CAIRD: Context-Aware Implicit Relation Discovery in Multi-Event Chains

The ability to understand implicit relationships between events plays an important role in higher-level natural language processing, though the current methods for event, or relationship extraction, struggle with multi-event chains where the logical relationship is unstated and spans multiple sentences or paragraphs. Current approaches generally rely on explicit indicators or preset relation types to identify event relationships and do not account for reasoning and common sense knowledge. In response to this gap we introduce Context-Aware Implicit Relation Discovery (CAIRD) - a framework for detecting and extracting unstated relationships that carry semantic importance across event sequences such as causal, temporal and conditional relationships. CAIRD comprises an Event Chain Context Encoder for sequential understanding, a Common Sense Knowledge Augmenter to incorporate outside knowledge, and an Implicit Relation Detector that learns a representation of relations within a continuous space, along with a Relation Fusion and Output Module. We also introduce the Implicit Narrative Graph dataset for annotating implicit relations, and the Eventual Common Sense knowledge graph for outside augmentation. Experiments show that CAIRD performs better than strong baselines for text relations and knowledge-enhanced approaches suggesting the utility of external knowledge, and the overall utility of the proposed framework architecture to capture both rich implicit logical properties of more complex, longer event sequences.