A new arXiv paper, "Calibrated Typed Decisions for Entity, Relation, and Event Extraction," addresses a well-known weakness in information extraction (IE): models perform strongly on the datasets they are trained on but struggle when applied to new domains or different annotation guidelines. The abstract notes that this transfer problem persists even when models are designed to accommodate user-specified types.

According to the abstract, supporting user-specified types alone is not enough to overcome the gap between training and deployment conditions. The paper's title points to "calibrated typed decisions" as the proposed direction, though the abstract is truncated and does not detail the method or experimental results.

The work highlights the need for IE systems to generalize beyond their training data, a key concern for real-world applications where annotation conventions vary. Readers looking for concrete techniques will need to consult the full paper, as the abstract only sets up the problem and the proposed approach at a high level.