Abstract
Pascal Denis, Jason Baldridge. Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference.
2007.
Introduction
Coreference resolution—identifying which mentions in a text (such as "he," "the president," or "Barack Obama") refer to the same real-world entity—is a fundamental task in natural language processing. Historically, systems have struggled with two primary hurdles: implicit anaphoricity and greedy local decision-making.
Most systems attempt to find an antecedent for every mention, leading to "false positives" for non-anaphoric mentions (e.g., the pleonastic "it" in "It is raining"). While some systems use an "anaphoricity filter" beforehand, this creates a brittle cascade where early errors propagate through the entire pipeline. Furthermore, making pairwise decisions independently ignores the global discourse context, often resulting in inconsistent or suboptimal coreference chains because the system cannot evaluate the document as a whole.
Methodology
The authors propose a global optimization framework using Integer Linear Programming (ILP) to jointly determine anaphoricity and coreference. This approach moves away from greedy heuristics toward a more holistic inference process.
Base Models
The system utilizes two standard Maximum Entropy (MaxEnt) classifiers:
- Coreference Classifier: Predicts the probability that a pair of mentions corefer.
- Anaphoricity Classifier: Predicts the probability that a single mention is anaphoric (i.e., it refers to a previous entity).
ILP Formulations
- COREF-ILP: This baseline ILP model finds the global assignment that maximizes the coreference classifier's scores, allowing for multiple antecedents and moving beyond simple "closest-first" approaches.
- JOINT-ILP: The primary contribution of this work. It combines the objective functions of both the anaphoricity and coreference classifiers into a single optimization problem. By introducing logical constraints, it ensures consistency:
- If a pair is marked as coreferent, the mention must be labeled anaphoric.
- If a mention is labeled anaphoric, it must have at least one antecedent.
- If a mention is labeled non-anaphoric, it cannot have any antecedents.
This joint optimization allows high-confidence coreference links to "overrule" weak anaphoricity predictions and vice versa.
Experimental Setup
The approach was evaluated on the ACE (Automated Content Extraction) Phase 2 corpus across three distinct genres: Broadcast News (BNEWS), Newspaper (NPAPER), and Newswire (NWIRE). To isolate the performance of the resolution logic from mention detection errors, the researchers focused specifically on "true" ACE mentions.
Results and Discussion
The JOINT-ILP system demonstrated superior performance compared to both the baseline pairwise classifier and the traditional cascaded approach across all datasets.
As shown in Figure 1, the cascaded system (AC-CASCADE) often underperforms the pairwise baseline in terms of F-score, highlighting the risks of hard filtering. The Oracle-Link results indicate a significant performance ceiling (~82% F-score), suggesting that while the optimization framework is effective, the underlying local features still have room for improvement.
Figure 2 illustrates the impact of the ILP framework. The JOINT-ILP model achieved absolute F-score gains of 3.7% to 5.3% over the base coreference classifier. A key driver of this improvement was a substantial increase in recall. For example, in the NPAPER dataset, recall rose from 58.1% to 68.0%. By optimizing globally, the system identifies coreference chains that local, greedy systems typically miss.
Key Insights
- Jointness vs. Cascades: Treating anaphoricity and coreference as co-dependent tasks allows for bidirectional information flow, avoiding the pitfalls of brittle pipelines.
- Declarative Constraints: ILP provides a powerful mechanism to enforce linguistic rules (e.g., "no antecedents for non-anaphors") within a mathematical optimization framework.
- Efficiency: The authors note that ILP is significantly more efficient than other global models like Conditional Random Fields (CRFs), particularly when modeling long-distance dependencies between mentions.
Conclusion
This work demonstrates that recasting coreference resolution as a global optimization problem using ILP significantly improves performance, particularly in terms of recall. Future research could further enhance this model by explicitly enforcing transitivity (if A=B and B=C, then A=C) and incorporating hard linguistic constraints—such as gender and number agreement—directly into the ILP objective function.
Summary updated