Abstract
This paper revitalizes Synchronous Hyperedge Replacement Grammars (SHRG) for semantic parsing by explicitly modeling the syntacto-semantic composition process. By inducing rules that synchronize string grammars with graph grammars, the authors bridge the gap between interpretable, linguistically-grounded frameworks and high-accuracy neural models. The proposed parser achieves state-of-the-art results on EDS and DMRS representations, demonstrating that structural compositionality remains a powerful tool in semantic analysis.
Introduction
Semantic parsing involves translating natural language into machine-readable logic or graphs. While sequence-to-sequence models are widely used, they often struggle to capture the deep, structural relationship between a sentence's syntax (word arrangement) and its semantics (meaning). Previous attempts to utilize SHRG—a mathematically elegant framework for graph generation—were often outperformed by greedy transition systems or pure neural architectures.
The authors argue that the limitations of previous SHRG applications stemmed from rule induction and utilization rather than the framework itself. They propose a neural SHRG parser that models the "syntacto-semantic" composition process, proving that linguistically-informed grammars can compete with and even outperform "black-box" neural models.
Methodology
The authors propose a two-stage neural SHRG parser consisting of grammar extraction and neural decoding.
Grammar Extraction
A novel algorithm induces SHRG rules from English Resource Semantics (ERS) by aligning "conceptual edges" in a semantic graph with specific spans in a syntactic tree. This process synchronizes a Context-Free Grammar (CFG) for strings with a Hyperedge Replacement Grammar (HRG) for graphs.
The Neural Parser
- Syntactic Parsing: An "LSTM-Minus" approach builds a constituent parser using a CKY decoder to identify the optimal syntactic tree based on vector representations of word spans.
- Semantic Interpretation: The model translates the syntactic tree into a semantic graph by assigning SHRG rules to each node. Decoders using Greedy Search and Bottom-Up Beam Search find the best rule combinations.
- Scoring: Multi-Layer Perceptrons (MLPs) score potential rules based on span embeddings and predicate-argument relationships.
Experimental Setup and Datasets
The study utilizes DeepBank, a gold-standard corpus of Wall Street Journal articles annotated with the English Resource Grammar (ERG). Evaluation focuses on two graph-structured representations:
- EDS (Elementary Dependency Structure)
- DMRS (Dependency-based Minimal Recursion Semantics)
These results are compared against benchmarks from the 2015 SemEval task on semantic dependency parsing.
Results
The SHRG-based parser established a new state-of-the-art for data-driven models, as shown in the table below. It achieved an Elementary Dependency Match (EDM) score of 90.35 on EDS (a 4.87-point improvement) and 89.51 on DMRS (a 5.45-point improvement). Notably, while traditional symbolic parsers often suffer from limited coverage (typically around 89%), this neural SHRG parser provides 100% coverage while maintaining high precision.
Discussion and Key Insights
- Compositionality: The success of SHRG is tied to the compositionality of the semantic representation. ERS is highly compositional, making it a natural fit for SHRG. In contrast, representations like AMR include non-compositional elements (e.g., co-reference) that pose greater challenges for grammar-based frameworks.
- Construction vs. Lexicalized Grammars: Experiments revealed that a construction-based approach (deriving meaning from phrases/rules) significantly outperformed lexicalized grammars (storing meaning in individual words). This suggests that modeling the process of combining words is more effective than analyzing words in isolation.
- Neural-Linguistic Synergy: The parser demonstrates that neural models can effectively learn deep linguistic structures, matching the accuracy of grammar-guided symbolic models while ensuring full coverage for any input sentence.
Limitations and Future Work
The model's reliance on high-quality alignments between surface strings and semantic concepts may limit its portability to languages or domains without resources like DeepBank. Future research may explore different syntactic formalisms beyond constituent-based trees and investigate methods to handle non-local semantic dependencies in less compositional representations like AMR.
Summary updated