CASBERT

We used CASBERT to create variable/constant embeding. CASBERT ¹⁸ is a tool for converting the composite annotation of an entity into an embedding by applying Sentence-BERT. ¹⁵ Sentence-BERT is used to convert textual properties related to ontology classes ( e.g. CHEBI:17138, OPB_00340) and predicates ( e.g. isPropertyOf, isVersionOf) to embeddings. Then, the embeddings is merged to create an embedding representing variable/constant.

Technically CASBERT can implement other approaches such as ColBERT ¹⁷ and poly-encoder ¹⁶ with the increase in computational and indexing complexity; the use of Sentence-BERT is preferred because of its practicality while still providing good performance. Other alternatives are Onto2Vec ²⁸ and OPA2Vec ²⁹ which use Word2Vec ³⁰ to translate the words in the ontology class to embeddings, and then merge them as an ontology embedding. However, a word embedding approach only considers word co-occurrence in the training data, so the embedding of a particular word will be the same for all sentences. Moreover, it does not consider the context of the sentence.