A Comparison of Personal Name Matching: Techniques and Practical Issues

TL;DR: The characteristics of personal names are discussed and potential sources of variations and errors are presented and a comprehensive number of commonly used, as well as some recently developed name matching techniques are overview.

Abstract: Finding and matching personal names is at the core of an increasing number of applications: from text and Web mining, search engines, to information extraction, dedupli- cation and data linkage systems Variations and errors in names make exact string matching problematic, and ap- proximate matching techniques have to be applied When compared to general text, however, personal names have different characteristics that need to be considered In this paper we discuss the characteristics of personal names and present potential sources of variations and errors We then overview a comprehensive number of commonly used, as well as some recently developed name matching techniques Experimental comparisons using four large name data sets indicate that there is no clear best matching technique

Summary

1. Introduction

• Increasingly large amounts of data are being created, communicated and stored by many individuals, organisations and businesses on a daily basis.
• At lot of this data contains some information about people, for example e-mails, customer and patient records, news articles, business and political memorandums.
• As reported in [28], the use of approximate comparison methods does improve the matching quality in these applications.
• Personal names have characteristics that makes them different to general text.
• While similar comparison studies on matching techniques have been done in the past [9, 17, 20, 25, 32, 34], none has analysed and compared such a comprehensive number of techniques specifically with application to personal names.

2. Personal name characteristics

• Even when only considering the English-speaking world, a name can have several different spelling forms for a variety of reasons.
• In the Anglo-Saxon region and most other Western countries, a personal name is usually made of a given name, an optional middle name, and a surname or family name [24].
• Other specific errors were differences in punctuation marks and whitespaces (for example ‘O’Connor’, ‘OConnor’ and ‘O Connor’) in 12% of errors, and different last names for female patients (8% of errors).
• Thus, there seem to be significant differences between general text and personal names, which have to be considered when name matching algorithm are being developed and used.

2.1 Sources of name variations

• Besides the variations in personal names discussed above, the nature of data entry [19] will determine the most likely types of errors and their distribution.
• Manual keyboard based data entry can result in wrongly typed neighbouring keys (for example ‘n’ and ‘m’, or ‘e’ and ‘r’).
• Finally, people themselves sometimes report their names differently depending upon the organisation they are in contact with, or deliberately provide wrong or modified names.
• When matching names, one has to deal with legitimate name variations (that should be preserved and matched), and errors introduced during data entry and recording (that should be corrected) [3].
• The challenge lies in distinguishing between these two sources of variations.

3. Matching techniques

• Name matching can be defined asthe process of determining whether two name strings are instances of the same name[24].
• In the following three subsections the authors present the most commonly used as well as several recently proposed new techniques.
• Without proper parsing and segmentation a name (even if stored in two fields as given- and surname) can contain several words separated by a hyphen, apostrophe, whitespace or other character.
• Frequency distributions of name values can also be used to improve the quality of name matching.
• I.e. only the basic techniques used to compare two names without taking any context information into account.the authors.

3.1 Phonetic encoding

• Most techniques – including all presented here – have been developed mainly with English in mind.
• The transformed name string is then encoded into a one-letter three-digits code (again removing zeros and duplicate numbers) using the following encoding table.
• It contains many rules that take the position within a name, as well as previous and following letters into account (similar to Phonix).
• When matching names, phonetic encoding can be used as a filtering step (calledblocking in data linkage [6, 30]), i.e. only names having the same phonetic code will be compared using a computationally more expensive pattern matching algorithm.

3.2 Pattern matching

• A similarity measure can be calculated by dividing the total length of the common sub-strings by the minimum, maximum or average lengths of the two original strings (similar to Smith-Waterman above).
• Positionalq-grams can be padded with start and end characters similar to non-positionalq-grams, and similarity measures can be calculated in the same three ways as with non-positionalq-grams.
• The Winkler algorithm therefore increases the Jaro similarity measure for agreeing initial characters (up to four).

