We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:


References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: Heterogeneous Graph Attention Network for Multi-hop Machine Reading Comprehension

Abstract: Multi-hop machine reading comprehension is a challenging task in natural language processing, which requires more reasoning ability and explainability. Spectral models based on graph convolutional networks grant the inferring abilities and lead to competitive results, however, part of them still face the challenge of analyzing the reasoning in a human-understandable way. Inspired by the concept of the Grandmother Cells in cognitive neuroscience, a spatial graph attention framework named crname, imitating the procedure was proposed. This model is designed to assemble the semantic features in multi-angle representations and automatically concentrate or alleviate the information for reasoning. The name "crname" is a metaphor for the pattern of the model: regard the subjects of queries as the start points of clues, take the reasoning entities as bridge points, and consider the latent candidate entities as the grandmother cells, and the clues end up in candidate entities. The proposed model allows us to visualize the reasoning graph and analyze the importance of edges connecting two entities and the selectivity in the mention and candidate nodes, which can be easier to be comprehended empirically. The official evaluations in open-domain multi-hop reading dataset WikiHop and Drug-drug Interactions dataset MedHop prove the validity of our approach and show the probability of the application of the model in the molecular biology domain.
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:2107.00841 [cs.CL]
  (or arXiv:2107.00841v1 [cs.CL] for this version)

Submission history

From: Peng Gao [view email]
[v1] Fri, 2 Jul 2021 05:29:39 GMT (28629kb,D)

Link back to: arXiv, form interface, contact.