Learning compositional structures for semantic graph parsing

Jonas Groschwitz, Meaghan Fowlie, Alexander Koller

Semantic parsing—the task of translating natural language into formal meaning representations like Abstract Meaning Representation (AMR)—is a cornerstone of natural language understanding. A leading technique in this field is AM (Apply-Modify) dependency parsing. This approach treats graph construction as a compositional process, predicting a tree of operations that builds a complex semantic graph from the meanings of individual words.

However, a significant bottleneck has persisted: training these parsers traditionally requires "gold" compositional trees. Since standard datasets only provide the sentence and the final graph, researchers previously had to write thousands of lines of expert-heavy, dataset-specific heuristics to reverse-engineer the necessary training steps. This paper introduces a neural latent-variable model that automates this process, discovering optimal compositional structures directly from data.

Methodology: The Latent-Variable Approach

The authors propose a model that treats the compositional tree as a hidden (latent) structure. The framework operates in three primary stages:

  1. Graph Decomposition: A graphbank-independent algorithm "unrolls" semantic graphs into trees. This identifies the basic shape of the composition but leaves specific "source names" (placeholders for slots like subject or object) undecided.
  2. Tree Automata: Since multiple valid sequences of operations can produce the same graph, the model uses a tree automaton to represent every possible valid compositional tree for a given graph.
  3. Joint Learning: Instead of training on a single "correct" tree, the neural parser is trained to maximize the inside score—the total probability of all valid trees that result in the target graph. This allows the model to gravitate toward structures that are consistent and easy to learn across the dataset.
Graph fragments (blobs) representing the semantic roles and argument structures for different lexical items, including active (G-charm) and passive (G-charmP) variants.

Key Insights and Results

Syntactic Consistency

One of the most striking findings is that the model does not merely choose the mathematically simplest trees. Instead, it learns structures that are remarkably "syntactically consistent." For example, it learns to assign the same semantic "slot" to a word's subject regardless of whether the sentence is active or passive. This suggests that the latent-variable approach naturally finds a "sweet spot" where semantic operations align with the underlying syntax of the language.

Performance and Generalization

The model was evaluated across four major semantic graphbanks: DM, PAS, PSD (from SemEval 2015), and AMRBank. The results demonstrate that this automated approach is highly competitive:

  • Accuracy: On the DM dataset, the model achieved an F1 score of 94.5, slightly exceeding the 93.9 achieved by previous heuristic-heavy systems.
  • Robustness: The model performed well on out-of-domain data, indicating that the learned structures capture genuine linguistic patterns rather than just dataset-specific noise.
  • Efficiency: By replacing manual rules with a latent-variable model, the authors significantly reduced the "expert burden." Only a minimal set of "blob" heuristics (defining which edges belong to which word) remains necessary.
A compositional structure where 'begin' and 'glow' share an argument ('fairy'), illustrating how the model links individual word meanings into a final semantic graph.

Limitations and Future Directions

While the model automates the bulk of the pipeline, it still relies on basic "blob" heuristics to partition edges to nodes. Future research may aim to automate this final manual step. Additionally, the model faces challenges with non-compositional phenomena like coreference in AMR, where performance drops slightly without pre-processing. Finally, while the "inside" algorithm is efficient, training currently takes roughly 2.5 times longer than standard supervised methods, making convergence speed a priority for future optimization.

In conclusion, this work represents a major step toward making semantic parsing more portable and less dependent on human expertise, proving that neural models can autonomously discover the compositional logic of human language.

Summary updated

Abstract

AM dependency parsing is a method for neural semantic graph parsing that exploits the principle of compositionality. While AM dependency parsers have been shown to be fast and accurate across several graphbanks, they require explicit annotations of the compositional tree structures for training. In the past, these were obtained using complex graphbank-specific heuristics written by experts. Here we show how they can instead be trained directly on the graphs with a neural latent-variable model, drastically reducing the amount and complexity of manual heuristics. We demonstrate that our model picks up on several linguistic phenomena on its own and achieves comparable accuracy to supervised training, greatly facilitating the use of AM dependency parsing for new sembanks.
BibTeX
@inproceedings{groschwitz2021learning,
  author = {Jonas Groschwitz and Meaghan Fowlie and Alexander Koller},
  title = {{Learning compositional structures for semantic graph parsing}},
  booktitle = {SPNLP @ ACL},
  year = {2021},
  doi = {10.18653/v1/2021.spnlp-1.3},
}

Loading related content...

Figure