3.3 Combined techniques

• Two techniques combine phonetic encoding and pattern matching with the aim to improve the matching quality.
• The edit costs in Editex are 0 if two letters are the same, 1 if they are in the same letter group, and 2 otherwise.
• Similar to basic edit distance, the time and space complexities of matching two stringss1 ands2 with Editex areO(|s1|×|s2|) andO(min(|s1|, |s2|)), respectively.
• This recently developed technique, calledSyllable Alignment Pattern Searching (SAPS)[13] is based on the idea of matching two names syllable by syllable, rather than character by character.
• The experimental results presented in [13] indicate that SAPS performs better than Editex, edit distance and Soundex on the same large name data set used in [25] (the COMPLETE data set the authors are using in their experiments as well).

4 Experiments and discussion

• In this section the authors discuss the results of a series of comparison experiments using four large name data sets.
• The aim of these experiments was to see which matching techniques achieve the best matching quality for different personal name types, and to compare their computational performance.
• All name matching techniques were implemented in Python as part of theFebrl (Freely Extensible Biomedical Record Linkage)3 data linkage system [5].

4.1 Name data sets

• Three of the test data sets were based on given- and surnames extracted from a health data set containing midwives’ records (women who gave birth) from the Australian state of New South Wales [4].
• A deduplication status in this data (indicating which records correspond to the same women) allowed us to extract true name pairs (known matches).
• The authors then created afull namedata set by concatenating given- with surnames (separated by a whitespace).
• The fourth data set was created in a similar way using the COMPLETE name database [13, 25] by forming surname pairs from 90 randomly chosen and manually matched queries.
• Table 2 shows the size of their four test data sets.

4.2 Distribution of edit distances

• In order to better understand their test data, the authors calculated the edit distances for all the known name pairs.
• This indicates the challenge 3http://datamining.anu.edu.au/linkage.html of name matching: how to correctly classify two names that are very different.

4.3 Matching results

• The two techniques that combine phonetic encoding with pattern matching (Editex and syllable alignment distance) do not perform as well as one might have expected, and neither do skip-grams.
• Details for the best performing pattern matching techniques on the four data sets can be seen in Figure 1.
• An optimal value for one data set and technique will very likely result in sub-optimal quality for another data set or technique.

4.4 Timing results

• As shown in Table 6, the phonetic encoding techniques (times shown include encoding of two names) are generally much faster than pattern matching, due to their complexity beingO(|s|) for a given strings.
• Phonix with its many rules is the slowest phonetic techniques (almost ten times as slow as others), while Smith-Waterman is the slowest pattern matching techniques.

5 Recommendations

• The mixed results presented in the previous section indicate that there is no single best name matching technique, and that the type of personal name data to be matched has to be considered when selecting a matching technique.
• The following recommendations will help with this.
• It is important to know the type of names to be matched, and if these names have been properly parsed and standardised [7], or if the name data potentially contains several words with various separators.
• Phonetic encoding followed by exact comparison of the phonetic codes should not be used.
• Even small changes of the threshold can result in dramatic drops in matching quality.

Figures

Figure 1. Best f-measure results for the four data sets (similarity measures on the horizontal and f-measures on the vertical axis).

Table 3. Distribution of edit distances for matched name pairs.

Table 1. Phonetic name encoding examples.

Table 2. Number of name pairs and single names in test data sets used for experiments.

Table 6. Timings results in milli-seconds (shortest times shown boldface and longest times underlined).

Table 4. Average f-measure values (best results shown boldface and worst results underlined).

