Benchmarking biorXiv

General

This is the end-to-end benchmarking result for GROBID version 0.7.0 against the bioRxiv test set (biorxiv-10k-test-2000), see the End-to-end evaluation page for explanations and for reproducing this evaluation.

The following end-to-end results are using: - BidLSTM-CRF-FEATURES as sequence labeling for the citation model - CRF Wapiti as sequence labelling engine for all other models.

Header extractions are consolidated by default with biblio-glutton service (the results with CrossRef REST API as consolidation service should be similar but much slower).

Other versions of these benchmarks with variants and Deep Learning models (e.g. newer master snapshots) are available here. Note that Deep Learning models might provide higher accuracy, but at the cost of slower runtime and more expensive CPU/GPU resources.

Evaluation on 1999 PDF preprints out of 2000 (1 PDF "too many blocks" interruption).

Runtime for processing 2000 PDF: 1169s (1,71 PDF per second) on Ubuntu 16.04, 4 CPU i7-4790K (8 threads), 16GB RAM (workstation bought in 2015 for 1600 euros) and with a GeForce GTX 1050 Ti GPU.

Header metadata

Evaluation on 1999 random PDF files out of 2000 PDF (ratio 1.0).

Strict Matching (exact matches)

Field-level results

label precision recall f1 support
abstract 2.24 2.16 2.2 1989
authors 84.06 83.13 83.59 1998
first_author 94.69 93.74 94.21 1996
keywords 59.91 60.91 60.4 839
title 86.84 84.14 85.47 1999
all fields (micro avg.) 66.58 65.39 65.98 8821
all fields (macro avg.) 65.54 64.82 65.17 8821

Soft Matching (ignoring punctuation, case and space characters mismatches)

Field-level results

label precision recall f1 support
abstract 57.42 55.46 56.42 1989
authors 84.62 83.68 84.15 1998
first_author 94.79 93.84 94.31 1996
keywords 65.77 66.87 66.31 839
title 92.36 89.49 90.9 1999
all fields (micro avg.) 80.78 79.33 80.05 8821
all fields (macro avg.) 78.99 77.87 78.42 8821

Levenshtein Matching (Minimum Levenshtein distance at 0.8)

Field-level results

label precision recall f1 support
abstract 76.83 74.21 75.5 1989
authors 92.46 91.44 91.95 1998
first_author 95.19 94.24 94.71 1996
keywords 78.31 79.62 78.96 839
title 95.25 92.3 93.75 1999
all fields (micro avg.) 88.85 87.26 88.05 8821
all fields (macro avg.) 87.61 86.36 86.97 8821

Ratcliff/Obershelp Matching (Minimum Ratcliff/Obershelp similarity at 0.95)

Field-level results

label precision recall f1 support
abstract 73.87 71.34 72.58 1989
authors 88.16 87.19 87.67 1998
first_author 94.69 93.74 94.21 1996
keywords 71.28 72.47 71.87 839
title 93.96 91.05 92.48 1999
all fields (micro avg.) 86.11 84.57 85.34 8821
all fields (macro avg.) 84.39 83.16 83.76 8821

Instance-level results

Total expected instances:   1999
Total correct instances:    34 (strict) 
Total correct instances:    753 (soft) 
Total correct instances:    1158 (Levenshtein) 
Total correct instances:    1026 (ObservedRatcliffObershelp) 

Instance-level recall:  1.7 (strict) 
Instance-level recall:  37.67   (soft) 
Instance-level recall:  57.93   (Levenshtein) 
Instance-level recall:  51.33   (RatcliffObershelp) 

Citation metadata

Evaluation on 1999 random PDF files out of 2000 PDF (ratio 1.0).

Strict Matching (exact matches)

Field-level results

label precision recall f1 support
authors 86.67 78.42 82.34 97138
date 91.41 83.27 87.15 97585
doi 72.61 80.38 76.3 16893
first_author 93.54 84.57 88.83 97138
inTitle 81.53 77.13 79.27 96384
issue 93.61 85.63 89.44 30282
page 96.52 78.67 86.69 88558
pmcid 63.26 63.57 63.41 807
pmid 66.96 75.63 71.03 2093
title 84.29 80.7 82.45 92423
volume 95.57 92.79 94.16 87671
all fields (micro avg.) 89.25 82.2 85.58 706972
all fields (macro avg.) 84.18 80.07 81.92 706972

Soft Matching (ignoring punctuation, case and space characters mismatches)

