engleski

Experiments with Neural Word Embeddings for Croatian

Word representations extracted from a large corpus have been shown to be very useful in a variety of natural language processing tasks. Recently, there has been much work on using neural networks to learn good word representations from raw text. We adopt this approach and train neural word embeddings from a large Croatian web corpus. We evaluate the embeddings on three lexico-semantic tasks: synonym detection, semantic relatedness, and analogy modeling. Results on all three tasks are remarkably good and some of them markedly above the state-of-the-art results for Croatian. In particular, on the synonym detection and semantic relatedness tasks, the model achieves an accuracy of 73% and a correlation of 0.67 with human judgments, respectively.

word embeddings ; lexical semantics ; distributional semantics ; Croatian language

nije evidentirano