Fig. 2. Schematic diagram for KV-PLM* finishing CHEMIchoice task.

For the given unfamiliar molecule entity, we get a versatile materials including structure and description, from which we know the correct sentence and randomly pick wrong sentences from the pool to form four choices. b Molecule structure and text of choices are fed into KV-PLM* and get their representations, based on which the confidence scores of choices are calculated by cosine similarity. c The tokenizers for structures and biomedical text are different. In this instance, KV-PLM* successfully finds out the correct description sentence for the given substance.