A Pattern-Based Approach to Name and Address Parsing with

Active Learning

Onais Khan Mohammed, Khizer Syed, John Talburt, Adeeba Tarannum,

Abdul Kareem Khan Kashif, Salman Khan, Najmudin Syed and Syed Yaser Mehdi

Department of Information Quality, University of Arkansas at Little Rock,

2801 S. University Ave., Little Rock, AR, U.S.A.

Keywords: Address Parsing, Active Learning, Data Standardization, Entity Resolution, Pattern Recognition, Knowledge

Graphs, Tokenization, US Postal Standards, Machine Learning Models in Parsing, Semantic Data Integration.

Abstract: Processing population data often requires parsing demographic items into a standard set of fields to achieve

metadata alignment. This paper describes a novel approach based on token pattern mappings augmented by

active learning. Input strings are tokenized and a token mask is created by replacing each token with a single-

character code indicating the token’s potential function in the input string. A user-created mapping then directs

each token represented in the mask to its correct functional category. Testing has shown the system to be as

accurate as, and in some cases, more accurate than comparable parsing systems. The primary advantage of

this approach over other systems is that it allows a user to easily add a new mapping when an input does not

conform to any previously encoded mappings instead of having to reprogram system parsing rules or retrain

a supervised parsing machine learning model.

1 INTRODUCTION

Many systems designed to process multiple sources

of the same information require each source to define

the same attributes for essential parts of the record

(Mohammed, Mahmood, 2022). This is especially

true when improving the quality of record linkage,

entity resolution, and data integration processes. For

example, one source may have a single field

containing the entire postal address of a person or

business whereas another source may separate the

street address from the city name and yet another even

may have separate fields for the components of the

street address such as street number and street name.

When different sources define attributes in different

ways, the process to standardize the attributes is

called metadata alignment. The most common

approach to standardization of attributes is to parse

(decompose) less structured attributes into the same

set of fundamental components (Elhamifar, E.,

Sapiro, G., Yang, A., & Sasrty, S. S., 2013). For

population (name and address) data these

fundamental components, while not universally

standardized, are generally well understood. These

components may vary by region or country, however.

at its core, the US Address Data Preparation Function

is designed to take an unstructured address string and

break it down into a series of meaningful components.

This includes identifying and separating the street

address, city, state, and postal code. These

components are then stored in a structured format,

allowing them to be easily retrieved and used in

various applications.

The use of HiPER indices, Boolean rules, and

scoring rules is one of the key benefits of the US

Address Data Preparation Function. HiPER indices

are used to search and retrieve data from large data

sets and are essential for high-performance data

processing applications. Boolean rules are used to

determine the validity of data, and scoring rules are

used to determine the relevance of data. (M.

Mohammed, 2021) These features allow for the

efficient processing of large amounts of data and

ensure that only accurate and relevant data is used in

applications. In addition to the US Address Data

Preparation Function, there are other data preparation

functions that require the definition of new XML

elements in the Entities Script. These functions are

designed to perform specific tasks, such as removing

duplicates, reformatting data, and transforming data

into a standardized format.

Mohammed, O. K., Syed, K., Talburt, J., Tarannum, A., Kashif, A. K. K., Khan, S., Syed, N. and Mehdi, S. Y.

A Pattern-Based Approach to Name and Address Parsing with Active Learning.

DOI: 10.5220/0013077500003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 70-77

ISBN: 978-989-758-737-5; ISSN: 2184-433X

In the context of record linkage, address parsing is a

critical step due to the ubiquity of address fields in

databases (Churches, T, 2002) This process involves

segmenting raw addresses into semantic fields, which

is essential for identifying records referring to the

same entity across different data collections. The

main challenge in address parsing arises from the

variability and inconsistency in address formats,

including differences in field orders and the presence

of errors. (Sorokine, A., Kaufman, 2002) The

evaluation of address parsing accuracy is focused on

the correct assignment of each element of an address

string into its appropriate field, with individual fields

like 'Street' being assessed separately. Traditional

rule-based parsing methods often struggle with the

complexity of real-world addresses, leading to a

growing preference for flexible, learning-based

approaches, such as machine learning models

(Elhamifar, 2013). Overall, effective address parsing

is crucial for ensuring high-quality data in record

linkage, especially given the diverse and imperfect

nature of address data encountered in various real-

world scenarios. Dictionary-based address parsing

uses the concept of string matching, where input

address strings are compared against a pre-compiled

dictionary containing known addresses or key address

components. This method is straightforward,

involving matching segments of an address with

dictionary entries to accurately identify and classify

address components.

