Search for: All records

The Transparent Research Object Vocabulary (TROV) is a key element of the Transparency Certified (TRACE) approach to ensuring research trustworthiness. In contrast with methods that entail repeating computations in part or in full to verify that the descriptions of methods included in a publication are sufficient to reproduce reported results, the TRACE approach depends on a controlled computing environment termed a Transparent Research System (TRS) to guarantee that accurate, sufficiently complete, and otherwise trustworthy records are captured when results are obtained in the first place. Records identifying (1) the digital artifacts and computations that yielded a research result, (2) the TRS that witnessed the artifacts and supervised the computations, and (3) the specific conditions enforced by the TRS that warrant trust in these records, together constitute a Transparent Research Object (TRO). Digital signatures provided by the TRS and by a trusted third-party timestamp authority (TSA) guarantee the integrity and authenticity of the TRO. The controlled vocabulary TROV provides means to declare and query the properties of a TRO, to enumerate the dimensions of trustworthiness the TRS asserts for a TRO, and to verify that each such assertion is warranted by the documented capabilities of the TRS. Our approach for describing, publishing, and working with TROs imposes no restrictions on how computational artifacts are packaged or otherwise shared, and aims to be interoperable with, rather than to replace, current and future Research Object standards, archival formats, and repository layouts. 

The quadratic computational and memory complexities of large Transformers have limited their scalability for long document summarization. In this paper, we propose HEPOS, a novel efficient encoder-decoder attention with head-wise positional strides to effectively pinpoint salient information from the source. We further conduct a systematic study of existing efficient self-attentions. Combined with HEPOS, we are able to process ten times more tokens than existing models that use full attentions. For evaluation, we present a new dataset, GOVREPORT, with significantly longer documents and summaries. Results show that our models produce significantly higher ROUGE scores than competitive comparisons, including new state-of-the-art results on PubMed. Human evaluation also shows that our models generate more informative summaries with fewer unfaithful errors.