Explainable Word-Embeddings for Medical Digital Libraries – a Context-Aware Approach

TitleExplainable Word-Embeddings for Medical Digital Libraries – a Context-Aware Approach
Publication TypeConference Paper
Year of Publication2020
AuthorsWawrzinek, J., S. A. R. Hussaini, O. Wiehr, J. M. G. Pinto, and W. - T. Balke
Conference NameACM/IEEE Joint Conference on Digital Libraries (JCDL’20)
Date Published08/2020
Conference LocationXi'an, Shaanxi, China

State of the Art Neural Language Models (NLMs) such as Word2Vec are becoming increasingly successful for important biomedical tasks such as the literature-based prediction of complex chemical properties or for finding novel drug-disease associations (DDAs). However, NLMs have the disadvantage of being hard to interpret. Therefore, it is notoriously difficult to explain why an artificial neural network learned or predicted some specific association.

Considering that digital libraries offer well-curated contexts, the challenge is to automatically create a reasonable explanation that is intuitively understandable for a user. For a pharmaceutical use case, we present a new method that generates pharmaceutical explanations for predicted DDAs in intuitively understandable sentences. In other words, our approach enables a context-aware access to embedded entities. We test the accuracy of our approach with a comprehensive retrospective analysis considering real DDA predictions. Our explanations can automatically determine the association type (Drug treats or induces a disease) of a predicted DDA with up to 83%. For existing DDAs, we even achieve accuracies up to 87%.  We show that we perform better than deep-learning approaches in this classification task by up to 9%.

Wawrzinek82.pdf692.25 KB