The effectiveness of rule-based approaches is

particularly pronounced with standard and well-

known address formats. Rule-based address parsing

adopts a structured approach, utilizing predefined

rules reflective of common address formats and

components. This method focuses on developing

techniques based on the unique traits of address

components and their organizational structure.

Parsing algorithms under this approach are designed

to recognize and interpret various address elements,

understanding their hierarchical and syntactical

relationships. This method’s flexibility allows it to

adapt to a variety of address structures, including

complex and non-standard formats. However, it

requires a detailed understanding of address

formatting rules and can be more complex to

implement.

2 METHODOLOGY

The system has two major components, the

automated parsing system, and a manual exception

handling system. In the method described here, the

system assumes that the input comprises only name

and address words and that the name words precede

the address words. The automated parsing system has

three major sub-components. The first component

reads the name and address input, parses it into

tokens, removes some punctuation, uppercases the

tokens, then creates a token mask by assigning each

token a character based on a small lookup table of

clue words.

Figure 1: Flow Diagram of the name and address parsing

using active learning.

2.1 Parsing Module

Note some punctuation, especially commas and

hyphens, are kept as important pattern clues. If a

token is not found in the lookup table it is classified

as either a “W” token type if it starts with a letter, or

an “N” token type if it starts with a digit. Based on an

analysis of the mask, the system divides the input

tokens into two groups, name tokens and address

tokens. The second sub-component processes the

name tokens. In this step, the name tokens are used to

generate a name token mask using a different clue

word table. The name clues identify tokens

commonly associated with the five name word

categories including prefix titles such as “MR” and

“MRS”, common given and family name,

generational suffixes, and suffix titles such as “PHD”

and “CPA”. As before, tokens not found are assigned

either “W” token type or “N” token type. If the

resulting name mask is found, then each token is

mapped to its appropriate name word category.

It is important to note that the token mappings are

created by domain experts, not automatically

assigned by token type. The expert can interpret the

overall meaning of the mask’s pattern. For example,

the token “JUNIOR” is tagged in the name clue table

by its most likely use as a generation suffix token, a

type “J” token. However, it could also occur as a

given name.

For example, assume that “JOHN” is in the table

as a common given name of “G” token type, and that

A Pattern-Based Approach to Name and Address Parsing with Active Learning

“DOE” is not in the table. The name “JOHN DOE,

JUNIOR” would generate the mask “GW,J” which

would be mapped by the domain expert as “G” to the

given name attribute, “DOE” to the surname attribute,

and “J’ to the generational suffix attribute. On the

other hand, the name “JUNIOR DOE” while

generating the mask “JW” would be mapped by the

domain expert as “JUNIOR” to the given name

attribute and “DOE” to the surname attribute despite

“J” indicating a generational suffix.

The third sub-component processes the address

tokens following a similar scheme. This component

has a separate address-specific clue word table. In

addition, there are currently 6 address token types

used to identify 15 address word categories as shown

in Table 1. So, for example the “D” token type

identifies a directional token such as “N” for north

and “SE” for southeast.

However, the token identified as type “D” may be

used as a predicational address attribute, e.g., “N

OAK ST” or a post directional address attribute, e.g.,

“E ST NORTH”. These examples again illustrate the

need for pattern interpretation of the mask by a

domain expert.

In the street address “N OAK ST”, the “N” token

is identified as a “D” type token and would be mapped

to the predicational attribute. However, in “E ST

NORTH”, both the “E” token and the “NORTH”

token would be identified as “D” type tokens, the “E”

token would be mapped to the street name attribute,

and the token “NORTH” would be mapped to the post

directional attribute.

