Version 0.6.2

Added

• Docker image covering both Deep Learning and CRF models, with GPU detection and preloading of embeddings
• For Deep Learning models, labeling is now done by batch: application of the citation DL model is 4 times faster for BidLSTM-CRF (with or without features) and 6 times faster for SciBERT
• More tests for sentence segmentation
• Add orcid of persons when available from the PDF or via consolidation (i.e. if in CrossRef metadata)
• Add BidLSTM-CRF-FEATURES header model (with feature channel)
• Add bioRxiv end-to-end evaluation
• Bounding boxes for optional section titles coordinates

Changed

• Reduce the size of docker images
• Improve end-to-end evaluation: multithreaded processing of PDF, progress bar, output the evaluation report in markdown format
• Update of several models covering CRF, BidLSTM-CRF and BidLSTM-CRF-FEATURES, mainly improving citation and author recognitions
• OpenNLP is the default optional sentence segmenter (similar result as Pragmatic Segmenter for scholar documents after benchmarking, but 30 times faster)
• Refine sentence segmentation to exploit layout information and predicted reference callouts
• Update jep version to 3.9.1

Fixed

• Ignore invalid utf-8 sequences
• Update CrossRef multithreaded calls to avoid using the unreliable time interval returned by the CrossRef REST API service, update usage of Crossref-Plus-API-Token and update the deprecated crossref field query.title
• Missing last table or figure when generating training data for the fulltext model
• Fix an error related to the feature value for the reference callout for the fulltext model
• Review/correct DeLFT configuration documentation, with a step-by-step configuration documentation
• Other minor fixes