Abstract
ConvAMR introduces a convolutional sequence-to-sequence (seq2seq) approach to Abstract Meaning Representation (AMR) parsing, addressing the computational inefficiencies of recurrent models. By employing a novel graph linearization technique and the parallelization capabilities of CNNs, the authors achieved a state-of-the-art Smatch score of 0.71 on standard benchmarks and established the first AMR baseline for legal documents with the JCivilCode-1.0 dataset.
Introduction
Abstract Meaning Representation (AMR) is a semantic formalism that describes the "who is doing what to whom" of a sentence using a rooted, directed graph. Converting natural language into these graphs—a process known as AMR parsing—is a significant challenge in NLP because graphs are inherently non-linear structures. While most existing neural parsers rely on Long Short-Term Memory (LSTM) networks, these models are often computationally slow and struggle with the long-distance dependencies created when a graph is flattened into a string.
The ConvAMR Methodology
The authors propose ConvAMR, a model that shifts the architecture from LSTMs to Convolutional Neural Networks (CNNs). This approach relies on two primary innovations:
- Modified Graph Linearization: To process graphs using a seq2seq model, they must be converted into a sequence of tokens. ConvAMR utilizes a depth-first search (DFS) traversal but simplifies the output by removing variables and signaling the end of a path by "doubling" the concept of the terminal node. This reduces the complexity of the target sequence compared to traditional bracketed notations.
- Convolutional Seq2Seq: CNNs allow for simultaneous processing of sequences, providing a shorter dependency path that aligns well with the local structures of a linearized graph. The authors found that a fully convolutional seq2seq model outperformed other configurations.
Datasets and Evaluation
The study evaluates ConvAMR on two distinct domains:
- LDC2014T12: The standard benchmark for AMR parsing, consisting primarily of news articles.
- JCivilCode-1.0: A new dataset introduced by the authors containing manually annotated AMR graphs for the English version of the Japan Civil Code. This represents the first public AMR resource specifically for the legal domain, which is characterized by highly complex and formal sentence structures.
Key Results and Insights
ConvAMR achieved a new state-of-the-art Smatch score of 0.71 on the LDC2014T12 dataset, a significant improvement over previous models like CAMR (0.66) and NeuralAMR (0.62). On the legal-specific JCivilCode-1.0 dataset, the model established a strong baseline score of 0.60.
Beyond raw performance, the authors introduced an "Information Loss" metric to measure the fidelity of the linearization process. They observed a loss of approximately 0.20–0.21, suggesting that the bottleneck in parsing may reside in the graph representation itself as much as the model architecture.
Limitations and Future Work
Despite its success, ConvAMR faces two main structural challenges. First is node collision, where distinct entities sharing the same concept are incorrectly collapsed into a single node during linearization. Second is syntax errors, where the model occasionally generates output that violates formal PENMAN notation. Future research will focus on refining the linearization process to preserve entity distinctions and enforcing stricter syntactic constraints during decoding.
Summary updated