For both the name parsing and the address parsing

to successfully complete, the token mask generated

by the lookup table process must be found in a

knowledge base of previously created mask

mappings. If either the name mask or address mask is

not found, then the parsing operation fails for that

input and the input and mask are both written to an

exceptions file that is the input to the Exception

Processing System, This mask mapping entry is then

inserted into the mask-mapping knowledge base so

that thereafter, any input generating the same mask

will be automatically processed by the automated

parsing system when a match for a token is not found

in the pre-defined dictionary of masks. This step is

important because not all addresses will conform to

the standard format used in the dictionary, and some

addresses may contain non-standard or ambiguous

components that cannot be matched to a specific

address field.

(i) Searching for a Match:

When a token is generated and the corresponding

mask does not match any of the pre-defined masks in

the dictionary, the program will search for a match by

comparing the token to a list of common address

components. This list may include common street

names, city names, and state abbreviations it also uses

the Levenshtein similarity scores to to find the closest

match for example Junior and ‘Junir’ wherein we

have a missing ‘o’ can help us achieve a robust

mechanism to assign token which is the epitome of

correctness.

(ii) Assigning to an Exception:

If a match is still not found after searching the list of

common components, the program will assign the

token to an exception. This is a catch-all category that

represents any component of the address that cannot

be matched to a specific field. Examples of

exceptions may include apartment or suite numbers,

building names, or unusual address formats.

(iii) Adding the Exception to the US Address:

Once the token has been assigned to an exception, the

program will add it to the US address components as

a separate field. This allows the exception to be

included in the final output, even if it cannot be

matched to a specific address field.

The fifth step of the program involves adding the

address tokens that were generated in the second step

to the US address components. This step builds on the

previous steps by combining the cleaned and

tokenized address components with the assigned

address fields to create a complete US address.

(i) Assigning Tokens to Address Fields:

Based on the comparison of masks in the third step,

the program assigns each token to a specific address

field, such as the street name, city, state, and zip code.

This creates a set of address fields, where each field

corresponds to a specific component of the address.

(ii) Adding Address Tokens to the Address Fields:

Once the tokens have been assigned to the address

fields, the program then adds these tokens to the

appropriate address field in the US address

components. For example, if the token "Main" was

assigned to the street name field, the program would

add "Main" to the street name component in the US

address.

(iii) Building the Complete US Address:

Once all tokens have been added to the appropriate

address fields, the program can then combine these

components to create a complete US address. This

involves concatenating the address components in the

correct order, and separating them with the

appropriate punctuation, such as commas and spaces.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

2.2 Annotated Dataset

We have tested our program by giving inputs like

attention line addresses, individual addresses,

highway addresses, university addresses, P.O. Box

addresses, Puerto Rico addresses and individual

names and generated truth files respectively. The

focus of our program is to generate as much of Mask

to Dictionaries, and that too by ensuring Data Quality.

A truth file is a reference file that contains known or

verified names and addresses. The parser uses this file

to compare the input name or address to the known or

verified names and addresses, helping to identify and

correct errors in the input data.

Table 1: An example of address parsing using active

learning.

Pos Token Code Dictionaries

1 123-1/2 N USAD

SNO

2 N D USAD_SPR

3 OA

W USAD_SNM

4 STREET F USAD_SFX

5 APT S USAD

ANM

6 3A N USAD

ANO

7 LITTLE W USAD_CTY

8 ROC

W USAD_CTY

9 AR

T USAD_STA

10 72203-4352 N USAD

ZIP

Parsing Puerto Rico addresses presented

significant challenges, particularly due to the

inclusion of Spanish street names, suffixes, and

regional dialects. For example, interpreting street

names like 'Callejón' or suffixes such as 'Esq.'

(Esquina, meaning corner) proved to be particularly

difficult. The complexity of these language-specific

elements often led to errors in address processing.

However, by employing active learning

techniques alongside a non-exhaustive token table

populated with contextual clue words, we were able

to overcome these obstacles. This method

significantly improved our ability to accurately parse

and process these complex addresses, demonstrating

the effectiveness of our approach.

