🏷️ v0.3.1

Data

Major improvement of intext.patent

Validation

Validation of the intext.patent table

Thanks

Special thanks to:

Gabriele Cristelli (EPFL)
Kyle Higham (Hitotsubashi University)
Lucas Violon (HEC Paris)

0.3.0

🏷 v0.3.0

Data

• Major improvement of bibliographical_reference schema (harmonize grobid & crossref) for seamless analysis
• Enrichment of intext.patent
• Add domain specific front page tables (norm_standard, database, wiki)

Community

• Revisit BQ project architecture
• Add Colab notebooks integration
• Revisit README.md

Code

• Lighter API
• Lighter dependencies

Models

• Add information extraction models
• Add models and training data DVC support

Validation

• Validation of in-text extraction models

Thanks

Special thanks to:

• Gabriele Cristelli (EPFL)
• Kyle Higham (Hitotsubashi University)
• Lucas Violon (HEC Paris)

v0.2-npl

🏷 v0.2-npl

The v0.2-npl introduces 2 major improvements:

• npl_class field. This field is predicted using a multi-class text classification model based on spaCy textCategorizer with the npl text as input. See focus and models binaries below.
• Propagate ISSN using title_j to bibliographical references with the same title_j but no match.

Focus on npl_class

en_core_web_sm_npl-class-ensemble-0.8

• ensemble model (bow+cnn with bagging)
• trained on 80% of the "gold" dataset and evaluated on remaining 20% (hold-out)

See in models/npl_class_training/ for more

Average performance

accuracy	precision	recall	f1
0.9	0.89	0.88	0.88

Class performance

precision	recall	f1	support
BIBLIOGRAPHICAL_REFERENCE	0.92	0.95	0.93
SEARCH_REPORT	1.0	0.92	0.96
OFFICE_ACTION	0.99	0.93	0.96
DATABASE	0.89	0.73	0.8
WEBPAGE	0.53	0.53	0.53
PATENT	0.91	0.94	0.93
NA	1.0	1.0	1.0
PRODUCT_DOCUMENTATION	0.44	0.43	0.44
NORM_STANDARD	0.86	0.6	0.71
LITIGATION	0.25	0.11	0.15

en_core_web_sm_npl-class-ensemble-1.0

Same as en_core_web_sm_npl-class-ensemble-1.0 but trained on full dataset to maximize performance. Model used to create the npl_class field.