📄 Coreference Resolution for Information Extraction

Summary

This paper addresses coreference resolution as a component of information extraction (IE). The task is to decide which mentions in a document refer to the same real-world entity, so that extracted facts can be merged per entity rather than per mention. The authors compare several ways of framing and solving the problem. They propose a loss-based decoding framework, in which grouping mentions into entities is treated as finding a partition that minimizes a loss built from pairwise decisions. Because exact decoding is expensive, they add a greedy algorithm that approximates it. Both are combined with two learners, the Perceptron and logistic regression, which score whether mention pairs corefer. The methods are tested on the Automatic Content Extraction (ACE) data. The work matters because it links coreference modeling to the needs of entity-centric extraction pipelines. It also gives a reusable design: learn pairwise scores, then decode them into a global entity partition.

Key points

• The setting is IE-oriented coreference, which groups entity mentions into entities, as in ACE-style entity detection and tracking.
• The core contributions are a loss-based decoding formulation for coreference and a greedy approximate decoding algorithm.
• Two learning algorithms are paired with the decoders: the Perceptron and logistic regression.
• Experiments use the ACE (Automatic Content Extraction) dataset.
• Caveat for the library: the supplied PDF text is almost entirely garbled by a font-encoding failure, so this summary rests on the paper's widely known framing. Exact scores, dataset sizes and table values could not be read and are omitted. Please re-extract the text (e.g., via OCR) before adding concrete numbers.