The approach of tokenizing addresses and

mapping tokens to specific categories like

USAD_SNO, USAD_CTY, or USAD_ZIP as shown

in Table 1 plays a vital role in standardizing and

systematizing the parsing process. In this

methodology, once a token pattern has been identified

and mapped, the same pattern can be automatically

recognized and categorized in future inputs without

requiring manual intervention each time. This

capability is particularly advantageous in applications

that deal with large volumes of address data, ensuring

consistency and efficiency

Table 2: An example of name parsing using active learning.

Pos Token Code Dictionaries

1 DR P PREFIX_TITLE

2 JOHN G FIRST_NAME

3 TABLURT L SURNAME

4 JR J GEN_SUFFIX

5 PHD Q SUFFIX

TITLE

For instance, when the pattern "123-1/2 N OAK

STREET APT 3A LITTLE ROCK ARK 72203-

4352" is first processed, it involves manual or semi-

automated categorization where each segment of the

address is assigned a specific dictionary based on its

identified role (numeric, directional, word, etc.).

After this initial classification, the system stores these

mappings in a knowledge base.

When a new address comes in that matches a

previously encountered pattern, such as another entry

starting with a similar structured numeric street

number followed by a directional indicator, the

system automatically applies the same categorization

rules. It recognizes that "123-1/2" should be classified

under USAD_SNO and "N" under USAD_SPR, and

so forth, based on the stored mappings. This pattern

recognition not only accelerates the parsing process

but also enhances accuracy by applying proven rules

to new data.

In the provided example for name parsing in Table

2, the tokenization process for a personal name "Dr.

John Tablurt Jr. IQCP" involves breaking down the

name into discrete elements and categorizing each

part using a specific code linked to a predefined

dictionary. The sequence starts with "DR," positioned

as the first token and categorized under the code 'P'

for PREFIX_TITLE, recognizing titles or honorifics.

Following this, 'JOHN,' the first or given name, is

assigned the code 'G' and classified under

FIRST_NAME, which is vital for personal

identification. 'TABLURT' functions as the surname

or last name, coded as 'L' and categorized under

SURNAME, playing a key role in family

identification. 'JR,' a generational suffix that indicates

lineage, is given the code 'J' and falls under

GENERATIONAL_SUFFIX. Lastly, 'PHD,'

representing a professional or academic qualification,

is coded as 'Q' and grouped under SUFFIX_TITLE.

Now, consider a scenario where the last name is

missing. The address parser, using context clues and

the code 'Q,' can intelligently determine that 'IQCP'

A Pattern-Based Approach to Name and Address Parsing with Active Learning

should not be interpreted as a last name but rather as

a suffix title. This structured categorization is

essential for accurately processing and analyzing

each name component for various administrative and

data management purposes.

Table 3: Comparison of Active Learning and USaddress

Performance on Puerto Rican Address Parsing.

Component Active Learning USAddress

Address

Number

1000 1000

Street Name PONCE DE LEON Avenida Ponce

de Leon

Street Suffix AVENIDA N/A

Occupancy

Type

STE Ste

Occupancy

Identifier

5 5

City Name SAN JUAN San Juan

State Name PR PR

Zip Code 907 907

3 RESULTS

We have conducted a comparison between the results

of an active learning approach and the usaddress

Python library, which employs a rule-based method

to parse U.S. addresses. This comparison highlights

the effectiveness of active learning, particularly in

handling complex addresses such as those found in

Puerto Rico.

The active learning method demonstrated superior

performance by accurately parsing each component

of the address, including the differentiation between

the street name, suffix, and occupancy type. For

example, while the usaddress library struggled to

parse certain elements like the street suffix

("AVENIDA"), the active learning model correctly

identified and categorized it.

This is particularly important in addresses from

Puerto Rico, where non-standard formats and Spanish

language elements often pose challenges for rule-

based parsers.

Table 4: Precision and Recall metrics: Active Learning Vs

Rule-Based System.

