Fig. 1. Conceptual diagram of knowledgeable and versatile machine reading.

Here we take salicylic acid as an example. Inspired by humans that versatilely learn meta-knowledge within and across different information, our machine reading system first serializes, a molecule structures via BPE on SMILES strings, then inserts the substrings into c. large-scale corpus and learns b fine-grained mapping between different semantic units by d mask language modeling. In this way, the system can perform e knowledgeable and versatile reading, achieving good performance on both mono-information downstream tasks and versatile reading tasks.