Abstract
Recognizing bridging anaphora is difficult due to the wide variation within the phenomenon, the resulting lack of easily identifiable surface markers and their relative rarity. We develop linguistically motivated discourse structure, lexico-semantic and genericity detection features and integrate these into a cascaded minority preference algorithm that models bridging recognition as a subtask of learning finegrained information status (IS). We substantially improve bridging recognition without impairing performance on other IS classes.
Introduction
Bridging anaphora (or associative anaphora) represents a significant challenge in Natural Language Processing. Unlike coreference resolution, where entities refer to the same referent (e.g., "the CEO" and "he"), bridging involves entities that are related but distinct. For instance, in the sequence "One building was upgraded... a resident called up the stairs," "the stairs" is a bridging anaphor linked to "building."
Recognizing these links is vital for document coherence but is hampered by three main factors:
- Lack of surface markers: No obvious keywords like "it" signal the link.
- Diversity: Relations range from part-whole to professional roles.
- Sparsity: Bridging is a "minority class," comprising only about 6% of noun phrases (NPs) in news text, often leading standard models to ignore them in favor of overall accuracy.
Methodology
The proposed solution involves a combination of a rich linguistic feature set and a specialized cascading classification algorithm.
Rich Linguistic Features
The authors introduced several novel feature groups to capture the nuances of bridging, including discourse structure (identifying "coherence gaps"), lexico-semantic features (using WordNet and the General Inquirer), and genericity detection to distinguish bridging from generic statements.
Cascading Collective Classification (CMPS)
To address the minority class problem, the authors moved beyond standard multi-class classification to a cascaded preference system. Five binary classifiers were trained for the rarest categories. The system evaluates the rarest class first; if a positive label is assigned, the process stops. Otherwise, it proceeds to the next rarest class, only defaulting to a standard multi-class model for common "majority" classes.
Experimental Setup
The system was evaluated on a corpus of 50 texts from the Wall Street Journal (WSJ) portion of OntoNotes, containing approximately 11,000 annotated NPs and 663 bridging instances. Experiments utilized gold-standard mentions to isolate classifier performance.
Results and Discussion
The cascading architecture and rich features led to a dramatic improvement in bridging recognition. The baseline F1-score of 18.9% was more than doubled to 42.2% with the full CMPS system. Overall accuracy also improved from 76.8% to 78.6%.
Key insights from the study include:
- Beyond Definites: Less than 40% of bridging anaphora were definite NPs, highlighting the importance of including indefinite and bare nouns.
- Coherence as a Signal: The "coherence gap" proved to be a powerful indicator of bridging.
- Algorithm Impact: While new features helped, the cascading architecture was crucial for forcing the model to prioritize rare classes rather than defaulting to more common labels.
Conclusion
This work demonstrates that combining deep linguistic features with a classification strategy tailored for imbalanced data can significantly improve the recognition of rare discourse phenomena like bridging. Future work will address the challenge of resolution (identifying the specific antecedent) and the noise inherent in automatically extracting complex features like genericity.
Summary updated