Improved Semantic Representation for Domain-Specific Entities

Most existing corpus-based approaches to semantic representation suffer from inaccurate modeling of domain-specific lexical items which either have low frequencies or are non-existent in open-domain corpora. We put forward a technique that improves word embeddings in specific domains by first transforming a given lexical item to a sorted list of representative words and then modeling the item by combining the embeddings of these words. Our experiments show that the proposed technique can significantly improve some of the recent word embedding techniques while modeling a set of lexical items in the biomedical domain, i.e., phenotypes.