Field-level results

label precision recall f1 support
authors 87.94 79.56 83.54 97138
date 91.41 83.27 87.15 97585
doi 77.11 85.35 81.02 16893
first_author 94 84.99 89.27 97138
inTitle 91.18 86.26 88.65 96384
issue 93.61 85.63 89.44 30282
page 96.52 78.67 86.69 88558
pmcid 73.74 74.1 73.92 807
pmid 71.36 80.6 75.7 2093
title 92.48 88.55 90.47 92423
volume 95.57 92.79 94.16 87671
all fields (micro avg.) 92.1 84.83 88.32 706972
all fields (macro avg.) 87.72 83.62 85.46 706972

Levenshtein Matching (Minimum Levenshtein distance at 0.8)

Field-level results

label precision recall f1 support
authors 92.69 83.86 88.06 97138
date 91.41 83.27 87.15 97585
doi 79.82 88.36 83.87 16893
first_author 94.15 85.12 89.41 97138
inTitle 92.13 87.16 89.58 96384
issue 93.61 85.63 89.44 30282
page 96.52 78.67 86.69 88558
pmcid 73.74 74.1 73.92 807
pmid 71.4 80.65 75.75 2093
title 95.33 91.27 93.26 92423
volume 95.57 92.79 94.16 87671
all fields (micro avg.) 93.36 85.99 89.53 706972
all fields (macro avg.) 88.76 84.63 86.48 706972

Ratcliff/Obershelp Matching (Minimum Ratcliff/Obershelp similarity at 0.95)

Field-level results

label precision recall f1 support
authors 89.79 81.23 85.3 97138
date 91.41 83.27 87.15 97585
doi 79.05 87.5 83.06 16893
first_author 93.59 84.61 88.87 97138
inTitle 89.87 85.03 87.38 96384
issue 93.61 85.63 89.44 30282
page 96.52 78.67 86.69 88558
pmcid 63.26 63.57 63.41 807
pmid 66.96 75.63 71.03 2093
title 94.52 90.5 92.47 92423
volume 95.57 92.79 94.16 87671
all fields (micro avg.) 92.42 85.12 88.62 706972
all fields (macro avg.) 86.74 82.59 84.45 706972

Instance-level results

Total expected instances:       98753
Total extracted instances:      103498
Total correct instances:        41273 (strict) 
Total correct instances:        51887 (soft) 
Total correct instances:        56012 (Levenshtein) 
Total correct instances:        52881 (RatcliffObershelp) 

Instance-level precision:   39.88 (strict) 
Instance-level precision:   50.13 (soft) 
Instance-level precision:   54.12 (Levenshtein) 
Instance-level precision:   51.09 (RatcliffObershelp) 

Instance-level recall:  41.79   (strict) 
Instance-level recall:  52.54   (soft) 
Instance-level recall:  56.72   (Levenshtein) 
Instance-level recall:  53.55   (RatcliffObershelp) 

Instance-level f-score: 40.81 (strict) 
Instance-level f-score: 51.31 (soft) 
Instance-level f-score: 55.39 (Levenshtein) 
Instance-level f-score: 52.29 (RatcliffObershelp) 

Matching 1 :    75680

Matching 2 :    4230

Matching 3 :    6093

Matching 4 :    2284

Total matches : 88287

Citation context resolution


Total expected references:   98712 - 49.38 references per article
Total predicted references:      103442 - 51.75 references per article

Total expected citation contexts:    142737 - 71.4 citation contexts per article
Total predicted citation contexts:   134945 - 67.51 citation contexts per article

Total correct predicted citation contexts:   111261 - 55.66 citation contexts per article
Total wrong predicted citation contexts:     23684 (wrong callout matching, callout missing in NLM, or matching with a bib. ref. not aligned with a bib.ref. in NLM)

Precision citation contexts:     82.45
Recall citation contexts:    77.95
fscore citation contexts:    80.14

Fulltext structures

Fulltext structure contents are complicated to capture from JATS NLM files. They are often normalized and different from the actual PDF content and are can be inconsistent from one document to another. The scores of the following metrics are thus not very meaningful in absolute term, in particular for the strict matching (textual content of the srtructure can be very long). As relative values for comparing different models, they seem however useful.

Evaluation on 1999 random PDF files out of 2000 PDF (ratio 1.0).

