KAFE: Knowledge and Frequency Adapted Embeddings

Abstract

AbstractWord embeddings are widely used in several Natural Language Processing (NLP) applications. The training process typically involves iterative gradient updates of each word vector. This makes word frequency a major factor in the quality of embedding, and in general the embedding of words with few training occurrences end up being of poor quality. This is problematic since rare and frequent words, albeit semantically similar, might end up far from each other in the embedding space.In this study, we develop KAFE (Knowledge And Frequency adapted Embeddings) which combines adversarial principles and knowledge graph to efficiently represent both frequent and rare words. The goal of adversarial training in KAFE is to minimize the spatial distinguishability (separability) of frequent and rare words in the embedding space. The knowledge graph encourages the embedding to follow the structure of the domain-specific hierarchy, providing an informative prior that is particularly important for words with low amount of training data. We demonstrate the performance of KAFE in representing clinical diagnoses using real-world Electronic Health Records (EHR) data coupled with a knowledge graph. EHRs are notorious for including ever-increasing numbers of rare concepts that are important to consider when defining the state of the patient for various downstream applications. Our experiments demonstrate better intelligibility through visualisation, as well as higher prediction and stability scores of KAFE over state-of-the-art.KeywordsWord embeddingsKnowledge graphsAdversarial learning