[Paper Review] Using Syntax-Based Machine Translation to Parse English into Abstract Meaning Representation

Michael Pust, Ulf Hermjakob+3 more|arXiv (Cornell University)|Apr 24, 2015

Natural Language Processing Techniques31 references13 citations

TL;DR

This paper proposes a syntax-based machine translation (SBMT) framework for parsing English into Abstract Meaning Representation (AMR), transforming AMR into a tree-like structure compatible with SBMT rule extraction and decoding. By introducing AMR-specific reordering, language modeling, and semantic resource integration, the method achieves a 7 Smatch point improvement over state-of-the-art results on the AMR 1.0 corpus.

ABSTRACT

We present a parser for Abstract Meaning Representation (AMR). We treat English-to-AMR conversion within the framework of string-to-tree, syntax-based machine translation (SBMT). To make this work, we transform the AMR structure into a form suitable for the mechanics of SBMT and useful for modeling. We introduce an AMR-specific language model and add data and features drawn from semantic resources. Our resulting AMR parser improves upon state-of-the-art results by 7 Smatch points.

Motivation & Objective

• To address the challenge of automatically parsing English sentences into Abstract Meaning Representation (AMR) structures.
• To adapt syntax-based machine translation (SBMT) techniques—originally designed for string-to-string translation—to the string-to-tree parsing task of AMR.
• To overcome structural and representational differences between SBMT and AMR, such as graph-structured targets, labeled edges, and unordered children.
• To improve parsing accuracy by integrating AMR-specific language models and external semantic resources like WordNet.
• To maximize Smatch score through tuned feature parameters and reordering strategies tailored to AMR’s unordered child nodes.

Proposed method

• Transforming AMR graphs into a syntax-style tree structure using intermediate preterminal nodes and role-based relabeling to enable GHKM rule extraction.
• Introducing a target-side reordering technique that exploits the unordered nature of AMR child nodes to improve decoding efficiency and accuracy.
• Developing an AMR-specific language model trained on the yield of AMR trees to better capture semantic and syntactic patterns.
• Integrating semantic knowledge from WordNet and other resources by propagating sense counts and using category-based scoring to guide concept labeling.
• Applying MIRA learning to tune feature weights for maximum Smatch score on the development set, using both unsupervised and rule-based alignments.
• Using a beamed bottom-up chart decoder with statistical and indicator features to generate optimal derivations from input English sentences.

Experimental results

Research questions

• RQ1 Can syntax-based machine translation be effectively adapted to the task of parsing English into Abstract Meaning Representation (AMR)?
• RQ2 How can the inherent structural differences between AMR (a directed acyclic graph) and SBMT (a tree-based system) be reconciled through representation transformation?
• RQ3 To what extent can AMR-specific language modeling and semantic knowledge sources improve parsing performance beyond standard SBMT components?
• RQ4 Does target-side reordering, leveraging unordered child nodes in AMR, lead to significant gains in Smatch score?
• RQ5 Can the integration of multiple alignment types (unsupervised and rule-based) further enhance parsing accuracy?

Key findings

• The proposed AMR parsing system achieves a Smatch score of 74.1 on the AMR 1.0 test set, representing a 7-point improvement over the prior state of the art.
• Transforming AMR into a tree-compatible format enables effective rule extraction and decoding using the GHKM algorithm, forming a viable baseline.
• The introduction of role-based restructuring and relabeling significantly enhances parsing quality, especially when combined with reordering.
• Incorporating an AMR-specific language model and semantic resources like WordNet leads to a measurable performance boost.
• Using both unsupervised and rule-based alignments in tuning improves results, with concatenated alignments yielding the best performance.
• The system demonstrates that SBMT can be rapidly adapted to semantic parsing tasks with minimal architectural changes, provided proper data transformation and domain-specific enhancements.