Variational Deep Logic Network for Joint Inference of Entities and Relations

Abstract Currently, deep learning models have been widely adopted and achieved promising results on various application domains. Despite their intriguing performance, most function as black boxes, lacking explicit reasoning capabilities explanations, which are usually essential for complex problems. Take joint inference in information extraction an example. This task requires the identification of multiple structured knowledge from texts, is inter-correlated, including entities, events, relationships between them. Various neural networks proposed to jointly perform entity relation prediction, only propagate implicitly via representation learning. However, they fail encode intensive correlations types relations enforce coexistence. On other hand, some approaches adopt rules explicitly constrain certain relational facts, although separation with restrains error propagation. Moreover, predefined inflexible might result negative effects when data noisy. To address these limitations, we propose a variational logic network that incorporates both EM algorithm. The model consists learn high-level features implicit interactions self-attention mechanism exploit target interactions. These two components trained interactively bring best worlds. We conduct extensive experiments ranging fine-grained sentiment terms extraction, end-to-end event demonstrate effectiveness our method.