Introduction
Natural Language Generation (NLG) involves mapping abstract meaning representations to human-readable text. For years, Abstract Meaning Representation (AMR) has served as a standard. However, AMR is inherently "lossy," omitting surface-level details such as tense, number, and articles. While beneficial for semantic parsing, this abstraction forces generators to guess missing grammatical information. This paper explores the use of Minimal Recursion Semantics (MRS), a richer representation that preserves these details, combined with the robustness of neural sequence-to-sequence models.
Methodology
The researchers utilize a two-layer bidirectional LSTM encoder-decoder architecture with several key modifications to handle rich semantics:
- Graph Linearization: Dependency MRS (DMRS) graphs are flattened into a linear string using the PENMAN format, enabling standard sequence-to-sequence processing.
- Attribute Embedding: Morpho-semantic attributes (e.g., tense, number) are embedded separately and concatenated with predicate embeddings.
- Semi-supervised Learning: To address data scarcity, a "silver" dataset of 1 million sentences from the Gigaword corpus was generated using the high-precision ACE parser.
- Pointer Mechanism: A pointer-generator network allows the model to copy tokens directly from the input, mitigating issues with unknown words.
Experimental Results
The model was evaluated using the Redwoods Treebank (Gold data) and the Gigaword corpus (Silver data). A primary finding is that the inclusion of grammatical "fiddly bits" significantly eases the generation task for neural networks.
As shown in the results, the DMRS-based model drastically outperforms AMR-based approaches. On a shared benchmark, the MRS-based model achieved a BLEU score of 75.8, compared to just 33.8 for the best AMR model. Furthermore, training on "silver" data provided a substantial 11.6 BLEU point boost over gold data alone, indicating that structural variety is as crucial as semantic accuracy.
Discussion and Conclusion
Ablation studies confirmed the importance of rich attributes; removing tense and number information resulted in a drop of over 12 BLEU points. The best-performing configuration (Gold + Silver data) reached a state-of-the-art BLEU score of 83.37 on the Wall Street Journal subset, with a 32% exact match rate.
Despite these gains, challenges remain. The "parsing bottleneck" is significant, as approximately 20% of sentences could not be automatically converted to MRS. Future work may involve moving from graph linearization to native Graph Neural Networks (GNNs) to better preserve structural relationships.
Summary updated