EVE: Explainable Vector Based Embedding Technique Using Wikipedia

摘要

We present an unsupervised explainable word embedding technique, called EVE,which is built upon the structure of Wikipedia. The proposed model defines thedimensions of a semantic vector representing a word using human-readablelabels, thereby it readily interpretable. Specifically, each vector isconstructed using the Wikipedia category graph structure together with theWikipedia article link structure. To test the effectiveness of the proposedword embedding model, we consider its usefulness in three fundamental tasks: 1)intruder detection - to evaluate its ability to identify a non-coherent vectorfrom a list of coherent vectors, 2) ability to cluster - to evaluate itstendency to group related vectors together while keeping unrelated vectors inseparate clusters, and 3) sorting relevant items first - to evaluate itsability to rank vectors (items) relevant to the query in the top order of theresult. For each task, we also propose a strategy to generate a task-specifichuman-interpretable explanation from the model. These demonstrate the overalleffectiveness of the explainable embeddings generated by EVE. Finally, wecompare EVE with the Word2Vec, FastText, and GloVe embedding techniques acrossthe three tasks, and report improvements over the state-of-the-art.