Abstract
Semantic Dependency Parsing (SDP) aims to capture complex predicate-argument relationships where words can have multiple heads or none at all. While biaffine parsers offer high speed and parallelizability, their independent arc-prediction mechanism often lacks global structural consistency. This paper proposes enhancing these parsers using simple auxiliary tasks—such as predicting head counts and label sets—during training. Evaluated on English and French benchmarks, these "side quests" force the model to learn richer internal representations, leading to systematic performance gains while maintaining the efficiency of the original $O(n^2)$ architecture.
The Challenge of Semantic Dependency Parsing
In standard Dependency Parsing (DP), every word typically has exactly one head, forming a tree structure. Semantic Dependency Parsing (SDP) is more flexible and complex: a single word can be the argument of multiple predicates or none.
Modern state-of-the-art biaffine parsers treat each potential edge between words as an independent classification decision. While this allows for $O(n^2)$ complexity and fast processing, it can lead to linguistically inconsistent graphs because the model lacks a global view of the word's role within the sentence.
Methods: Global Awareness via Auxiliary Tasks
To address this "tunnel vision," the author introduces four auxiliary tasks that act as global constraints during training, forcing the model's shared representations to capture interdependencies between arcs:
- Task H (Heads): Predicts the total number of incoming arcs (governors) for a word.
- Task D (Dependents): Predicts the total number of outgoing arcs (dependents).
- Task S (Label Set): Predicts the specific combination of incoming labels for a word.
- Task B (Bag of Labels): Predicts a vector representing the counts of each possible label type.
These tasks are integrated using Task Uncertainty weighting to automatically balance loss contributions and Stack Propagation to feed auxiliary hidden layers into the main parser.
Experimental Setup
The method was tested on two diverse benchmarks:
- English SemEval 2015 Task 18: Covering three formalisms: DM (DELPH-IN MRS), PAS (Predicate-Argument Structure), and PSD (Prague Semantic Dependencies).
- French Sequoia Treebank: Featuring deep syntactic graphs with high levels of cyclicity.
Key Results
The inclusion of auxiliary tasks led to systematic improvements across all datasets. The combination of Task B (Bag of Labels) and Task H (Number of Heads) proved particularly effective.
On the French dataset, the B+H configuration achieved a statistically significant increase of +0.53 to +0.56 in Labeled F-score (LF). Furthermore, graph consistency improved; for English, the percentage of words assigned the correct number of heads rose from 94.9% to 95.5%. Crucially, these gains come with minimal computational overhead, as the auxiliary tasks primarily serve to "shape" the model during the training phase.
Conclusion
By forcing biaffine parsers to consider local counting properties, simple auxiliary tasks can effectively mitigate the independence assumptions that lead to structural errors. While the gains are modest compared to computationally expensive higher-order parsers, they represent a robust and efficient way to improve the state-of-the-art in graph-based semantic parsing. Future research may explore applying these techniques to other graph-based tasks like AMR parsing.
Summary updated