This tool retrieves papers by measuring similarity between queries and sentences in the full text of papers in
CORD19 corpus using a similarity metric derived from BioSentVec
We also use the tool BeFree
for entity linking [1,2].
Genes are linked to UniProt
and diseases to medgen
We display relationships between genes and diseases from DisGeNET
Highlighting formatting is based on displaCy
This tool was developed by Manzil Zaheer (Google), Nicholas Monath (UMass Amherst),
Shehzaad Dhuliawala (Microsoft Research, Montreal), Taamannae Taabassum (University of Toronto), Rajarshi Das (UMass Amherst), Bhuwan Dhingra (CMU), and Andrew McCallum (UMass Amherst).
Overview of our system
We build our tool based on the (36K) research papers in the COVID-19 Open Research Dataset (CORD-19).
It retrieves papers by measuring similarity between queries and sentences in the full text of papers in CORD19 corpus. We also extract the bio-medical entities (diseases, genes, drugs) and link them to various ontologies. Lastly we also display various gene-disease association between the retrieved entities. This tool is a work-in-progress and below we provide a simple description of different components of our system.
We first preprocess the corpus by splitting the text of research papers into sentences using the sentence tokenizer in nltk.
Retrieving relevant evidence for a query
Given a query, our tool retrieves papers and highlights relevant sentences within it. We do this by assigning a score to all sentences in the corpus. The score reflects the similarity of the sentence w.r.t the query. We also assign score to each individual research paper as the maximum relevance score of a sentence in the paper. Our tool also displays context surrounding the high scoring sentence as it might contain relevant information as well. Next, we briefly describe how we score each sentence w.r.t a given query.
Scoring sentences using BioSentVec
The score of each sentence is computed as the inner product between the sentence and the query vector. We create vector representation of each sentence using a pre-trained BioSentVec model . The BioSentVec model is essentially a sent2vec model  which has been trained on the pubmed corpus and outputs a sentence embedding by averaging the embeddings of the words in them. Although simple, sent2vec has shown to be effective for many downstream applications.
We compute all sentences vectors and store them offline. Given a query, we first compute its distributed representation and then find the K-nearest sentence vectors (i.e. the top-K scoring sentences) from the corpus. As mentioned before, we score a document as the maximum score of any sentence within it. The figure below summarizes our system.
Extracting Knowledge Graphs (KGs) from the evidence
The main aim for building this tool is to help bio-medical researchers find relevant information. However, it is unlikely that all the answers that they are looking for would be answered by our initial set of retrieved papers. Often times, the initial retrieved evidence provides a good starting point for retrieving more evidence (e.g. think about those instances where you click an anchor link in a wikipedia page to go into another wiki page).
We hypothesize that the bio-medical entities (viruses, genes, drugs etc) which are mentioned in the retrieved research papers would be useful for the researchers. Therefore, we identify the entities present in the papers, and link them to knowledge bases from which the researchers can browse further information. We also display relationship edges between the entities (e.g. associations between genes and diseases). Below we briefly summarize each component
Identifying entities in text
The first step is to identify the entities that are present in the retrieved papers. We use a simple approach of fuzzy string matching to match to a dictionary containing various entity names and their aliases. The dictionary is derived from the BeFree system. After finding the entities, we link the genes to the UniProt and diseases to the MedGen ontologies. We are also adding entity linking drugs to Drugbank soon.
Finding relations between entities
There are several knowledge bases (KBs) that capture relation between entities. These KBs are either manually curated from the findings of several research papers over many years or have been automatically inferred from text using various relation extraction approaches. Finding and displaying these relations in our tool will allow researchers to get insights about findings that are not restricted to the retrieved paper. For our tool, we use the gene-disease association present in DisGeNET. We are working on adding relations between genes, diseases and drugs as well. The overview figure summarizes this component.
 À. Bravo, M. Cases, N. Queralt-Rosinach, F. Sanz, and L. I. Furlong, "A Knowledge-Driven Approach to Extract Disease-Related Biomarkers from the Literature", BioMed Research International, vol. 2014, Article ID 253128, 11 pages, 2014. doi:10.1155/2014/253128. (Article, for the "Big Data and Network Biology" special issue at BioMed Research International).
 À. Bravo, J. Piñero, N. Queralt, M. Rautschka and L.I. Furlong, "Extraction of relations between genes and diseases from text and large-scale data analysis: implications for translational research". BMC Bioinformatics 2015, Article, doi:10.1186/s12859-015-0472-9.
 Janet Piñero, Juan Manuel Ramírez-Anguita, Josep Saüch-Pitarch, Francesco Ronzano, Emilio Centeno, Ferran Sanz, Laura I Furlong.
The DisGeNET knowledge platform for disease genomics: 2019 update. Nucl. Acids Res. (2019)
 Chen Q, Peng Y, Lu Z. BioSentVec: creating sentence embeddings for biomedical texts. The 7th IEEE International Conference on Healthcare Informatics. 2019.
 Prakhar Gupta, Matteo Pagliardini, Martin Jaggi Better Word Embeddings by Disentangling Contextual n-Gram Information. NAACL 2019