Full text

TR-CS-06-02
A Comparison of Personal Name
Matching: Techniques and Practical
Issues
Peter Christen
September 2006
Joint Computer Science Technical Report Series
Department of Computer Science
Faculty of Engineering and Information Technology
Computer Sciences Laboratory
Research School of Information Sciences and Engineering

This technical report series is published jointly by the Department of
Computer Science, Faculty of Engineering and Information Technology,
and the Computer Sciences Laboratory, Research School of Information
Sciences and Engineering, The Australian National University.
Please direct correspondence regarding this series to:
Technical Reports
Department of Computer Science
Faculty of Engineering and Information Technology
The Australian National University
Canberra ACT 0200
Australia
or send email to:
Technical-DOT-Reports-AT-cs-DOT-anu.edu.au
A list of technical reports, including some abstracts and copies of some full
reports may be found at:
http://cs.anu.edu.au/techreports/
Recent reports in this series:
TR-CS-06-01 Stephen M Blackburn, Robin Garner, Chris Hoffmann,
Asjad M Khan, Kathryn S McKinley, Rotem Bentzur, Amer
Diwan, Daniel Feinberg, Daniel Frampton, Samuel Z Guyer,
Martin Hirzel, Antony Hosking, Maria Jump, Han Lee,
J Eliot B Moss, Aashish Phansalkar, Darko Stefanovi
´
c,
Thomas VanDrunen, Daniel von Dincklage, and Ben
Wiedermann. The DaCapo benchmarks: Java benchmarking
development and analysis (Extended Version). September
2006.
TR-CS-05-01 Peter Strazdins. CycleCounter: an efficient and accurate
UltraSPARC III CPU simulation module. May 2005.
TR-CS-04-04 C. W. Johnson and Ian Barnes. Redesigning the intermediate
course in software design. November 2004.
TR-CS-04-03 Alonso Marquez. Efficient implementation of design patterns
in Java programs. February 2004.
TR-CS-04-02 Bill Clarke. Solemn: Solaris emulation mode for Sparc Sulima.
February 2004.
TR-CS-04-01 Peter Strazdins and John Uhlmann. Local scheduling
out-performs gang scheduling on a Beowulf cluster. January
2004.

A Comparison of Personal Name Matching: Techniques and Practical Issues
Peter Christen
Department of Computer Science, The Australian National University
Canberra ACT 0200, Australia
Peter.Christen@anu.edu.au
Abstract
Finding and matching personal names is at the core of an
increasing number of applications: from text and Web min-
ing, information retrieval and extraction, search engines,
to deduplication and data linkage systems. Variations and
errors in names make exact string matching problematic,
and approximate matching techniques based on phonetic
encoding or pattern matching have to be applied. When
compared to general text, however, personal names have
different characteristics that need to be considered.
In this paper we discuss the characteristics of personal
names and present potential sources of variations and er-
rors. We overview a comprehensive number of commonly
used, as well as some recently developed name matching
techniques. Experimental comparisons on four large name
data sets indicate that there is no clear best technique.
We provide a series of recommendations that will help re-
searchers and practitioners to select a name matching tech-
nique suitable for a given data set.
1. Introduction
Increasingly large amounts of data are being created,
communicated and stored by many individuals, organisa-
tions and businesses on a daily basis. At lot of this data con-
tains some information about people, for example e-mails,
customer and patient records, news articles, business and
political memorandums. Even most scientific and techni-
cal documents contain details about their authors. Personal
names are often used to search for documents in large col-
lections. Examples include Web searches (the most popular
query in the last few years on Google has always been a
celebrity name, with another four or five names ranked in
the top ten queries
1
), retrieval of medical patient records,
or bibliographic searches (using author names). Names are
also important pieces of information when databases are
1
http://www.google.com/press/zeitgeist.html
deduplicated (e.g. to find and remove duplicate customer
records), and when two data sets are linked or integrated
and no unique entity identifiers are available [5, 6, 30]. As
reported in [28], the use of approximate comparison meth-
ods does improvethe matching quality in these applications.
Personal names have characteristics that makes them
different to general text. While there is only one cor-
rect spelling for many words, there are often several valid
spelling variations for personal names, for example ‘Gail’,
‘Gale’ and ‘Gayle’. People also frequently use (or are
given) nicknames in daily life, for example ‘Bill’ rather
than the more formal ‘William’. Personal names sometimes
change over time, for example when somebody gets mar-
ried. Names are also heavily influenced by people’s cul-
tural backgrounds. These issues make matching of personal
names more challenging compared to matching of general
text [3, 24].
As names are often recorded with different spellings, ap-
plying exact matching leads to poor results. In [11], for
example, the percentage of name mismatches in three large
hospital databases ranged between 23% and 36%. To im-
prove matching accuracy, many different techniques for ap-
proximate name matching have been developed in the last
four decades [15, 20, 25, 34], and new techniques are still
being invented [13, 18]. Most techniques are based on a
pattern matching, phonetic encoding, or a combination of
these two approaches.
Computational complexity has to be considered when
name matching is done on very large data sets. The time
needed to determine if two names match is crucial for the
overall performance of an application (besides data struc-
tures that allow to efficiently extract candidate name pairs
while filtering out likely non-matches [23]). Matching
speed is vital when quick response times are needed, for
example in search engines, or crime and biomedical emer-
gency response systems, where an answer should be avail-
able within a couple of seconds.
While similar comparison studies on matching tech-
niques have been done in the past [9, 17, 20, 25, 32, 34],
none has analysed and compared such a comprehensive

