Advanced citation matching and large-scale cited reference extraction

Advanced citation matching and large-scale cited reference extraction

Nees Jan van Eck

Centre for Science and Technology Studies (CWTS), Leiden University

EXCITE Workshop 2017: “Challenges in Extracting and Managing References”

Cologne, Germany, March 30, 2017

Outline

• Citation matching– Comparison of the accuracy of the Web of Science, CWTS, and

iFQ citation matching algorithms

• Cited reference extraction– Assessment of the accuracy of cited references in Web of Science

based on Elsevier ScienceDirect data

1

Accuracy of the WoS, CWTS, and iFQcitation matching algorithms

2

3

Citation matching problem

4

…

References

[1] Hirsch, JE (2005)PNAS, 102, p.16569

[2] Egghe, L (2006)Scientist, 20, p.15

…

An index to quantify an individual's scientific research output

Hirsch, JE

PNAS, 102(46), p.16569-72UT: 000233462900010

Abstract…

How to improve the h-index

Egghe, L

The Scientist, 20(3), p.15UT: 000235634200013

Abstract…

Bibliographic databaseWoS, Scopus

A

B

C

Why is citation matching difficult?

• ‘Big data’ problem– No. of publications: 50 million– No. of cited references: 1 billion

• Little data available on cited references in WoS– First author (last name and initials)– Source title (abbreviated)– Publication year– Volume number– First page number– (DOI)

• Errors in data– Citation extraction errors

• OCR errors• Interpretation errors due to different citation styles

– Typos and other human errors

5

/A Olensky, M/Y 2015/W J ASS INFORM SCI TEC/V 67/P 2550

Citation matching algorithms of WoS

• Little is known about the citation matching algorithm used in WoS

• Larsen et al. (2007) concluded from their investigation of missed matches in WoS that the algorithm is quite conservative and does not allow for any variations

6

Citation matching algorithm of CWTS

• The aim is to overcome the problem of missed citation matches in WoS

• Iterative, rule-based algorithm:1. Preprocessing

2. Start with the most restrictive matching rules

3. Continue with less restrictive matching rules

• Less restrictive matching rules allow for various types of inaccuracies in the cited reference data

7

Example matching rules

• Most restrictive matching rule:– Exact match on

• first author• publication year• publication name• volume number• starting page number• DOI

• Less restrictive matching rule:– Match on

• Soundex encoding of the last name of the first author• publication year plus or minus one• volume number• starting page number

8

Citation matching algorithms of iFQ

• Iterative, rule-based algorithm

• Allows non-unique matches of a single cited reference with several target articles

9

Data collection (1)

• Builds on data collected by Olensky (2015)

• Sample of 300 publications (cited pubs)– 2 science domains

– 6 disciplines

– 2 languages

– 2 publication years

• 3975 corresponding cited references in WoS– Times cited used to find cited references that are linked in WoS

– Cited reference search used to find cited references that are not linked in WoS

10

Data collection (2)

11

300 cited pubs

3975 citing pubs

not linkedin WoS

linkedin WoS

Results

12

All matches WoS CWTS iFQ

# % # % # %

Correct matches 3664 99.2 3855 98.8 3856 98.5

Incorrect matches 29 0.8 45 1.2 57 1.5

All citations WoS CWTS iFQ

# % # % # %

Correct matches 3664 93.8 3855 98.6 3856 98.7

Missed matches 244 6.2 53 1.4 52 1.3

Recall

Precision

WoS CWTS iFQ

F1 score 96.4 98.7 98.6

Qualitative analysis of missed matches(WOS: 244; CWTS: 53; iFQ: 52)

13

Changes in CWTS citation matching algorithm

• Introduction of a matching rules in which:1. Volume and issue number are interchanged

2. Volume and first page number are interchanged

• Small change in the order in which the matching rules are applied

14

Improved results

15

All matches CWTS (current) CWTS (revised)

# % # %

Correct matches 3888 99.6 3906 99.8

Incorrect matches 16 0.4 9 0.2

All citations CWTS (current) CWTS (revised)

# % # %

Correct matches 3888 98.7 3906 99.1

Missed matches 53 1.3 35 0.9

Recall

Precision

CWTS (current) CWTS (revised)

F1 score 99.1 99.4

Conclusions

• A significant number of citation matches are missing in WoS

• Substantial improvement in recall is possible, but at the cost of a small decrease in precision

