📄 RelaxCor: A Global Relaxation Labeling Approach to Coreference Resolution

Summary

The paper presents RelaxCor, the system entered by Sapena, Padró and Turmo in SemEval-2010 Task 1, which covered intra-document coreference resolution in several languages. The system treats coreference as graph partitioning. Mentions become vertices, and edges between mention pairs carry weights built from learned constraints. To obtain these constraints, the authors train C4.5 decision trees on pairwise feature vectors and convert them into rules. Separate models are built for pronoun, named-entity and nominal anaphors. Each rule's weight reflects how often the pairs it covers actually corefer. A relaxation labeling algorithm then iteratively assigns each mention to a partition, considering all mentions jointly. This keeps decisions consistent and enforces transitivity.

RelaxCor took part only in the gold-mention setting, for Catalan, Spanish and English. It used no external resources apart from WordNet in the English-open run. The global formulation yields very high precision, which gives strong CEAF and B³ scores, but low recall hurts MUC. Even so, it achieved several of the best results in the task, showing that one language-independent method can be competitive across languages.

Key points

• Constraint weights are calculated as the fraction of coreferent training pairs a rule covers, minus 0.5, so weights can be negative. In the graph, a mention with index i can only take one of the first i partition labels.
• Negative pairs made up roughly 94% of the training examples. To make learning feasible, negatives were kept only if they differed from some positive example in at most d = 3 features.
• Test CEAF F1 was 70.5 for Catalan, 66.6 for Spanish, 75.6 for English-closed and 75.8 for English-open. B³ F1 was 79.9, 78.2, 84.5 and 84.6, while MUC F1 stayed low, ranging from 24.7 to 42.5.
• RelaxCor ranked best on Catalan (CEAF, B³), English-closed (CEAF, B³), English-open (B³) and Spanish (B³). It had the highest precision in nearly all metric and language combinations.
• Adding WordNet for English gave only slight, non-significant gains. The authors identify missing semantic and world knowledge, especially for nominal mentions, as the main limit on recall.