Introduction: The Challenge of Fragmented Formalisms
Semantic parsing—the task of translating natural language into formal meaning representations—is currently a fragmented field. Different linguistic theories have produced distinct "formalisms" for representing meaning:
- Span-based (e.g., FrameNet): Meaning is tied to specific chunks of text (spans) that fill roles in a "frame." For example, in "The cat sat on the mat," "The cat" might be labeled as the Agent.
- Dependency-based (e.g., DM): Meaning is represented as a graph of directed, labeled arcs between individual words.
Ideally, a model should learn from all available data to understand language better. However, most joint learning techniques require parallel data (the same sentences annotated in both styles). In reality, most datasets are disjoint: FrameNet annotators and DM annotators worked on entirely different sets of documents. This paper tackles the challenge of how to perform joint learning across these structurally divergent formalisms using non-overlapping data.
Methodology: Joint Learning via Latent Variables
The authors propose a joint model that scores both a frame-semantic parse and a semantic dependency graph simultaneously.
Latent Variable Formulation
When training on a FrameNet sentence (which lacks DM labels), the model treats the missing DM graph as a latent structured variable. It essentially "guesses" the most likely dependency graph to help inform the frame-semantic prediction.
Cross-Task Scoring and Neural Architecture
The core innovation is a scoring function that explicitly relates spans to dependencies. Inspired by linguistic "head rules," the model scores the affinity between a FrameNet argument span and the dependency arcs connected to the "head" word of that span. The architecture utilizes bidirectional LSTMs (biLSTMs) to create contextualized word representations, which are then fed into multi-layer perceptrons (MLPs) to score predicates, arguments, and arcs.
Sparsity for Computational Efficiency
Jointly decoding two different structures is computationally expensive (quadratic complexity). To address this, the authors apply an $\ell_1$ penalty to the cross-task scores. This encourages the model to ignore unlikely connections, resulting in a 14x speedup during inference without sacrificing accuracy.
Datasets and Evaluation
The model was evaluated on standard benchmarks:
- FrameNet 1.5 & 1.7: Standard benchmarks for span-based semantic role labeling.
- SemEval 2015 Task 18 (DM): A dependency-based formalism derived from DeepBank.
- Exemplar Sentences: Additional single-target annotations from the FrameNet project.
Crucially, the authors ensured there was no overlap between the FrameNet training data and the DM test sets to demonstrate the model's ability to generalize from disjoint sources.
Key Results and Insights
The joint model (FULL) demonstrated superior performance across multiple metrics.
- State-of-the-Art on FrameNet: The joint model outperformed all previous single-model and ensemble systems. On FN 1.5, it achieved an F1 score of 77.4, an 0.8% absolute improvement over the previous best.
- Improved Generalization: The joint model showed significant gains on out-of-domain data (the Brown corpus), suggesting that learning from multiple formalisms helps the model capture "deeper" linguistic patterns.
- Synergy in Divergence: Even though the data structures differ, they describe the same underlying reality. A dependency arc often points to the head of a semantic span. By modeling this connection, the two tasks provide "mutual supervision."
Conclusion and Future Directions
This work demonstrates that you don't need gold labels for every task to learn effectively. By forcing the model to find a dependency structure that "fits" a FrameNet annotation, it learns more robust features. While the current approach uses simplified head-finding rules and assumes pre-identified targets, future work could extend this latent variable approach to even more distinct representations like Abstract Meaning Representation (AMR) or Universal Dependencies (UD).
Summary updated