Neighborhood hash graph kernel for protein–protein interaction extraction

Automated extraction of protein–protein interactions (PPIs) from biomedical literatures is an important topic of biomedical text mining. In this paper, we propose an approach based on neighborhood hash graph kernel for this task. In contrast to the existing graph kernel-based approaches for PPI extraction, the proposed approach not only has the capability to make use of full dependency graphs to represent the sentence structure but also effectively control the computational complexity. We evaluate the proposed approach on five publicly available PPI corpora and perform detailed comparisons with other approaches. The experimental result shows that our approach is comparable to the state-of-the-art PPI extraction system and much faster than all-path graph kernel approach on all five PPI corpora.

Figure optionsDownload as PowerPoint slideHighlights
► We propose a neighbor hash kernel method for protein–protein interaction extraction.
► Neighbor hash kernel can efficiently represent sentence dependency graphs.
► Neighbor hash kernel method achieves state-of-the-art performance on five corpora.
► Neighbor hash kernel method is much faster than all-path kernel method.