Entities, structures, reactions — the three data formats of chemistry, each with a completely different processing path. Structures travel as images, entities carry multiple names, and reaction information is a set of bound relations. This page explains what makes each of them hard, and exactly which sub-tasks we handle.
Chemical information is mainly presented in the form of structure diagrams, and a structure diagram is a graph, not text. Automated recognition tools perform well on neatly laid-out patent tables, but the moment you reach indefinite structures, Markush formulas, structures spanning pages, or obscured substituents, someone with a chemistry background has to make the call.
What makes it harder is that errors raise no error flag: a model can output a structure that is syntactically valid but chemically impossible, such as pentavalent carbon. Such results enter the dataset quietly, and are only found by tracing back when model behaviour becomes abnormal.
Systematic names, trivial names, abbreviations and research codes — the same substance may be written completely differently across documents and patent versions. This is not merely a difference of expression — it determines whether these records can be recognized as the same entity in a database.
Starting materials, intermediates, products, reagents, solvents and yields, plus reaction conditions such as temperature, time and catalysts, form a set of mutually bound relations. Annotating any single item on its own cannot reconstruct the reaction; they must be processed as a group.
Mixed languages and range-value descriptions are both within scope. Chinese patents, English patents, academic literature and experimental records differ considerably in layout, so processing paths are designed separately, but they share the same schema design method.
Extract and correct chemical structure diagrams from patents, literature and experimental records: full compound names, Markush formulas, indefinite structures and complex substituents. Delivers a structured SMILES / InChI dataset in which every record is traceable.
OCSRMarkushStructure drawingThe difference is that the data does not exist as text alone. Structures are graphs and require the double processing of recognition plus judgment; reaction information is relational and must be annotated in groups. Text annotation has room for error tolerance — here there is none: get one bond wrong and the identity of the compound changes.
Systematic names, trivial names, abbreviations and research codes are all kept as written and recorded separately, while links pointing to the same entity are established. We do not force a single unique name — different versions within a patent family using different designations is normal, and replacing them would remove the entry point for trace-back verification.
By project agreement, the original wording is preserved and a separate normalized field is created. The name, number and example number of the same substance may differ between the Chinese and English versions; only by keeping the original can each be checked one by one.
Because reaction information is relational: starting materials, intermediates, products, reagents, solvents and yields form a set of bound relations, and annotating any single item on its own cannot reconstruct the reaction. Only group annotation lets downstream users apply it directly to route retrieval and synthesis planning.
We deliver JSON / CSV / Excel, with fields mapping one-to-one to the co-designed schema; structures are output as SMILES / InChI, entities and relations are split into linkable tables, and the source location in the original text is preserved.
Send us a representative patent or paper and we can run a small trial annotation first, so you can see directly how structure judgment and entity normalization are handled.