📄 Learning-based Multi-Sieve Co-reference Resolution with Knowledge

Summary

The paper tackles co-reference resolution: deciding which mentions in a text point to the same entity. Many errors come from missing world knowledge, such as not knowing that "Kursk" is a submarine and can therefore be called a "vessel." The authors link mentions to Wikipedia pages using an existing context-sensitive disambiguation system. From the linked pages they derive a small, high-precision set of attributes: fine-grained semantic categories, gender and nationality. These attributes feed knowledge-based compatibility features in the pairwise model of Bengtson and Roth (2008). The authors also recast rule-based multi-sieve resolution as a sequence of learned classifiers ordered from easy to hard decisions, with two key properties:

• a later sieve may revise an earlier decision;
• entity-based features let mentions that have already been grouped share evidence.

On the non-transcript part of ACE 2004, the full system beats this strong baseline by about 3 MUC, 2 B3 and 2 CEAF F1 points. The work shows that disambiguated encyclopedic knowledge and staged learned inference can improve a competitive resolver.

Key points

• Baseline: Bengtson and Roth (2008), a pairwise averaged-perceptron scorer with Best-Link inference. It had the top B3 and BLANC scores at the CoNLL 2011 shared task. On the authors' split it reaches roughly 78.39 MUC, 83.03 B3 and 77.52 CEAF.
• Category extraction: A simple heuristic takes category heads (plural nouns before relation words like "of" or "in") and drops rare tokens. This yields 2088 fine-grained types, which are manually mapped to ACE-style classes (ORG, LOC, GPE, FACILITY).
• Gender: Gender is inferred from the pronouns in each article's first paragraph.
• Nationality: Nationality is taken from country names in the page's categories.
• Robustness of gains: The improvements hold across all decision thresholds, with a minimum of 3 MUC, 2 B3 and 1.25 CEAF F1. The extracted attributes are released publicly.
• Limitations: The system relies on gold mention boundaries and exhaustive annotation, so it was not applied to OntoNotes. The authors also describe their division into sieves as preliminary.