number of techniques specifically with application to per-
sonal names. The contributions of this paper are a detailed
discussion of the characteristics of personal names and pos-
sible sources of variations and errors in them, an overview
of a range of name matching techniques, and a comparison
of their performance using several large real world data sets
containing personal names.
We start in Section 2 with a discussion of personal name
characteristics and sources of variations. In Section 3 we
first look at different situations and contexts of name match-
ing, and then present a comprehensive number of name
matching techniques. The results of experimental compar-
isons are discussed in Section 4, and a series of recommen-
dations is given in Section 5 that will help researchers and
practitioners who are faced with the problem of selecting a
name matching technique. Finally, conclusions and an out-
look to future work is discussed in Section 6.
2. Personal name characteristics
Even when only considering the English-speaking
world, a name can have several different spelling forms for
a variety of reasons. In the Anglo-Saxon region and most
other Western countries, a personal name is usually made
of a given name, an optional middle name, and a surname
or family name [24]. Both ‘Gail Vest’ and ‘Gayle West’
might refer to the same person, while ‘Tina Smith’ might be
recorded in the same database as ‘Christine J. Smith’ and as
‘C.J. Smith-Miller’. People change their name over time,
most commonly when somebody gets married (in which
case there are different cultural conventions and laws of
how a person’s name is changed). Compound names are
often used by married women, while in certain countries
husbands can take on the surname of their wives.
In daily life, people often use (or are given) nicknames.
These can be short forms of their given names (like ‘Bob’
for ‘Robert’, or ‘Liz’ for ‘Elizabeth’), they can be varia-
tions of their surname (like ‘Vesty’ for ‘Vest’) or they might
relate to some life event, character sketch or physical char-
acteristics of a person [3]. While having one given and one
middle name is common for Anglo-Saxon names, several
European countries favour compound given names instead,
for example ‘Hans-Peter’ or ‘Jean-Pierre’. In general, there
are no legal regulations of what constitutes a name [3].
In today’s multi-cultural societies and worldwide data
collections (e.g. global online businesses or international
crime and terrorism databases), the challenge is to be able to
match names coming from different cultural backgrounds.
For Asian names, for example, there exist several translit-
eration systems into the Roman alphabet [24], the surname
traditionally appears before the given name, and frequently
a Western given name is added. Hispanic names can con-
tain two surnames, while Arabic names are often made of
several components and contain various affixes that can be
separated by hyphens or whitespaces.
An early study [10] on spelling errors in general words
found that over 80% of errors were single errors – either a
letter was deleted, an extra letter was inserted, a letter was
substituted for another letter, or two adjacent letters were
transposed. Substitutions were the most common errors,
followed by deletions, then insertions and finally transpo-
sitions, followed by multiple errors in one word. Other
studies [15, 19, 27] reported similar results. However, in
a study [11] that looked at patient names within hospital
databases, different types and distributions of errors were
found. With 36%, insertion of an additional name word,
initial or title were the most common errors. This was fol-
lowed in 14% of errors by several different letters in a name
due to nicknames or spelling variations. Other specific er-
rors were differences in punctuation marks and whitespaces
(for example ‘O’Connor’, ‘OConnor’ and ‘O Connor’) in
12% of errors, and different last names for female patients
(8% of errors). Single errors in this study accounted for
39% of all errors, only around half compared to the 80%
reported in [10]. Thus, there seem to be significant differ-
ences between general text and personal names, which have
to be considered when name matching algorithm are being
developed and used. According to [20] the most common
name variations can be categorised as
• spelling variations (like ‘Meier’ and ‘Meyer’) due to
typographical errors that do not affect the phonetical
structure of a name but still post a problem for match-
ing;
• phonetic variations (like ‘Sinclair’ and ’St. Clair’)
where the phonemes are modified and the structure of
a name is changed substantially;
• compound names (like ’Hans-Peter’ or ‘Smith Miller’)
that might be given in full (potentially with differ-
ent separators), one component only, or components
swapped;
• alternative names (like nicknames, married names or
other deliberate name changes); and
• initials only (mainly for given and middle names).
In [19] character level (or non-word) misspellings are
classified into (1) typographical errors, where it is assumed
that the person doing the data entry does know the correct
spelling of a word but makes a typing error (e.g. ‘Sydeny’
instead of ‘Sydney’); (2) cognitive errors, assumed to come
from a lack of knowledge or misconceptions; and (3) pho-
netic errors, coming from substituting a correct spelling
with a similar sounding one. The combination of phonet-
ical and spelling variations, as well as potentially totally
changed name words, make name matching challenging.