Index

P-

Active

Learni

P-

Rule-

Based

R-

Active

Learni

R-

Rule -

Based

USAD_SNO 0.9914 0.9167 1 0.6

USAD_SPR 0.9231 0 1 0

USAD_SNM 0.9854 0.6329 1 0.2924

USAD_SFX 0.985 0.6912 1 0.3672

USAD_CTY 0.9931 0.5165 0.9797 0.5639

USAD_STA 0.9932 0.7748 1 0.5911

USAD_ZIP 0.9968 0.8341 0.9968 0.5981

USAD_ORG 1 0 0.9796 0

USAD_MGN 1 0 1 0

USAD_HNO 1 0.7123 0.9859 0.7324

USAD_ANM 1 1 0.9892 0.0114

USAD_MDG 1 0 1 0

USAD_HNM 1 0.7692 1 1

USAD_SPT 0.9231 0 1 0

USAD_RNM 0.98 0.96 1 0.9796

USAD_BNO 1 1 0.9864 0.9592

USAD_ANO 1 0.44 0.9892 0.1264

USAD_RNO 0.9804 0.875 1 0.98

Overall

Accuracy

0.9934 0.4755 - -

Micro

Average

0.9934 0.7586 0.9934 0.4755

Weighted

Average

0.9935 0.6046 0.9934 0.4755

In this evaluation, we compare the performance of

two approaches—Active Learning and a Rule-Based

System—across various address components, as

defined by the USAD Conversion Dictionary. This

dictionary maps human-readable address parts to

standardized USAD codes, including components

like street number (USAD_SNO), city

(USAD_CTY), and organization (USAD_ORG).

Active Learning demonstrated significantly better

performance across all metrics. For instance, it

achieved near-perfect precision, recall, and F1-score

for street numbers (USAD_SNO), with values of

0.9914, 1.0000, and 0.9957, respectively. In contrast,

the Rule-Based System struggled with a precision of

0.9167 and a recall of 0.6000, resulting in an F1-score

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

of 0.7253. This indicates that the rule-based approach

could not adequately capture the variations in street

numbers, likely due to rigid or incomplete rule

definitions.

For pre-directionals (USAD_SPR), the

discrepancy was even more pronounced. The Active

Learning model managed a perfect recall and a

precision of 0.9231, leading to an F1-score of 0.9600.

Meanwhile, the Rule-Based System completely failed

in this category, with all metrics at zero, suggesting a

lack of adequate rules for recognizing pre

directionals. The street name (USAD_SNM)

component also showed a clear advantage for Active

Learning, with an F1-score of 0.9926 compared to

0.4000 for the Rule-Based System. The latter’s

performance was hindered by a low recall of 0.2924

and a precision of 0.6329, indicating that it could

neither consistently identify nor correctly label street

names.

Similarly, in identifying city names

(USAD_CTY), the Active Learning approach

exhibited high precision and recall, achieving a F1-

score of 0.9863. The Rule-Based System lagged

significantly, with metrics indicating a much lower

accuracy in parsing this component. Overall, the

Active Learning method

4 CONCLUSION

In this paper, we have presented a method for address

parsing that employs an active learning approach,

particularly effective for handling complex cases

such as Puerto Rican addresses. It is important to

clarify that our use of "active learning" refers to an

interactive, user-driven process rather than a machine

learning-based approach. Our system is primarily

dictionary-based, relying on a robust set of predefined

rules and dictionaries to parse addresses. What sets

our approach apart is the integration of user feedback

to handle exceptions and edge cases. When the

system encounters an address that doesn’t fit the

existing rules, it prompts the user to provide

feedback. This feedback is used to refine the rules and

update the dictionaries, allowing the system to adapt

to new or unusual address formats.

Rather than employing complex machine learning

algorithms, our system uses simple distance metrics

to identify potential matches and discrepancies within

the address components. The feedback loop is crucial

here—by continually refining the parsing rules based

on user input, the system becomes more accurate over

time, even without the use of machine learning. This

iterative process enables the system to handle a wide

variety of address formats, including those that

deviate from standard patterns, making it both

flexible and reliable.