• Citation matching algorithm of CWTS performs quite well

• During the analysis, various problems were detected in WoS cited reference extraction

16

Accuracy of WoS cited reference extraction

17

Introduction

• Aim: To determine the accuracy of WoS cited references data

• Approach: Comparison of the cited references extracted from the full text of Elsevier publications with the cited references available in WoS

18

Data

• Elsevier full text data– ScienceDirect API

– Subscription-based journal publications in the period 1987-2016

• WoS meta data– Document types ‘article’ and ‘review’

• Matching of Elsevier full-text data and WoS meta data at the level of individual publications

19

20

21

22

Validation of missing cited references in WoS

• Publication year: 2015

• Number of missing cited references: 73,536

• Sample size: 60– Missing cited reference: 33 (55.0%)

– Incorrect cited reference: 10 (16.7%)

– Error in meta data of cited reference (e.g., incorrect publication year or incorrect volume number): 16 (26.7%)

– Correct cited reference: 1 (1.5%)

23

Missing cited references in WoS (1)

24

Missing cited references in WoS (2)

25

Missing cited references in WoS (3)

26

Missing cited references in WoS (4)

27

Missing cited references in WoS (5)

28

Missing cited references in WoS (6)

29

Missing cited references in WoS (7)

30

Missing cited references in WoS (8)

31

Missing cited references in WoS (9)

32

Incorrect cited references in WoS (1)

33

Incorrect cited references in WoS (2)

34

Incorrect cited references in WoS (3)

35

Incorrect cited references in WoS (4)

36

Incorrect cited references in WoS (5)

37

Incorrect cited references in WoS (6)

38

WoS cited reference Original cited reference in publication

WANG J, 2006, CHINESE CHEM LETT, V17, P49

J. Wang, J.K. Carson, M.F. North, D.J. Cleland, Int. J. Heat Mass Transfer 49 (17) (2006) 3075–3083.

KANBER B, 2013, CEREBROVASC DIS S2, V35, P21

Kanber B, Hartshorne TC, Horsfield MA, Naylor AR, Robinson TG, Ramnarine KV. Dynamic variations in the ultrasound gray-scale median of carotid artery plaques. Cardiovasc Ultrasound 2013a;11:21.

EVANS P, 2010, TLS-TIMES LIT S 0326, P30

Evans PD, Chowdhury MJA. Photoprotection of wood using polyester-type UVabsorbersderived from the reaction of 2 hydroxy-4(2,3-epoxypropoxy)-benzophenone with dicarboxylic acid anhydrides. J Wood ChemTechnol 2010;30:186e204.

Incorrect cited references in WoS (7)

39

WoS cited reference Original cited reference in publication

CAO X, 2010, IEEE GLOBECOMM 2010, V2010, P1

Cao, X., Zong, Z., Ju, X., Sun, Y., Dai, C., Liu, Q., Jiang, J., 2010. Molecular cloning, characterization and function analysis of the gene encoding HMG-CoA reductase from Euphorbia Pekinensis Rupr. Mol. Biol. Rep. 37, 1559e1567.

LI XY, 2013, NANJING NONGYE DAXUE, V36, P36

X. Li, S. Wang, Y. Chen, G. Liu, X. Yang, Overexpression of CD40 in sacral chordomasand its correlation with low tumor recurrence, Onkologie 36 (10) (2013) 567–571

ZHANG K, 2014, IEEE T PATTERN ANAL, V1, P1

K. Zhang, H. Chen, G. Wu, K. Chen, H. Yang, High expression of SPHK1 in sacral chordomaand association with patients’ poor prognosis, Med. Oncol. 31 (11) (2014) 247.

More cited references in WoS (1)

40

More cited references in WoS (2)

41

More cited references in WoS (3)

42

More cited references in WoS (4)

43

More cited references in WoS (5)

44

More cited references in WoS (6)

45

Conclusions

• About 0.3% of cited references are missing in WoS

• About 0.2% of cited references in WoS have minor errors (e.g., incorrect publication year or volume number)

• About 0.1% of cited references in WoS have major errors (i.e., reference to completely incorrect target document)

• WoS does a good job in handling references pointing to multiple target documents

• These results are based on Elsevier publications only; publications from other publishers may yield different outcomes 46

Thank you for your attention!

47