AMR dependency parsing with a typed semantic algebra

Jonas Groschwitz, Matthias Lindemann, Meaghan Fowlie, Mark Johnson, Alexander Koller

Abstract

We present a semantic parser for Abstract Meaning Representations which learns to parse sentences 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 semantic relationships in a sentence, specifically "who is doing what to whom." Unlike traditional syntactic trees, AMRs are directed acyclic graphs. This graph structure is necessary to represent reentrancy—instances where a single entity participates in multiple semantic roles. For example, in the sentence "The boy wants to sleep," the entity "boy" is the subject of both "wanting" and "sleeping."

Parsing strings directly into graphs is computationally complex and often lacks the structural guarantees provided by tree-based parsing. This paper addresses the challenge by bridging the "graph-tree gap" using a "typed semantic algebra."

The Apply-Modify (AM) Algebra

The core innovation of this work is the Apply-Modify (AM) algebra, which decomposes complex AMR graphs into a set of atomic fragments called "as-graphs." These fragments are combined using two primary operations:

  1. Apply (APP): Fills a required argument slot (e.g., a subject or object).
  2. Modify (MOD): Attaches a modifier (e.g., an adverb or adjective).
Atomic graph fragments (as-graphs) representing lexical items like 'want', 'write', and 'sleep'.

Each fragment is assigned a "type" that acts as a valency constraint, ensuring that only semantically valid combinations are formed. By representing the sequence of these operations as a dependency tree over the words in a sentence, the authors transform the graph-parsing problem into a more tractable dependency parsing task.

Neural Architecture and Decoding

The system utilizes a neural pipeline consisting of a BiLSTM-based supertagger to predict the most likely atomic fragment for each word and a neural dependency parser to predict the APP or MOD relationships between them.

Because finding the optimal well-typed tree is NP-complete, the authors propose two decoding approximations:

  • Fixed-tree decoder: Labels a pre-determined unlabeled dependency tree.
  • Projective decoder: Uses dynamic programming to find the best tree under the assumption of projectivity (no crossing edges).

Results and Insights

The model was evaluated on the LDC2015E86 and LDC2017T10 benchmarks using the Smatch f-score. On the 2017 dataset, the parser achieved a score of 71.0, matching the state-of-the-art. On the 2015 dataset, it reached 70.2, outperforming most previous models.

A key insight is how the AM algebra handles reentrancy. By lexicalizing reentrancy within the atomic fragments themselves, the global structure of the derivation remains a tree, even if the resulting AMR is a graph.

A composed AMR graph for 'want to sleep soundly', illustrating how local reentrancy (the 's' node) is resolved.

Conclusion

By combining the efficiency of dependency parsing with a linguistically principled type system, this approach provides a robust framework for semantic parsing. While challenges remain in handling long-range coreference and improving decoding speed for long sentences, the AM algebra demonstrates that explicit compositionality is a powerful tool for generating complex semantic representations.

Summary updated

Abstract

We present a semantic parser for Abstract Meaning Representations 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.
BibTeX
@inproceedings{groschwitz2018amr,
  author = {Jonas Groschwitz and Matthias Lindemann and Meaghan Fowlie and Mark Johnson and Alexander Koller},
  title = {{AMR dependency parsing with a typed semantic algebra}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2018},
  url = {https://openreview.net/forum?id=rk4hBslubS},
}
Figure