Abstract
Transition-based Abstract Meaning Representation (AMR) parsing often suffers from an expansive search space and noisy word-to-concept alignments. This paper introduces a refined search space through a "compact" graph representation and a hybrid alignment tool. By merging non-lexical subgraphs and enforcing structural constraints, the proposed parser achieves state-of-the-art performance, including a 74.0 Smatch score on the LDC2014T12 benchmark, while requiring fewer transitions and minimal external features.
Introduction
Abstract Meaning Representation (AMR) captures the semantic structure of a sentence—"who is doing what to whom"—using a rooted, directed graph. Unlike dependency trees, AMR graphs can feature reentrancy (nodes with multiple parents) and concepts that do not map one-to-one to surface tokens. Transition-based parsers, which build these graphs incrementally, often struggle with the resulting massive search space. This paper argues that the primary bottleneck in AMR parsing is an inefficiently defined search space and proposes two key refinements: a compact graph representation and a hybrid alignment model.
Methodology
Compact AMR Graph Representation
Standard AMR graphs often contain "non-lexical" subgraphs that represent entities or concepts not directly corresponding to individual words. For example, a generic country node might be linked to a specific name like United Kingdom. To streamline the parsing process, we compress these subgraphs into single "compact" nodes, such as ENcountry.
This compaction allows the parser to map a contiguous span of text to a single node, reducing the number of required actions by approximately 9%. Additionally, we impose structural constraints—such as acyclicity and restricted reentrancy—to prune impossible paths from the search space, guiding the parser toward valid semantic structures.
Hybrid Alignment and Transition System
Effective training of a transition-based parser requires accurate alignments between sentence words and graph nodes. We introduce a Hybrid Aligner that combines rule-based logic with unsupervised word alignment models (IBM models). This approach prioritizes "locality," ensuring that words near each other in the sentence are likely to be proximal in the semantic graph.
The parser employs a transition system with nine specific actions (e.g., SHIFT, REDUCE, MERGE, ENTITY) powered by Stack LSTMs. An attention mechanism allows the model to look ahead at words remaining in the buffer, facilitating more informed transition decisions.
Results and Discussion
The model was evaluated on standard benchmarks, including LDC2014T12, LDC2015E86, and LDC2017T10. On the LDC2014T12 Newswire dataset, the parser achieved a Smatch score of 74.0, outperforming the previous state-of-the-art by 3 points.
Key insights from the evaluation include:
- Efficiency through Compaction: Shorter action sequences reduce the opportunity for the model to drift, leading to more robust parsing.
- Alignment Accuracy: The hybrid aligner significantly reduced "null" alignments from 5.1% to 2.0%, providing cleaner training data.
- Minimal Feature Dependency: The system achieves high performance using only POS tags, whereas competitors often rely on complex pipelines involving dependency trees and named entity recognizers.
Conclusion
By refining the search space through graph compaction and improved alignment, we significantly enhance the performance and efficiency of transition-based AMR parsing. Future work will focus on character-based representations to better handle morphological cues and negation, as well as exploring latent alignment models to further integrate the alignment and parsing phases.
Summary updated