The Role of Reentrancies in Abstract Meaning Representation Parsing

Marco Damonte, Ida Szubert, Shay B. Cohen, Mark Steedman

Abstract

Abstract Meaning Representation (AMR) models sentences as rooted, directed acyclic graphs to capture semantic relationships. However, modern parsers frequently struggle with "reentrancies"—nodes with multiple incoming edges representing a single concept playing multiple semantic roles. This paper identifies the linguistic triggers behind these structures and demonstrates through "oracle" experiments that fixing reentrancy errors alone can boost parsing performance (Smatch) by up to 5% and reentrancy recovery by nearly 20%.

Introduction

In Natural Language Processing, AMR is a popular formalism that represents the "who did what to whom" of a sentence. Unlike simpler tree structures, AMR allows for reentrancies. For example, in the sentence "I want to believe," the concept "I" is both the wanter and the believer, necessitating two incoming edges to the same node. Despite their importance, reentrancies remain the "Achilles' heel" of modern AMR parsers. They are difficult to predict, and the field has lacked a comprehensive breakdown of their linguistic origins. This study bridges that gap by categorizing reentrancy triggers, quantifying their frequency, and measuring the potential performance gains from perfect reentrancy handling.

A Linguistic Taxonomy of Reentrancies

The researchers classified reentrancy triggers into three primary categories to better understand why they occur:

  1. Syntactic: Forced by sentence structure. This includes coordination (e.g., "She ate and drank"), control verbs (e.g., "I asked you to sing"), and relative clauses.
  2. Pragmatic: Derived from context or world knowledge, primarily coreference where different words (e.g., "the team" and "FC Barcelona") refer to the same entity.
  3. AMR-Specific: Artifacts of the formalism itself, such as "verbalization" (converting nouns like "instructions" into predicate nodes) or specific handling of inverse roles.
Table 1: Examples of linguistic phenomena and their corresponding AMR graph structures, highlighting reentrancies.

Automated heuristics using Universal Dependency (UD) parses and coreference resolution tools were developed to label these reentrancies in standard benchmarks like LDC2015E86 and LDC2017T10. Coreference emerged as the single most common cause, appearing in roughly 37% of cases.

The Oracle Experiment

To find the "upper bound" of potential improvement, the authors created an "oracle" system. This system takes a parser's output and "fixes" specific reentrancy errors—such as missing edges or collapsed nodes—based on the ground-truth gold standard.

A key finding was the "ADD" bottleneck: the most frequent error parsers make is simply failing to draw an edge between two existing nodes. Fixing just this one type of error accounts for more than half of the potential performance gains identified in the study.

Key Results and Performance Gains

By using the oracle to correct reentrancy-related errors in a state-of-the-art parser (Lyu and Titov, 2018), the researchers observed significant improvements:

  • Smatch Score: Improved by 4.4 to 4.6 points, a substantial jump for this metric.
  • Reentrancy Recovery: The specific metric for reentrancy prediction saw a massive increase of 18 to 18.8 points.
  • Baseline Struggles: Simple automatic post-processors (using Random and Seq2Seq models) failed to improve upon the vanilla parser, suggesting that reentrancy requires deeper integration into the parsing process rather than a "quick fix."

Conclusion

The study highlights that while phenomena like "control" and "coordination" are well-known, others like adjunct control and verbalization are significant but often overlooked sources of reentrancy. The primary limitation remains the reliance on noisy external tools; automated heuristics failed to classify 46% of reentrancies due to imperfect UD parses or coreference systems. Future AMR parsers should move away from post-processing and toward models that are inherently informed by these specific linguistic triggers to bridge the identified performance gap.

Summary updated

BibTeX
@inproceedings{damonte2020role,
  author = {Marco Damonte and Ida Szubert and Shay B. Cohen and Mark Steedman},
  title = {{The Role of Reentrancies in Abstract Meaning Representation Parsing}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  doi = {10.18653/v1/2020.findings-emnlp.199},
}
Figure