Abstract
Determining whether an event actually occurred based on a full news article is a complex task prone to noise. This paper introduces RMHAN, an end-to-end model that employs hierarchical reinforcement learning to filter out irrelevant text and focus on key evidence at both the word and sentence levels. By "pruning" documents to isolate factual cues, RMHAN achieves state-of-the-art results on bilingual datasets, outperforming standard BERT-based models.
Introduction
In Natural Language Processing, Event Factuality Identification (EFI) determines the factual status of an event—whether it definitely happened (Certain Positive), is a possibility (Possible Positive), or was explicitly denied (Certain Negative). While most research focuses on single sentences (SEFI), real-world information is often distributed across an entire document (DEFI). Document-level analysis is challenging because articles may contain contradictory statements, speculative quotes, or irrelevant background noise. Previous DEFI models often relied on manual "gold" annotations for triggers or struggled with distinguishing relevant sentences from "fluff."
Methodology: The RMHAN Architecture
The proposed Reinforced Multi-Granularity Hierarchical Attention Network (RMHAN) addresses these challenges through three primary components:
- Hierarchical Encoding: The model processes text at two levels, encoding tokens into sentence representations and then sentences into a document-level representation.
- Multi-Granularity Sub-Encoders: Specialized encoders capture different nuances: IS-SubEnc for internal sentence semantics, T-SubEnc for lead-sentence context, and E-SubEnc to integrate specific event descriptions.
- Reinforced Text Selection: Using Hierarchical Reinforcement Learning (HRL), the model employs two policy networks: TS-PNet for token selection and SS-PNet for sentence selection.
Unlike standard "soft" attention mechanisms that assign weights to all words, RMHAN performs a "hard" selection, learning to discard noisy text that does not contribute to factual prediction.
Experimental Evaluation
The model was evaluated on the DLEF-v2 corpus, a benchmark for document-level event factuality containing English (4,530 events) and Chinese (4,650 events) sub-corpora. Events are labeled across five categories: Certain Positive (CT+), Certain Negative (CT-), Possible Positive (PS+), Possible Negative (PS-), and Underspecified (Uu). Evaluation primarily focused on CT-, PS+, and CT+, which represent the majority of the data.
Results and Discussion
RMHAN consistently outperformed all baselines, including Graph Convolutional Networks (SGCN), BERT-Base, and previous reinforcement learning models like RLSTM.
Key findings include:
- Hard Selection Efficacy: The reinforcement learning policy effectively cleans data before final classification, proving superior to standard attention.
- Structural Advantage: A hierarchical approach significantly outperforms flat input models (like the RMLAN variant), mirroring human processing of document structure.
- Ablation Insights: Removing the Sentence Selection (SS-PNet) module led to a significant drop in F1 scores (approx. 3-4 points), confirming that filtering irrelevant sentences is critical for DEFI.
Conclusion
RMHAN demonstrates that an end-to-end reinforced hierarchical model can achieve state-of-the-art performance in document-level event factuality identification without requiring manual cue annotations. Future work will explore cross-document EFI and improved temporal logic to better handle complex timelines and conflicting reports across multiple sources.
Summary updated