Abstract
Abstract Meaning Representation (AMR) parsing and generation have traditionally been treated as distinct tasks requiring complex, heuristic-driven pipelines. We present SPRING (Symmetric PaRsIng aNd Generation), a unified sequence-to-sequence approach based on the pretrained BART model. By employing a novel DFS-based graph linearization and removing hand-crafted graph recategorization rules, SPRING achieves state-of-the-art results in both parsing and generation across multiple benchmarks, demonstrating superior robustness in out-of-distribution scenarios.
Introduction
Abstract Meaning Representation (AMR) captures "who is doing what to whom" in a sentence using a rooted, directed graph. Historically, the field has treated AMR parsing (text-to-graph) and AMR generation (graph-to-text) as asymmetrical problems. While generation often utilized sequence-to-sequence (seq2seq) models, parsing typically involved intricate pipelines, external syntactic features, and "graph recategorization"—heuristics designed to simplify graphs for models. This paper introduces SPRING, a symmetric architecture that handles both directions efficiently by leveraging large-scale pretraining and a refined linearization strategy.
Methods
The SPRING model is built upon three primary pillars:
BART as the Backbone
The researchers utilize BART, a pretrained Transformer encoder-decoder. Unlike models that only pretrain the encoder (e.g., BERT), pretraining the decoder is essential for generating the structured, formal language required by AMR graphs.
DFS-based Linearization
To process graphs within a seq2seq framework, they must be "flattened" into strings. SPRING employs a Depth-First Search (DFS) approach using special pointer tokens (e.g., <R0>, <R1>) to manage co-reference (nodes appearing multiple times). This representation is isomorphic, ensuring the graph can be perfectly reconstructed without information loss.
Eliminating Heuristics
SPRING avoids traditional "graph recategorization." By treating AMR labels directly and relying on an expanded vocabulary, the model avoids the brittleness associated with hand-coded rules, leading to better generalization.
Experimental Setup
The model was trained on standard benchmarks (AMR 2.0/3.0) and silver data from the Gigaword corpus. The training configuration followed specific hyperparameter settings to ensure stability and performance.
Results and Insights
SPRING achieved significant improvements over previous state-of-the-art (SOTA) systems:
- Parsing: Reached 83.8 Smatch on AMR 2.0, a 3.6-point gain.
- Generation: Achieved 45.3 BLEU on AMR 2.0, an 11.2-point increase.
- Robustness: On the "Bio" out-of-distribution dataset, SPRING outperformed heuristic-based models by over 20 Smatch points.
The authors found that DFS-based linearization is more "language-like" and compact than alternatives. Crucially, they demonstrated that while recategorization helps on standard benchmarks, it degrades performance on out-of-distribution data.
Conclusion
SPRING proves that a single, symmetric seq2seq architecture can outperform complex, task-specific pipelines in AMR. By leveraging pretrained language models and isomorphic linearization, it provides a robust foundation for semantic parsing and generation. Future work includes extending this approach to cross-lingual settings and other semantic formalisms.
Summary updated