The WEAVE 2.0 Corpus: Role Labelled Synthetic Chemical Procedures from Patents with Chemical Named Entities

Shubhangi Dutta, Manish Shrivastava, Prabhakar BhimalapuramPACLIC 2023 · OpenReview

Introduction

Chemical patents represent a vast repository of experimental data essential for drug discovery and materials science. However, this information is largely trapped in unstructured text. Traditional Named Entity Recognition (NER) systems can identify chemical names (e.g., "Ethylene Glycol"), but they often fail to capture the functional role of the chemical within a specific reaction context. Understanding whether a substance acts as a reactant, solvent, or catalyst is crucial for reconstructing complex chemical reaction pathways. This paper introduces WEAVE 2.0, a dataset and modeling framework designed to bridge the gap between simple entity recognition and functional role labeling.

The WEAVE 2.0 Dataset

WEAVE 2.0 is an extension of the original WEAVE dataset, comprising 33 manually annotated patent documents containing 17,177 entities. The corpus utilizes a sophisticated tagging scheme with 71 unique labels, combining seven naming types (e.g., systematic, formula, abbreviation) with various reaction roles. This granularity allows models to distinguish between a chemical's identity and its specific purpose in a procedure.

Example of WEAVE 2.0 annotation showing chemical types and functional roles within a patent excerpt.

Methodology

To address the joint task of identifying chemical types and roles, the researchers developed several neural architectures:

  1. ChemicalBERT Embeddings: The authors fine-tuned SciBERT on a massive corpus of 230 million lines of patent data to create domain-specific embeddings.
  2. Architectural Variants:
    • Joint Model: A multi-head architecture predicting type and role simultaneously.
  3. Two-Step Model: A pipeline where type prediction informs a subsequent role-labeling stage.
  4. Attention-Based Model: This architecture replaces the standard Conditional Random Field (CRF) layer with an attention mechanism to better decode BiLSTM encodings by focusing on contextual cues like surrounding verbs and prepositions.
The Attention-Based Model architecture, which achieved the highest performance by leveraging attention layers for decoding.
  1. Data Augmentation: To mitigate class imbalance and the "long tail" of rare labels, the authors employed strategies such as shuffling sentences and replacing named entities with different chemicals of the same role. The latter proved most effective, helping the model learn the structural patterns of chemical procedures.

Results and Discussion

The ChemicalBERT + Attention Model emerged as the top performer, achieving an F1 score of 0.87 on the WEAVE 2.0 dataset. The success of the attention mechanism underscores the importance of context; a chemical's role is often defined by its proximity to specific action-oriented words (e.g., "added to," "dissolved in").

Data augmentation was vital for performance, raising the baseline F1 scores from approximately 0.80 to the final 0.87. Furthermore, the architecture demonstrated strong generalizability, achieving an F1 score of 0.95 on the standard CHEMDNER benchmark.

Conclusion

The WEAVE 2.0 corpus and the accompanying attention-based models provide a robust foundation for automated chemical data extraction. While challenges remain regarding extreme class imbalance for rare labels, this work moves the field closer to full chemical pathway generation—enabling machines to autonomously read patents and generate executable "recipes" for chemical synthesis.

Summary updated

BibTeX
@inproceedings{dutta2023weave,
  author = {Shubhangi Dutta and Manish Shrivastava and Prabhakar Bhimalapuram},
  title = {{The WEAVE 2.0 Corpus: Role Labelled Synthetic Chemical Procedures from Patents with Chemical Named Entities}},
  booktitle = {PACLIC},
  year = {2023},
  url = {https://openreview.net/forum?id=7NSHw9LWhi},
}
Figure