Dynamic Relation Extraction with A Learnable Temporal Encoding Method

The need to judge the relations between two entities at a specific time arises in many natural language understanding and knowledge graph related tasks, where the traditional relation extraction (RE) task without considering specific time is not feasible. Therefore, it is an important task to extract the dynamic relation from sentences containing the two entities. However, existing studies focus on extracting the static relation while ignoring temporal information in sentences or encode temporal information as a sequence to infer the relation. Considering these limitations of existing studies, we propose a Learnable Temporal Encoding (LTE) model, which encodes explicit temporal information in sentences. Specifically, we introduce a key-value memory network in LTE to identify the relation between an entity pair at a specific time. Through experiments on a general temporal relation extraction dataset, we show that the proposed model outperforms other state-of-the-art baselines, which demonstrate the effectiveness of LTE for dynamic relation extraction. We also conduct visual analysis to demonstrate that our model can fully represent the temporal information in the embedding space for any time spots.