Active learning offers significant advantages over

traditional machine learning models, particularly in

scenarios that require the flexibility to make

incremental improvements without the need to

overhaul the entire dataset or system rules. This

characteristic is especially beneficial in parsing

systems, like those used for processing names and

addresses, where variability and exceptions are

common. In traditional deterministic or machine

learning models, if an error is identified, addressing

this error typically requires retraining the model or

revising the entire rule set. This process can be time-

consuming and resource-intensive, as it may involve

recalibrating the model parameters or rewriting rules

based on the new data.

Moreover, such adjustments risk affecting the

performance on other parts of the dataset that were

previously correct, a phenomenon known as the

"stability-plasticity dilemma." In contrast, active

learning allows for more targeted interventions.

When an error is detected in active learning systems,

only the specific part of the system—such as a

particular mask or dictionary entry associated with

the error—needs to be adjusted. This is done by either

updating the existing entries or adding new ones to

the dictionary to handle the exception. This selective

updating preserves the integrity and accuracy of the

other parts of the system, ensuring that previous

learning and correct mappings are not disturbed by

changes made to address new or outlier cases.

This method enhances the system's adaptability

and efficiency, enabling it to improve continuously as

it processes new information without requiring

comprehensive retraining or global rule adjustments.

Such a model is not only more scalable but also more

maintainable, as it allows for fine-grained

improvements that directly address specific issues or

adapt to new types of data encountered in operational

environments. Active learning thus provides a

practical approach in dynamic settings where data can

change frequently, or new patterns emerge over time,

making it a superior choice for applications that

benefit from ongoing learning and adaptation without

the need for frequent, broad-scale modifications.

A Pattern-Based Approach to Name and Address Parsing with Active Learning

Table 5: Address Parsing Improvements with Active

Learning vs. Rule-Based System.

Use Case Input

Example

Rule-Based

Parsing

Result

Active Learning

Parsing Result

Puerto Rican

Address

Parsing

"1000

Avenida

Juan Ponce

de León, Apt

3C, San

Juan, PR

00907"

Struggles with

stName vs.

suffix,

misclassified

"Avenida,"

and

incorrectly

parses apt.

number.

Correctly

identifies

"Avenida Juan

Ponce de León" as

the street name

and accurately

parses "Apt 3C."

Puerto Rican

Address with

Directional

Component

"123 Calle

de la Rosa

Oeste, Apt

2B,

Guaynabo,

PR 00966"

Will

misinterpret

"Oeste"

(West) as part

of the street

name instead

of a

directional

suffix.

Correctly

identifies "Calle

de la Rosa" as the

street name and

"Oeste" as the

directional suffix,

accurately parsing

"Apt 2B."

Pre-

directional

and Post-

directional

Components

"123 North

East Street

Drive, Little

Rock, AR

30003"

Will classify

"North" as the

only pre-

directional

and "East

Street" as the

street name,

failing to

recognize

"Street" as a

suffix.

Correctly

identifies "North

East" as a

combined pre-

directional

("NE"), "Street" as

a suffix, and

"Drive" as a post-

directional.

Complex

Urban

Address

"456 West

Market

Street Plaza,

Chicago, IL

60605"

Struggles to

differentiate

between

"Market

Street" and

"Plaza,"

leading to

incorrect

parsing.

Successfully

identifies "West

Market Street" as

the street name

and "Plaza" as the

post-directional

suffix.

Puerto Rican

Address with

Multiple

Components

"789 Calle

San

Francisco,

Piso 4,

Oficina 12,

Old San

Juan, PR

00901"

Will

incorrectly

parse "Piso 4"

(Floor 4) and

"Oficina 12"

(Office 12),

potentially

merging these

with other

components.

Accurately

identifies "Calle

San Francisco" as

the street name,

and correctly

parses "Piso 4"

and "Oficina 12"

as distinct

occupancy

identifiers.

ACKNOWLEDGEMENTS

This research was partially supported by the

Department of Commerce US Census Cooperative

Agreement CB21RMD0160002. We express our

profound gratitude to Dr. John R. Talburt, the Project

Coordinator. His invaluable guidance and support

have been instrumental in the progression and success

of this project. We also extend our thanks to Syed

Yaser Mehdi for his comprehensive research on the

literature review which has significantly enriched our

understanding and findings. We are deeply grateful to

the University of Arkansas at Little Rock for

providing us with the necessary resources and an

inspiring platform to conduct our research. Their

commitment to fostering an environment of academic

excellence and innovation has been crucial to our

work. Additionally, we would like to acknowledge

the US Census Bureau for sponsoring this project.

