A generating model for Finnish nominal inflection using distributional semantics

Abstract Finnish nouns are characterized by rich inflectional variation, with obligatory marking of case and number, optional possessive suffixes the possibility further cliticization. We present a model for conceptualization inflected nouns, using pre-compiled fasttext embeddings (300-dimensional semantic vectors that approximate words’ meanings). Instead deriving vector an word from another in its paradigm, we propose is conceptualized means summation latent representing meanings lexeme features. tested this on 2,000 most frequent their forms corpus (84 million tokens). Visualization space t-SNE clarified ‘main effects’ additive does not do justice to semantics inflection. In Finnish, how number realized turns out vary substantially case. Further interactions emerged clitics. By taking these into account, accuracy our model, evaluated as gold standard, improved 76% 89%. Analyses errors made 7.5% due overabundance (and hence true errors), 16.5% involved exchanges semantically highly similar stems (lexemes). Our results indicate, first, noun inflection more intricate than assumed thus far, second, intricacies can be captured surprisingly high simple generating based imputed lexemes, features,