu uporabimo enega izmed uspešnih angleških pristopov, poimenovanega UnifiedQA, in preizkusimo njegovo delovanje za slovenski jezik. Naučimo generativni model za odgovarjanje na vprašanja, ki pokriva štiri različne tipe vprašanj - da/ne, večizbirni, abstraktni in ekstraktivni tip. Za učenje uporabimo obstoječe podatkovne zbirke BoolQ, COPA, MultiRC in SQuAD 2.0 ter strojno prevedemo podatkovno zbirko MCTest. Pokažemo, da je splošni model sposoben odgovarjati na vprašanja v različnih formatih, saj deluje vsaj tako dobro kot namenski modeli, z vnosom angleškega znanja pa rezultate še izboljšamo. There have been many studies in the field of question answering for English language. Less attention has been devoted to low-resource languages, such as Slovene. In this work, we use one of successful English approaches, named UnifiedQA, and test its viability for Slovene language. We finetune a generative model for question answering, covering four different question formats - yes/no, multiple choice, abstractive and extractive format. For finetuning, we use existing datasets BoolQ, COPA, MultiRC and SQuAD 2.0 and machine translate the MCTest dataset. We show that a general model is capable of answering questions in different formats at least as well as specialized models. The results are further improved using examples in English language. |
