Abstract: Yuanliang Meng, Anna Rumshisky. Proceedings of the 27th International Conference on Computational Linguistics (COLING).
2018.
Introduction
Coreference resolution is a fundamental task in natural language processing that involves identifying all mentions in a text that refer to the same real-world entity. For instance, in the sequence "Obama finished his speech; the president then sat down," a system must recognize that all three bolded terms are coreferent.
Traditionally, coreference systems rely on a pairwise (dyadic) approach, evaluating pairs of mentions in isolation to determine their relationship. However, this method is inherently myopic. Language often exhibits transitivity: if mention A refers to the same entity as B, and B refers to C, then A must refer to C. While traditional systems attempt to enforce these logical constraints during a post-processing clustering phase, they rarely incorporate these three-way dependencies during the initial feature extraction and representation learning phase. This paper introduces a Triad-based Neural Network to address this "low-level" blindness by evaluating triplets of mentions simultaneously.
Methodology
The proposed architecture moves beyond simple pairs by feeding three mentions ($i, j, k$) into the model at once. The process follows several key stages:
- Feature Extraction: Mentions are represented using GloVe word embeddings and Part-of-Speech (POS) tags, including a context window of five words before and after each mention. These features are processed through shared bidirectional LSTMs.
- Pairwise and Triad Representations: Within each triad, the model calculates internal representations for the three possible pairs ($ij, jk, ki$). Crucially, these are summed into a single "triad context" vector. This shared context allows the decision for any specific pair (e.g., $ij$) to be informed by the relationships observed in the other two pairs ($jk$ and $ki$).
- Decoding and Clustering: Three decoders produce a 3D binary vector representing the links between the triad members. To resolve entities across a full document, the system uses hierarchical clustering. Since a single pair $(a, b)$ may appear in multiple triads, the system averages scores from all relevant triads to produce a final "affinity score."
Experimental Evaluation
The model was evaluated on the English portion of the CoNLL 2012 Shared Task, the standard benchmark for coreference resolution. The Triad model achieved state-of-the-art (SOTA) results on two of the three primary metrics: MUC and B³.
As shown in the results table, the Triad model significantly outperformed previous benchmarks in MUC (80.92 F1) and B³ (65.65 F1). However, it showed lower performance on the CEAFφ₄ metric. The authors attribute this to their "single linkage" clustering strategy, which tends to over-merge entities—a behavior tolerated better by MUC and B³ than by CEAF.
Key Insights and Ablation Study
A critical finding of the research is that transitivity should be treated as a learnable feature rather than a hard-coded rule. By considering three mentions together, the neural network learns how a third "bridge" mention can clarify ambiguous relationships between the first two.
To validate the necessity of the triad structure, the authors conducted an ablation study comparing the Triad model to a "Dyad" (pairwise) version of the same architecture.
The comparison demonstrates that the Triad approach consistently outperforms the Dyad approach across almost all metrics, regardless of the maximum distance allowed between mentions. This confirms that capturing multi-mention dependencies at the representation level is the primary driver of the system's performance gains.
Limitations and Future Work
While effective, the triad approach introduces computational challenges, as the number of possible triads grows cubically ($O(n^3)$) with the number of mentions. To maintain practicality, the authors limited the distance between mentions in a triad to a window of 40. Additionally, the current clustering method's tendency to over-merge suggests that future work could benefit from more sophisticated ranking algorithms or refined clustering logic to improve CEAF scores. The authors also suggest that the logic could theoretically be extended to higher-order "polyads" (e.g., tetrads), though the computational cost would require even stricter candidate pruning.
Summary updated