Modeling temporal self and interactive evolution for biomedical hypothesis generation

Hypothesis generation (HG) aims to reveal meaningful hidden relationships between scientific terms from literature for accelerating innovation in drug discovery, disease prognosis and treatment. Recent studies have successfully employed the dynamic nature of term-pair relations for HG. However, the existing methods focus on capturing the evolution of term pairs by modeling the temporal meaning of terms themselves, which is hard to accurately model intricate spatio-temporal relations between term pairs.

In this paper, a Temporal Self and Interactive Evolution (TSIE) modeling method is proposed to accurately characterize complex dynamics of term-pair relations in HG. Specifically, for each term pair, we first employ Gated Recurrent Unit (GRU) to model its Temporal Self-evolution (TSE) and Temporal Interactive Evolution (TIE) for learning its TSE Embedding (TSE_emb) and TIE Embedding (TIE_emb), respectively. Then, we adopt a dual-tower Transformer to further model the temporal dependencies of both TSE_emb and TIE_emb, which are finally integrated by a gated fusion layer for inferring the future connectivity of the term pair.

Experiments on three real-world datasets Immunotherapy, Virology, and Neurology demonstrate that TSIE can effectively capture complex evolutional patterns for biomedical hypothesis generation and achieve the state-of-the-art performance.

This paper proposes a novel TSIE method to learn temporal interactive difference features and enhance the model's understanding of temporal relation inference. Our TSIE learns both TSE and TIE to effectively model the dynamic relationship between terms. By adapting a dual-tower Transformer encoder, TSIE can further model the temporal dependencies of TSE and TIE.