Introduction
In Natural Language Processing, extracting "who did what to whom" (predicate-argument tuples) is a fundamental task. While standard tools excel at identifying explicit arguments (e.g., "John ate the apple"), they often fail to capture implicit arguments. These are participants mentioned elsewhere in a text that are not grammatically linked to the specific action. For instance, in the sentence "a decline then set in," a human reader understands what declined based on the preceding context, but a machine perceives a missing data point.
This research re-frames the challenge of predicting these missing participants as a reading comprehension (RC) task. By treating the missing argument as a query and the surrounding document as context, the authors propose a neural model that "reads" the text to identify the most likely filler for these roles.
Methodology: The Pointer Attentive Reader
The core innovation of this work is the Pointer Attentive Reader (PAR). The model operates by taking a predicate-argument tuple with a missing slot (e.g., reduce-pred, TARGET-subj, cost-dobj) and treating it as a question. The preceding sentences in the document serve as the search space.
The PAR model combines two key components:
- Attentive Reader: Uses an attention mechanism to weigh different parts of the document based on their relevance to the query.
- Pointer Network: Instead of selecting a word from a predefined vocabulary, the model "points" directly to a specific word or mention already present in the document context.
For complex scenarios involving multiple missing arguments, the authors implement a Multi-hop Attention approach. In the first "hop," the model scans the document to gather information about all missing pieces, building a more informed query. In the second hop, it makes the final prediction for the target argument.
Experimental Results
The model was evaluated on several datasets, including modified versions of OntoNotes (ON-SHORT and ON-LONG) and the manually annotated G&C (Gerber and Chai) dataset. Large-scale training was also conducted using a 25-million-pair dataset derived from English Wikipedia.
The PAR model demonstrated significant performance gains:
- OntoNotes Superiority: On the ON-LONG dataset, PAR achieved 51.52% accuracy, more than doubling the performance of the previous state-of-the-art "EVENTCOMP" model (21.26%).
- Contextual Efficiency: The model's BiGRU architecture proved highly effective at maintaining long-distance dependencies, leading to better performance on longer documents compared to baselines.
- Robustness: Unlike previous models that relied on high entity frequency (salience) to make accurate guesses, PAR showed steady performance across both rare and frequent entities. It learns "implicit salience" naturally through sequential document encoding.
Conclusion and Future Work
The transition to a reading comprehension framework allows for more sophisticated reasoning in implicit argument prediction. While the PAR model sets a new benchmark, limitations remain. Currently, the model only considers preceding context, whereas fillers may occasionally appear later in a narrative. Future work may involve integrating more advanced RC architectures like BiDAF and exploring latent coreference modeling to further bridge the gap between verbal and nominal predicates.
Summary updated