Distributional memory explainable word embeddings in continuous space

Abstract

Natural Language Processing (NLP) is a key processing step in fusion systems that need to process unstructured -and possibly human generated- text in natural language. Recent developments in Deep Learning have greatly increased the performance of NLP tasks. In particular, learned word representations have the form of high dimensional real valued vectors, called word embeddings, that have a number of amenable properties such as representing similar words with similar values of their vectorial representation, and capturing semantic regularities that correspond to geometric properties in the continuous high dimensional space. However, word embeddings have the drawback of being non interpretable. That is, their dimensions cannot be clearly associated to linguistic features. In this work, we propose real valued explicit linguistic word vectors that enjoy the properties of learned word embeddings while being human understandable.