Evaluating Word Embeddings in Multi-label Classification Using Fine-grained Name Typing

Embedding models typically associate each word with a single real-valued vector, representing its different properties. Evaluation methods, therefore, need to analyze the accuracy and completeness of these properties in embeddings. This requires fine-grained analysis of embedding subspaces.
Multi-label classification is an appropriate way to do so. We propose a new evaluation method for word embeddings based on multi-label classification given a
word embedding. The task we use is fine-grained name typing: given a large corpus, find all types that a name can refer to based on the name embedding. Given the scale of entities in knowledge bases, we can build datasets for this task that are complementary to the current embedding evaluation datasets in: they are very large, contain fine-grained classes, and allow the direct evaluation of embeddings without confounding factors like sentence context