Strict Matching (exact matches)

Field-level results

label precision recall f1 support
figure_title 4.13 3.57 3.83 13162
reference_citation 70.63 70.48 70.55 147404
reference_figure 73.72 65.91 69.6 47965
reference_table 48.12 80.66 60.28 5951
section_title 71.28 71.04 71.16 32384
table_title 4.54 4.09 4.3 2957
all fields (micro avg.) 66.55 65.6 66.07 249823
all fields (macro avg.) 45.4 49.29 46.62 249823

Soft Matching (ignoring punctuation, case and space characters mismatches)

Field-level results

label precision recall f1 support
figure_title 67.36 58.3 62.51 13162
reference_citation 82.28 82.1 82.19 147404
reference_figure 74.42 66.53 70.25 47965
reference_table 48.55 81.38 60.81 5951
section_title 75.06 74.8 74.93 32384
table_title 50.73 45.72 48.1 2957
all fields (micro avg.) 77.57 76.47 77.01 249823
all fields (macro avg.) 66.4 68.14 66.47 249823

Evaluation metrics produced in 1132.55 seconds

Header metadata

Evaluation on 1998 random PDF files out of 1998 PDF (ratio 1.0).

Strict Matching (exact matches)

Field-level results

label precision recall f1 support
abstract 2.19 2.06 2.12 1988
authors 73.78 70.31 72 1997
first_author 93.12 88.82 90.92 1995
keywords 55.41 54.95 55.18 839
title 78.64 71.32 74.8 1998
all fields (micro avg.) 61.3 57.88 59.54 8817
all fields (macro avg.) 60.63 57.49 59 8817

Soft Matching (ignoring punctuation, case and space characters mismatches)

Field-level results

label precision recall f1 support
abstract 55.12 51.96 53.5 1988
authors 75.56 72.01 73.74 1997
first_author 93.69 89.37 91.48 1995
keywords 60.22 59.71 59.96 839
title 90.01 81.63 85.62 1998
all fields (micro avg.) 76.72 72.43 74.51 8817
all fields (macro avg.) 74.92 70.94 72.86 8817

Levenshtein Matching (Minimum Levenshtein distance at 0.8)

Field-level results

label precision recall f1 support
abstract 73.91 69.67 71.72 1988
authors 90.54 86.28 88.36 1997
first_author 94.48 90.13 92.25 1995
keywords 72.36 71.75 72.05 839
title 93.65 84.93 89.08 1998
all fields (micro avg.) 86.56 81.72 84.07 8817
all fields (macro avg.) 84.99 80.55 82.69 8817

Ratcliff/Obershelp Matching (Minimum Ratcliff/Obershelp similarity at 0.95)

Field-level results

label precision recall f1 support
abstract 70.7 66.65 68.62 1988
authors 83.34 79.42 81.33 1997
first_author 93.17 88.87 90.97 1995
keywords 65.87 65.32 65.59 839
title 92.05 83.48 87.56 1998
all fields (micro avg.) 82.89 78.26 80.51 8817
all fields (macro avg.) 81.03 76.75 78.81 8817

Instance-level results

Total expected instances:   1998
Total correct instances:    30 (strict) 
Total correct instances:    596 (soft) 
Total correct instances:    1000 (Levenshtein) 
Total correct instances:    862 (ObservedRatcliffObershelp) 

Instance-level recall:  1.5 (strict) 
Instance-level recall:  29.83   (soft) 
Instance-level recall:  50.05   (Levenshtein) 
Instance-level recall:  43.14   (RatcliffObershelp) 

Citation metadata

Evaluation on 1998 random PDF files out of 1998 PDF (ratio 1.0).

Strict Matching (exact matches)

Field-level results

label precision recall f1 support
authors 86.86 70.58 77.88 97092
date 90.05 74.78 81.71 97539
doi 72.86 72.76 72.81 16893
first_author 93.63 76.04 83.92 97092
inTitle 80.98 69.64 74.88 96338
issue 95.02 78.68 86.08 30279
page 95.07 71.56 81.66 88516
pmcid 66.07 59.11 62.39 807
pmid 67.75 67.56 67.66 2093
title 84.05 71.76 77.42 92377
volume 94.83 83.41 88.75 87626
all fields (micro avg.) 88.83 74.03 80.76 706652
all fields (macro avg.) 84.29 72.35 77.74 706652

Soft Matching (ignoring punctuation, case and space characters mismatches)

Field-level results

label precision recall f1 support
authors 88.12 71.61 79.01 97092
date 90.05 74.78 81.71 97539
doi 77.64 77.54 77.59 16893
first_author 94.08 76.4 84.33 97092
inTitle 90.26 77.63 83.47 96338
issue 95.02 78.68 86.08 30279
page 95.07 71.56 81.66 88516
pmcid 76.87 68.77 72.6 807
pmid 71.78 71.57 71.67 2093
title 92.17 78.7 84.9 92377
volume 94.83 83.41 88.75 87626
all fields (micro avg.) 91.62 76.35 83.29 706652
all fields (macro avg.) 87.81 75.51 81.07 706652

Levenshtein Matching (Minimum Levenshtein distance at 0.8)

