DISAMBIGUATING SEMANTIC ROLES IN SWEDISH COMPOUNDS with Swedish FrameNet and SALDO

Abstract

The compounding of words in Swedish is productive, recursive, and frequent in both text and speech. Compounds can be ambiguous on many levels, and the processing of them involves segmentation, lemma disambiguation, word sense disambiguation, and semantic analysis. In this thesis, we focus on the latter. We concretise the semantic analysis as semantic role disambiguation, meaning the automatic analysis of the relationship between the two parts of a compound (prefix and suffix) given a set of semantic roles selected by the suffix. The system architecture revolves around lexical resources such as the Swedish FrameNet (SweFN) and SALDO. In two experimental rounds, we train on (1) chunked and semantic role-analysed sentences, and (2) compounds marked up using the frames and semantic roles of SweFN. For instance, laxröra ‘salmon casserole’ is analysed as Constituent_parts+LU (LU=lexical unit) in the Food frame. The training data of tagged sentences used in predicting compound semantic roles is deemed too sparse, and produces only a small improvement over a most-frequent-class baseline. In our final experiments, we use a narrowed down set of frames and compounds as both train and test data. We reach a best classification accuracy of 62% against a 33% baseline on 100 unseen compounds.