Their collaboration and the data provided have been

pivotal in shaping the research outcomes. This project

not only benefited from the contributions mentioned

above but also reflected the collaborative spirit and

the shared goal of advancing knowledge in the field

of Information Quality.

REFERENCES

Evaluation of Automation Techniques for Data Quality

Assessment for Party and Product Master Data.

Mohammed, Mahmood. University of Arkansas at

Little Rock ProQuest Dissertations Publishing, 2022.

29206326.

Churches, T., Christen, P., Lim, K. et al. Preparation of

name and address data for record linkage using hidden

Markov models. BMC Med Inform Decis Mak 2, 9

(2002). https://doi.org/10.1186/1472-6947-2-9

Sorokine, A., Kaufman, J., Piburn, J., & Stewart, R. (2020).

Active Learning Approach to Record Linking in Large

Geodatasets. Oak Ridge National Lab.(ORNL), Oak

Ridge, TN (United States).

Elhamifar, E., Sapiro, G., Yang, A., & Sasrty, S. S. (2013).

A convex optimization framework for active learning.

In Proceedings of the IEEE International Conference

on Computer Vision (pp. 209-216).

Hogan, H., Cantwell, P. J., Devine, J., Mule, V. T., and

Velkoff, V. “Quality and the 2010 Census.” Population

Research and Policy Review, 32(5):637–662 (2013).

Qian, K., Popa, L., & Sen, P. (2017, November). Active

learning for large-scale entity resolution. In

Proceedings of the 2017 ACM on Conference on

Information and Knowledge Management (pp. 1379-

1388).

Li, X., Kardes, H., Wang, X., & Sun, A. (2014, November).

Hmm-based address parsing with massive synthetic

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

training data generation. In Proceedings of the 4th

International Workshop on Location and the Web (pp.

33-36).

Li, X., Talburt, J. R., & Li, T. (2018, December). Scoring

matrix for unstandardized data in entity resolution. In

2018 International Conference on Computational

Science and Computational Intelligence (CSCI) (pp.

1087-1092). IEEE.

Maddodi, S., Attigeri, G. V., & Karunakar, A. K. (2010,

November). Data deduplication techniques and

analysis. In 2010 3rd International Conference on

Emerging Trends in Engineering and Technology (pp.

664-668). IEEE.

Meduri, V. V., Popa, L., Sen, P., & Sarwat, M. (2020, June).

A comprehensive benchmark framework for active

learning methods in entity matching. In Proceedings of

the 2020 ACM SIGMOD International Conference on

Management of Data (pp. 1133-1147).

Nafa, Y., Chen, Q., Chen, Z., Lu, X., He, H., Duan, T., &

Li, Z. (2022). Active deep learning on entity resolution

by risk sampling. Knowledge-Based Systems, 236,

107729.

Simonini, G., Zecchini, L., Bergamaschi, S., & Felix, N.

(2022). Entity Resolution On-Demand.

Tu, J., Han, X., Fan, J., Tang, N., Chai, C., Li, G., & Du, X.

(2022). DADER: hands-off entity resolution with

domain adaptation. Proceedings of the VLDB

Endowment, 15(12), 3666-3669.

Ye, Y., & Talburt, J. R. (2019). Generating synthetic data to

support entity resolution education and research. Journal

of Computing Sciences in Colleges, 34(7), 12-19.

M. Mohammed, J. R. Talburt, S. Dagtas and M.

Hollingsworth, "A Zero Trust Model Based Framework

For Data Quality Assessment," 2021 International

Conference on Computational Science and

Computational Intelligence (CSCI), Las Vegas,

NV, USA, 2021, pp. 305-307, doi:

10.1109/CSCI54926.2021.00123.

A Pattern-Based Approach to Name and Address Parsing with Active Learning