Abstract
We present a semantic parser for Abstract Meaning Representations (AMR) which learns to parse strings into tree representations of the compositional structure of an AMR graph. This allows us to use standard neural techniques for supertagging and dependency tree parsing, constrained by a linguistically principled type system. We present two approximative decoding algorithms, which achieve state-of-the-art accuracy and outperform strong baselines.
Introduction
Abstract Meaning Representation (AMR) is a popular framework for capturing the semantic structure of a sentence—the "who did what to whom." Unlike syntactic trees, AMRs are graphs that include "reentrancies," where a single concept participates in multiple relations. For example, in the sentence "The boy wants to sleep," the concept "boy" is both the wanter and the sleeper.
Parsing strings into graphs is significantly more difficult than parsing them into trees. Traditional grammar-based approaches often struggle with the complexity of graph decomposition, while neural sequence-to-sequence models may fail to respect inherent structural constraints. This paper addresses these challenges by making the compositional structure of the AMR explicit, allowing the use of efficient, well-understood tree-parsing techniques to produce complex semantic graphs.
Methodology: The Typed Semantic Algebra
The core innovation of this work is the Apply-Modify (AM) algebra, which treats AMR construction as a dependency parsing task.
Algebraic Terms as Trees
The authors view an AMR as a composition of "atomic graphs," each representing an individual word. These building blocks are combined using two primary operations:
- Apply: Fills an argument slot (e.g., providing a subject for a verb).
- Modify: Adds an adjunct or modifier (e.g., an adverb modifying a verb).
These operations form an "AM dependency tree," where nodes are words and edges are algebraic operations. This transformation effectively reduces graph parsing to a tree-parsing problem.
The Type System
To ensure the resulting graphs are linguistically and structurally valid, the parser uses a principled type system. For instance, a transitive verb is assigned a type indicating it requires both a subject and an object. This system acts as a constraint, preventing the parser from making structurally "illegal" moves.
Neural Architecture and Decoding
The model employs a Bi-LSTM Supertagger to predict the atomic graph and type for each word, and a Dependency Parser to predict the algebraic operations between them. Since finding the optimal well-typed tree is NP-complete, the authors propose two approximation algorithms: a Fixed-tree decoder and a Projective decoder based on dynamic programming.
Key Results and Insights
The system was evaluated on standard AMR benchmarks (LDC2015E86 and LDC2017T10), achieving state-of-the-art Smatch F-scores of up to 71.0.
Key insights from the evaluation include:
- Type-Awareness is Essential: A baseline parser without the type system performed poorly (Smatch ~26), proving that semantic types are the "glue" of this approach.
- Projectivity Assumptions: Despite many gold-standard trees being non-projective, the projective decoder performed remarkably well, often finding alternative type assignments to compensate for structural limitations.
- Efficiency: By leveraging dependency parsing, the system benefits from the speed and maturity of existing NLP toolkits.
Conclusion and Future Work
This paper demonstrates that a typed semantic algebra can effectively bridge the gap between string-to-tree parsing and complex graph generation. Future work will explore "latent alignment" to learn token-to-node mappings during training and aim to improve the handling of long-range reentrancies caused by coreference, which currently remain a challenge for compositional models.
Summary updated