2.1 Sources of name variations
Besides the variations in personal names discussed
above, the nature of data entry [19] will determine the most
likely types of errors and their distribution.
• When handwritten forms are scanned and optical char-
acter recognition (OCR) is applied [15, 27], the most
likely types of errors will be substitutions between
similar looking characters (like ‘q’ and ‘g’), or sub-
stitutions of one character with a similar looking char-
acter sequence (like ‘m’ and ‘r n’, or ‘b’ and ‘l i’).
• Manual keyboard based data entry can result in
wrongly typed neighbouring keys (for example ‘n’ and
‘m’, or ‘e’ and ‘r’). While in some cases this is quickly
corrected by the person doing the data entry, such er-
rors are often not recognised, possibly due to limited
time or by distractions to the person doing the data
entry (imagine a busy receptionist in a hospital emer-
gency department). The likelihood of letter substitu-
tions obviously depends upon the keyboard layout.
• Data entry over the telephone (for example as part of
a survey study) is a confounding factor to manual key-
board entry. The person doing the data entry might not
request the correct spelling, but rather assume a default
spelling which is based on the person’s knowledge and
cultural background. Generally, errors are more likely
for names that come from a culture that is different to
the one of the person doing the data entry, or if names
are long or complicated (like ‘Kyzwieslowski’) [11].
• Limitations in the maximum length of input fields can
force people to use abbreviations, initials only, or even
disregard some parts of a name.
• Finally, people themselves sometimes report their
names differently depending upon the organisation
they are in contact with, or deliberately provide wrong
or modified names. Or, while somebody might report
her or his name consistently in good faith, others report
it inconsistently or wrongly for various reasons.
If data from various sources is used, for example in a text
mining, information retrieval or data linkage system, then
the variability and error distribution will likely be larger
than if the names to be matched come from one source only.
This will also limit the use of trained name matching algo-
rithms [2, 9, 31] that are adapted to deal with certain types
of variations and errors. Having meta-data that describes
the data entry process for all data to be used can be valuable
when assessing data quality.
As discussed previously, while there is only one cor-
rect spelling for most general words, there are often no
wrong name spellings, just several valid name variations.
For this reason, in many cases it is not possible to disre-
gard a name as wrong if it is not found in a dictionary
of known names. When matching names, one has to deal
with legitimate name variations (that should be preserved
and matched), and errors introduced during data entry and
recording (that should be corrected) [3]. The challenge lies
in distinguishing between these two sources of variations.
3. Matching techniques
Name matching can be defined as the process of deter-
mining whether two name strings are instances of the same
name [24]. As name variations and errors are quite com-
mon [11], exact name comparison will not result in good
matching quality. Rather, an approximate measure of how
similar to names are is desired. Generally, a normalised
similarity measure between 1.0 (two names are identical)
and 0.0 (two names are totally different) is used.
The two main approaches for matching names are pho-
netic encoding and pattern matching. Different techniques
have been developed for both approaches, and several tech-
niques combine the two with the aim to improve the match-
ing quality. In the following three subsections we present
the most commonly used as well as several recently pro-
posed new techniques.
Matching two names can be viewed as an isolated prob-
lem or within a wider database or application context. Four
different situations can be considered.
1. The matching of two names that consist of a single
word each, not containing whitespaces or other sep-
arators like hyphens or commas. This is normally the
situation when names have been parsed and segmented
into components (individual words) [7], and all sepa-
rators have been removed. Full names are split into
their components and stored into fields like title, given
name, middle name, surname and alternative surname.
Parsing errors, however, can result in a name word
being put into the wrong field, thereby increasing the
likelihood of wrong matching.
2. Without proper parsing and segmentation a name (even
if stored in two fields as given- and surname) can con-
tain several words separated by a hyphen, apostrophe,
whitespace or other character. Examples include com-
pound given names, born surname and married name,
name pre- and suffixes, and title words (like ‘Ms’, ‘Mr’
or ‘Dr’). In this situation, besides variations in a single
word, parts of a name might be in a different order or
missing, and there might be different separators. All
this will complicate the name matching task.
3. In the first two situations names were matched indi-
vidually without taking any context information into