Field-level results

label precision recall f1 support
authors 92.75 75.36 83.16 97092
date 90.05 74.78 81.71 97539
doi 79.08 78.98 79.03 16893
first_author 94.24 76.53 84.46 97092
inTitle 91.1 78.34 84.24 96338
issue 95.02 78.68 86.08 30279
page 95.07 71.56 81.66 88516
pmcid 76.87 68.77 72.6 807
pmid 71.83 71.62 71.72 2093
title 95.31 81.38 87.79 92377
volume 94.83 83.41 88.75 87626
all fields (micro avg.) 92.83 77.37 84.4 706652
all fields (macro avg.) 88.74 76.31 81.93 706652

Ratcliff/Obershelp Matching (Minimum Ratcliff/Obershelp similarity at 0.95)

Field-level results

label precision recall f1 support
authors 89.91 73.06 80.61 97092
date 90.05 74.78 81.71 97539
doi 78.23 78.13 78.18 16893
first_author 93.68 76.08 83.97 97092
inTitle 89.01 76.55 82.31 96338
issue 95.02 78.68 86.08 30279
page 95.07 71.56 81.66 88516
pmcid 66.07 59.11 62.39 807
pmid 67.75 67.56 67.66 2093
title 94.33 80.54 86.89 92377
volume 94.83 83.41 88.75 87626
all fields (micro avg.) 91.9 76.59 83.55 706652
all fields (macro avg.) 86.72 74.5 80.02 706652

Instance-level results

Total expected instances:       98707
Total extracted instances:      96483
Total correct instances:        37769 (strict) 
Total correct instances:        47362 (soft) 
Total correct instances:        50792 (Levenshtein) 
Total correct instances:        48107 (RatcliffObershelp) 

Instance-level precision:   39.15 (strict) 
Instance-level precision:   49.09 (soft) 
Instance-level precision:   52.64 (Levenshtein) 
Instance-level precision:   49.86 (RatcliffObershelp) 

Instance-level recall:  38.26   (strict) 
Instance-level recall:  47.98   (soft) 
Instance-level recall:  51.46   (Levenshtein) 
Instance-level recall:  48.74   (RatcliffObershelp) 

Instance-level f-score: 38.7 (strict) 
Instance-level f-score: 48.53 (soft) 
Instance-level f-score: 52.04 (Levenshtein) 
Instance-level f-score: 49.29 (RatcliffObershelp) 

Matching 1 :    67251

Matching 2 :    4362

Matching 3 :    5614

Matching 4 :    2042

Total matches : 79269

Citation context resolution


Total expected references:   98705 - 49.4 references per article
Total predicted references:      96483 - 48.29 references per article

Total expected citation contexts:    142688 - 71.42 citation contexts per article
Total predicted citation contexts:   123504 - 61.81 citation contexts per article

Total correct predicted citation contexts:   98549 - 49.32 citation contexts per article
Total wrong predicted citation contexts:     24955 (wrong callout matching, callout missing in NLM, or matching with a bib. ref. not aligned with a bib.ref. in NLM)

Precision citation contexts:     79.79
Recall citation contexts:    69.07
fscore citation contexts:    74.04

Fulltext structures

Fulltext structure contents are complicated to capture from JATS NLM files. They are often normalized and different from the actual PDF content and are can be inconsistent from one document to another. The scores of the following metrics are thus not very meaningful in absolute term, in particular for the strict matching (textual content of the srtructure can be very long). As relative values for comparing different models, they seem however useful.

Evaluation on 1998 random PDF files out of 1998 PDF (ratio 1.0).

Strict Matching (exact matches)

Field-level results

label precision recall f1 support
figure_title 3.29 2.73 2.99 13147
reference_citation 70.52 66.38 68.39 147296
reference_figure 73.83 67.2 70.36 47901
reference_table 49.41 78.97 60.79 5938
section_title 69.31 67.48 68.38 32360
table_title 3.52 2.91 3.19 2951
all fields (micro avg.) 66.33 62.87 64.56 249593
all fields (macro avg.) 44.98 47.61 45.68 249593

Soft Matching (ignoring punctuation, case and space characters mismatches)

Field-level results

label precision recall f1 support
figure_title 64.44 53.45 58.43 13147
reference_citation 82.57 77.72 80.07 147296
reference_figure 74.45 67.76 70.95 47901
reference_table 49.76 79.52 61.21 5938
section_title 73.07 71.14 72.1 32360
table_title 48.36 40.02 43.8 2951
all fields (micro avg.) 77.31 73.28 75.24 249593
all fields (macro avg.) 65.44 64.94 64.43 249593

Evaluation metrics produced in 1133.739 seconds