Multifaceted Domain-Specific Document Embeddings
This page corresponds to our demo paper submission titled "Multifaceted Domain-Specific Document Embeddings" by Julian Risch, Philipp Hager, and Ralf Krestel. The submission is currently under review at NAACL'21. It is based on a Master's thesis by Philipp Hager.
To show the practical feasibility of our approach, we implemented a demo, which can be accessed here. Our source code and the evaluation datasets are available on GitHub and a screencast is on YouTube.
Word and document embeddings are Natural Language Processing (NLP) techniques that map words to fixed-length numerical vectors in an embedding space. Current embedding algorithms work well when trained on large text corpora, but fail to produce high-quality vectors when given a small number of documents or are confronted with uncommon terms, as is often the case in specialized domains. Secondly, it is common to blend the entire document into a single embedding vector, making it hard to find documents relating only to a specific piece of information or to explain why two documents are considered similar. In this work, we propose a novel approach to train document embeddings for domain-specific texts. We use a siamese neural network architecture in combination with knowledge graphs to train document embeddings on a small number of training examples from the medical domain. The model identifies different types of domain knowledge and encodes them into separate dimensions of our embedding, thereby enabling multiple ways of finding and comparing related documents in vector space. We evaluate our approach on medical journal articles. An interactive demo, our source code, and the evaluation datasets are available online: https://hpi.de/naumann//s/multifaceted-embeddings.