Introduction
Semantic parsing—the task of translating natural language into machine-readable logical forms—is essential for advanced natural language understanding. While frameworks like Abstract Meaning Representation (AMR) are popular, they often abstract away linguistic nuances such as tense, aspect, and non-intersective modifiers. Unscoped Episodic Logical Forms (ULF) preserve these details, offering a richer semantic foundation. However, ULF parsing has historically relied on rule-based systems that lack scalability. This paper introduces the first neural transition-based parser for ULF, addressing the challenges of complex representation and limited annotated data.
Methodology: Node Generative Transition System
The authors propose a novel transition-based architecture designed for the compositional nature of ULF. Unlike standard systems that use a fixed vocabulary, this "node generative" system derives logical atoms directly from input words in the buffer.
System Architecture
The model employs a multi-component neural network:
- Encoders: A BiLSTM processes the input sentence, while a separate LSTM encodes the sequence of generated logical symbols.
- Action Decoder: A third LSTM predicts parsing actions (e.g., GEN, PUSH, ARC) based on the current state and "hard attention" over the input encoders.
- Transition Logic: The system uses a modified cache to manage local dependencies, building the logical tree in a bottom-up fashion. This approach ensures that complete constituents are formed before higher-level structural decisions are made.
Constrained Decoding
To mitigate data sparsity, the researchers explored two types of constraints: a lexicon to restrict symbol generation to known mappings, and a type-composition grammar to ensure the parser only constructs logically valid structures. Interestingly, while these constraints increased precision, they often hurt recall by being too rigid, causing the parser to produce fragmented results on complex sentences.
Dataset: The First ULF Corpus
A significant contribution of this work is the release of the first official annotated ULF dataset. Comprising 1,738 sentences, the corpus draws from diverse sources including Tatoeba (varied translations), Project Gutenberg (literature), UIUC Question Classification (factoid questions), and Discourse Graphbank (news). This variety ensures the model is exposed to a wide range of linguistic structures.
Results and Discussion
The proposed model establishes a strong baseline for ULF parsing, significantly outperforming existing semantic parsers adapted for the task.
Quantitative Performance
The "best model" achieved a SEMBLEU score of 47.4 and an EL-SMATCH F1 score of 59.8. As shown in the table below, it substantially exceeds the performance of state-of-the-art AMR parsers when applied to ULF data.
Insights on Features
Ablation studies revealed that traditional features like Part-of-Speech (POS) tags and GloVe embeddings are crucial for performance in this domain. Conversely, large pre-trained models like RoBERTa provided only marginal gains, suggesting that the specific structural requirements of ULF parsing may benefit more from specialized linguistic features than general-purpose embeddings in low-data regimes.
Conclusion
This work marks a significant step forward in deep semantic parsing by providing both a robust neural architecture and a foundational dataset for ULF. While data scarcity and the rigidity of logical constraints remain challenges, the transition-based approach demonstrates that capturing rich linguistic nuances is achievable through specialized machine-learning models. Future research will focus on expanding the annotated corpus and developing "soft" constraints to guide the parser more flexibly.
Summary updated