Combining embedding methods for a word intrusion task

We report a new baseline for a Danish word intrusion task by combining pre-trained offthe-shelf word, subword and knowledge graph embedding models. We test fastText, Byte-Pair Encoding, BERT and the knowledge graph embedding in Wembedder, finding fastText as the individual model with the superior performance, while a simple combination of the fastText with other models can slightly improve the accuracy of finding the odd-one-out words in the wordintrusion task.