A simple topic-model browser for SciXplore (focusing on Covid-19 Open Research Corpus Analysis )

Our goal in this demo is to analyze COVID-19 Open Research Dataset in order to:

• Demonstrate related domain specific analysis capability of SciXplore
• Identify intepretable topics (described by WIKIPedia concepts, MeSH, words & texts)
• Identify and annotate publication with WikiPedia or MeSH related concepts
• Analyze related topic trends
• Analyze similarity among topics, among publications or other entities
• Demonstrate topic & concept based exploration & information retrieval capabilities

The analysis has been based on SciXplore multi-view topic modelling, representation learning and semantic annotation framework. SciXplore provides a lot of advantages on top of text only topic modeling alternatives (e.g., LDA) producing more interpetable, multi-view, topics as well as vector representations (embeddings) for all involving entities (including topics, words & concepts). Therefore, by combining multi-view topic modelling with skip-gram based topic2vec representation learning infers topics of better quality and facilitate similarity analysis & recommendation. SciXplore also provides a semantic annotation engine for DBPedia (Wiki) based annotation which extends the well-established DBPedia Spotlight framework

The visualization has been based on agoldst.github.io/dfr-browser, use d3 and it relies entirely on static html and javascript files. We have created a python scipt which analyzes the results of SciXplore and exports required data files in a format which is compatible with dfrbrowser. To better support the requirements of SciXplore & Covid-19 dataset we have implemented several changes on top of the dfrbrowser such as:

• Change the way we are calculating trends and more specifically the normalization: topic activity per year is based on the ratio of the docs of a topic in a given year / total docs for a given topic (and total docs of all topics for a given year)
• Change html / CSS in several places to handle longer 'word' descriptions (which correspond to concepts)
• Support autocompletion & multi-selection in word & concept search
• Show a topic relations view (through a Force-Directed Layout)

Our analysis has been based on the following steps:

• Consume and pre-process publications from COVID-19 Open Research Dataset
• Enrich them (e.g., getting MeSH terms) utilizing MEDLINE/PubMed dumps
• Annotate all docs with Wikipedia terms using SciXplore semantic annotation engine
• Analyze text, metadata, MeSH & Wikipedia concepts with SciXplore inferring:
  • Intepretable multi-view topics
  • Topics per document or other entities
  • Vector representations per topics or other entities
  • Analyze trends & similarities
  • Create a WEB GUI for data exploration (adapting agoldst.github.io/dfr-browser)
  • Give titles to the topics

There are several things that we can do on top (or that they aren't properly addressed through this visualization) especially in relation to search & recommendation, such as:

• besides silimiarities among topics, identify and analyze similarities among concepts or documents
• content-based recommendation: given a publication, topic or concept recommend similar entities
• improvements on concept-based search (e.g., search using multiple topics or articles which are related to some concepts)
• merge corresponding WIKI & MeSH concepts (and provide links)
• journal or publisher specific analysis (e.g., search by journal or publisher)

For more information or feedback please contact Omiros Metaxas: ometaxas@atypon.com