Abstract
Joint models of syntactic and semantic parsing have the potential to improve performance on both tasks—but to date, the best results have been achieved with pipelines. We introduce a joint model using CCG, which is motivated by the close link between CCG syntax and semantics. Semantic roles are recovered by labelling the deep dependency structures produced by the grammar. Furthermore, because CCG is lexicalized, we show it is possible to factor the parsing model over words and introduce a new A* parsing algorithm—which we demonstrate is faster and more accurate than adaptive supertagging. Our joint model is the first to substantially improve both syntactic and semantic accuracy over a comparable pipeline, and also achieves state-of-the-art results for a non-ensemble semantic role labelling model.
The Challenge of Pipeline Architectures
In traditional natural language processing, systems are typically constructed as pipelines. First, a syntactic parser determines the grammatical structure of a sentence; subsequently, a Semantic Role Labeling (SRL) system assigns semantic labels (such as Agent or Patient) to these structures.
The primary drawback of this approach is error propagation. If the initial parser makes an incorrect attachment decision in a complex sentence, the downstream SRL system lacks the mechanism to recover from that error. While a joint model—where syntactic and semantic information are processed simultaneously—should theoretically resolve these ambiguities, they have historically been slower and less accurate than well-tuned pipelines due to the computationally explosive search space of combined structures.
Leveraging CCG for Joint Modeling
This research utilizes Combinatory Categorial Grammar (CCG) as its foundation. CCG is a highly lexicalized formalism where words carry complex categories that define their combinatory potential. Because CCG categories inherently define "slots" for arguments, the joint task can be framed as labeling these syntactic slots with semantic roles (e.g., ARG0, ARG1) or a null label ($\emptyset$).
As shown in Figure 1, CCG is uniquely suited for SRL because it naturally handles long-range dependencies. In the example "He refused to confirm or deny reports," the model correctly identifies "He" as the ARG0 (subject) for "refused," "confirm," and "deny" simultaneously. This deep dependency structure allows semantic information to "talk back" to the parser, helping to resolve structural ambiguities that are difficult to solve with syntax alone.
Lexical Factorization and A* Search
A core innovation of this work is the factorization of the model into extended lexical entries. The researchers treat a parse as a sequence of decisions for each word: selecting its CCG category, determining the semantic roles for its arguments, and deciding where those arguments attach.
To make this joint model efficient, the authors introduce a novel A* parsing algorithm. Standard parsers often rely on the CKY algorithm, which exhaustively fills tables of all possible spans. By using the lexical factorization shown in Figure 4, the authors compute a heuristic—an upper bound on the score of the remaining parse. Because the best possible roles and attachments can be pre-calculated for individual words, the A* search prunes the search space aggressively without sacrificing the optimal solution.
Experimental Results
The model was evaluated on several benchmarks, including PropBank for SRL and CCGbank for syntactic accuracy.
- Superior Accuracy: The joint model achieved an F1 score of 83.5 on PropBank, significantly outperforming the pipeline baseline of 76.4 and exceeding previous single-parser state-of-the-art models.
- Syntactic Improvement: Syntactic parsing accuracy improved by 1.6 F1 points over the pipeline, confirming that semantic constraints actually assist the syntactic parser.
- Efficiency: The A algorithm proved to be five times faster* than CKY parsing on the same model, making joint parsing practically viable for large-scale applications.
- Generalization: The model maintained strong performance on the out-of-domain Brown corpus, demonstrating robustness across different writing styles.
Conclusion and Future Work
This work demonstrates that joint syntactic and semantic parsing is not only possible but superior to traditional pipelines when using a well-suited formalism like CCG and an efficient search algorithm like A*. While attachment errors still account for roughly a third of the model's mistakes, the success of this approach suggests a path forward for even deeper semantic parsers that produce full logical forms, bridging the gap between broad-coverage parsing and formal semantics.
Summary updated