Correction to Automated Chemical Reaction Extraction from Scientific Literature

Unleashing the Power of Knowledge Extraction from Scientific Literature in Catalysis

“…Apart from the training data and optimization method, the initialization of model parameters is also an important factor for the final performance, which is also demonstrated in the chemical reaction extraction experiment …”

“…This is very helpful for active learning (as described in Section ) since it allows us to directly locate uncertain span predictions. On the other hand, the commonly used CRF models , can only assign a confidence score for all span predictions in the same sentence. When only specific types of entities are of interest, such global confidence scores are not useful.…”

“…In this work, we use a state-of-the-art transformer-based model, BERT, to extract span representations. Since its proposal, BERT has claimed the best performance over a variety of natural language processing benchmarks. , As shown in a recent ChEMU benchmark, , BERT-based models ,− have been widely used for chemical information extraction and become the dominating method. Lin et al provide two reasons why BERT can work well on our catalysis information extraction task.…”

“…Its data source is a domain with a very different writing style from scientific literature, and the task is not to extract catalyst-related entities. More recent studies centered around the articles related to materials science , and focused on extracting information such as summary-level information, synthesis route, and procedure parameters. − These studies were mostly constrained to certain types of paragraphs, like abstract or synthesis-related, − ,− which might greatly hinder the extraction performance. Thus, there is no suitable data set for catalysis-related information extraction from scientific literature.…”

Accelerating manufacturing for biomass conversion via integrated process and bench digitalization: a perspective

“…In contrast, Laino and co-workers 56 used deep-learning to convert experimental procedures to action sequences without human involvement. Operating between these two extremes, Barzilay and co-workers 57 recently used human intervention to validate the automated classifier of reactants, products, and operating conditions. Considering the nascency of and the complexity inherent to biomass catalysis, such a supervised learning approach could be the first step forward.…”