A repository to explore the use of ontology-based graphs (generated with MOWL) as background knowledge for scientific articles similarity.

• RELISH corpus
• RELISH corpus pre-annotated with terms from the MeSH vocabulary provided by ZB MED
• MOWL library to generate graphs from ontologies
• Libraries to create graph embeddings

1. Using MOWL to generate a graph corresponding to the MeSH vocabulary. The input for this step is the MeSH vocabulary and the output is a graph.
2. Using the RELISH corpus pre-annotated with MeSH terms (i.e., any entity from the MeSH vocabulary recognized in the text will have been replaced by the corresponding term ID) create a mini-graph for each document. Use co-occurrence to add weights to the edges. This graph will be most likely kind of disconnected (e.g., not too many edges). The input for this step will be a TSV file where each row is a pre-annotated article from the RELISH corpus (see example below) while the output will be a graph per document.
3. For all nodes in a document mini-graph, identify the direct nodes coming from the MeSH-based graph and add them to the mini-graph. This will give you a more connected graph, with background knowledge coming from the MeSH vocabulary. The inputs for this step are the document mini-graphs while the outputs will be document enriched mini-graphs.
4. Create graph embeddings for the documents in the RELISH corpus.
5. Calculate the cosine similarity for each existing pair of documents in the corpus.
6. Analyze correlation of the cosine similarities using this approach versus another one created in the OntoClue framework.
7. Discuss results (lessons learnt, possible variations, limitations).

• We are not capturing relations from the text between two recognized.
• Some biologically relevant entities, see P53 in the example, are not part of the MeSH vocabulary (however this approach is for a recommendation system and not for a specific biological task so might still be ok).

• Explore Named Entity Recognition approaches other than simple/direct lexical matching.
• Using sentiment analysis to add positive/negative connotation to the relation betweeen two entities recognized in the text.

• Nelson David Quiñones Virgen, co-lead, orcid:0000-0002-5037-0443
• Leyla Jael Castro, co-lead, orcid:0000-0003-3986-0510
• Asmaa Alqhtani, participant, orcid:0000-0002-3119-5109
• Rahaf Alayed, participant, orcid:
• Layan Aljohani, participant, orcid:0000-0001-6954-2973
• Maria G Gomez Castillo, participant, orcid:

• Day 1: Formation of the team, discussion on the approach, distribution of initial tasks.
  • Task 1: Asmaa, Rahaf, Layan, Maria. Getting a graph out of MeSH using mOWL
  • Task 2: Nelson. Explore strategies to create mini-graphs representing the publications
• Day 2
  • Task 1 done, using the KAUST cloud, output saved in a pkl file, the projection used was DL2Vec for taxonomy only in two directions.
  • Task 2: MeSH terms extracted to a JSON file using the PMID as key for each article, invlufinh the co-occurrence count.
  • Task 3: Maria and Rahaf. From the MeSH graph created in Task 1, use a graph embedding algorithm to obtain the graph embeddings, each concept will be assigned a vector
  • Task 4a: Layan and Maria. Define and implement a strategy to get all the concept embeddings for a document and move to document embeddings. Once the MeSH graph embeddings are ready, apply to the actual data.
  • Task 4b: Nelson. Define and implement a strategy to create mini-graphs out the concepts identified in each document and the MeSH graph embeddings.
• Day 3
  • Leyla Jael. Add information to the mid-term report. Done
  • Task 4a. Asma, Rahaf, Maria, Layan. Word embeddings for the graph are done. Continuing with the document embeddings for RELISH corpus.
  • Task 4b. Nelson. Exploring options based on lowest common ancestor.

This project started as part of the BioHackathon MENA 2023