Frequently asked questions

Q1. What are the contributions mentioned in the paper "A comparison of personal name matching: techniques and practical issues" ?

In this paper the authors discuss the characteristics of personal names and present potential sources of variations and errors. The authors provide a series of recommendations that will help researchers and practitioners to select a name matching technique suitable for a given data set. 

Q2. What is the method for a string to be sorted?

If a string contains more than one word (i.e. it contains at least one whitespace or other separator), then the words are first sorted alphabetically before the Winkler technique is applied (to the full strings). 

Q3. What are the likely types of errors in handwritten forms?

When handwritten forms are scanned and optical character recognition (OCR) is applied [15, 27], the most likely types of errors will be substitutions between similar looking characters (like ‘q’ and ‘g’), or substitutions of one character with a similar looking character sequence (like ‘m’ and ‘r n’, or ‘b’ and ‘l i’).• 

Q4. How did the authors extract the names from the records that did not have duplicates?

The authors also extracted single names from records that did not have duplicates, and randomly created name pairs (the same number as known matched pairs in order to get balanced test data sets). 

Q5. What is the importance of determining if two names match?

The time needed to determine if two names match is crucial for the overall performance of an application (besides data structures that allow to efficiently extract candidate name pairs while filtering out likely non-matches [23]). 

Q6. What is the algorithm for a name that contains initials only?

As it allows for gaps, the Smith-Waterman algorithm should be especially suited for compound names that contain initials only or abbreviated names 

Q7. What is the q-gram similarity measure between two strings?

A q-gram similarity measure between two strings is calculated by counting the number of q-grams in common (i.e. q-grams contained in both strings) and divide by either the number of q-grams in the shorter string (called Overlap coefficient2), the number in the longer string (called Jaccard similarity) or the average number of q-grams in both strings (called the Dice coefficient). 

Q8. How many people are creating, communicating and storing data?

Increasingly large amounts of data are being created, communicated and stored by many individuals, organisations and businesses on a daily basis. 

Q9. What is the technique for removing obvious non-matches?

As expected, the Bag distance is very fast (followed by simple q-grams), making it suitable as a filtering technique to remove obvious non-matches. 

Q10. What are the results of their experiments with skip-grams?

Their experiments with skip-grams using multi-lingual texts from different European languages show improved results compared to bigrams, trigrams, edit distance and the longest common sub-string technique.•