Abstract
Verb Phrase Ellipsis (VPE) is a linguistic phenomenon where a verb phrase is omitted because it is implied by the context. Resolving VPE is critical for natural language understanding tasks such as event extraction and coreference resolution, which often fail when faced with incomplete syntactic structures. This paper proposes a high-performance two-stage pipeline for VPE resolution. By combining discriminative machine learning with linguistically-informed features—specifically c-command relationships and structural alignment—the authors achieved a state-of-the-art 80.78% F1 score in VPE detection and a robust 65% accuracy in antecedent identification.
Introduction
In natural language, humans frequently use ellipsis to avoid redundancy. For example, in the sentence, "The government includes money spent on residential renovation; Dodge does not," the reader understands that Dodge does not include money spent on residential renovation. For AI systems, however, these omissions often lead to "hallucinated" or incomplete data representations. A naive system might extract a nonsensical relationship like (Dodge, agent, do).
Resolving VPE requires addressing two distinct challenges:
- Detection: Determining if a word (e.g., "does", "to", "can") serves as a trigger for an ellipsis or is used in a standard verbal capacity.
- Identification: Locating the exact "antecedent" (the original verb phrase) within the preceding text.
Methodology
The authors developed a two-stage pipeline to handle these challenges sequentially.
1. VPE Detection
Detection is framed as a binary classification problem. While standard lexical features like words and part-of-speech (POS) tags are used, the system's performance is significantly enhanced by Syntactic Features. Specifically, the concept of c-command—a structural relationship within a parse tree—is used to identify triggers positioned in a way that suggests an omitted phrase follows. This linguistic grounding allows the model to outperform purely statistical approaches.
2. Antecedent Identification
Once a trigger is detected, the system must select the correct antecedent from multiple overlapping candidates. The authors adapted the Margin-Infused Relaxed Algorithm (MIRA) for this task. MIRA is particularly effective here because it handles "partial loss," allowing the model to distinguish between candidates that are entirely incorrect and those that are nearly correct but have slightly misaligned boundaries.
A key innovation in this stage is the Alignment Feature. This feature measures structural parallelism between the clause containing the trigger and the clause containing the candidate antecedent.
Experimental Setup and Results
The system was evaluated using the Wall Street Journal (WSJ) dataset (Bos and Spenader, 2011), which contains 554 VPE instances. Unlike previous research focused solely on "do-VPE," this study included a wider range of triggers, including modals and the challenging "to-VPE."
- Detection: The model achieved an F1 score of 80.78%, a significant improvement over the previous state-of-the-art of 69.52%.
- Identification: The MIRA-based approach reached 65% accuracy, outperforming a "nearest verb phrase" baseline by 29%.
- End-to-End: The full pipeline achieved an F1 of 51.96%, setting a new benchmark for comprehensive VPE resolution.
Conclusion
This research demonstrates that reintroducing classical linguistic concepts like c-command and structural parallelism into modern machine learning frameworks can yield substantial gains in complex discourse tasks. While detection has reached high levels of performance, identifying exact antecedent boundaries remains a challenge due to data scarcity. Future work will focus on integrating VPE resolution into joint models for broader discourse analysis.
Summary updated