Public repositories containing a diverse chemical and biological data are one of the main sources of knowledge for further biomedical research. Unfortunately, extraction and transforming these data into a well-interpretable form is a complex task. Ongoing community efforts are mainly focused on the analysis of co-occurrences of terms, text annotation based on terms similarity and related tasks.<br/>Here we present an approach based on natural-language processing techniques, which is intended to shift the focus of a search for similar texts on chemical topics from word- to document-level. PubMed records were used to implement word2vec and doc2vec models. Generated text representations can be used to search for similar abstracts; similarity is more dependent on this representation than co-presence of the certain terms (neighbor compounds, similar publication date, etc.).<br/>Document-level clustering was also implemented to provide insight into PubMed text corpus structure. This approach can serve as an alternative to standard topic modeling techniques for discovery of the hidden semantic features in unsupervised manner.

Abstract recommendation system: beyond word-level representations

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.