Entity Matching in OCRed Documents with Redundant Databases
Nihel Kooli and Abdel Bela
¨
ıd
LORIA Campus, scientifique BP 239 54506, Vandoeuvre-l
`
es-Nancy Cedex, France
Keywords:
Entity Recognition, Entity Resolution, Entity Matching, OCRed Documents, Redundant Databases.
Abstract:
This paper presents an entity recognition approach on documents recognized by OCR (Optical Character
Recognition). The recognition is formulated as a task of matching entities in a database with their representa-
tions in a document. A pre-processing step of entity resolution is performed on the database to provide a better
representation of the entities. For this, a statistical model based on record linkage and record merge phases is
used. Furthermore, documents recognized by OCR can contain noisy data and altered structure. An adapted
method is proposed to retrieve the entities from their structures by tolerating possible OCR errors. A modified
version of EROCS is applied to this problem by adapting the notion of segments to blocks provided by the
OCR. It handles document segments to match the document to its corresponding entities. For efficiency, a
process of data labeling in the document is applied in order to filter the compared entities and segments. The
evaluation on business documents shows a significant improvement of matching rates compared to those of
EROCS.
1 INTRODUCTION
With the growth of industrial data and the multitude of
inflow sources in the companies, the processed docu-
ments can be of different types (electronic, scanned,
etc.) and different classes (purchase orders, invoices,
etc.). These documents often contain unstructured
textual data. For a better management, it is necessary
to automate the task of data processing in documents
by identifying and extracting contained information
and then structuring it.
The data contained in administrative documents
are often predefined in structured catalogs: databases
prepared by experts. An entity contained in a catalog
is represented by a series of fields: text values whose
semantics and types are informed by the headers of
the database.
Entity Recognition is the process of identifying
and locating a term or phrase in an unstructured tex-
tual document referring a particular entity such as a
person, a place, an organization, etc. In this context,
the problem can be reformulated as a task of match-
ing a mentioned entity in the document with its repre-
sentation in the database. This task is far from being
solved by a simple intersection between words in doc-
ument and those in database records and it is a more
complex one since:
• Documents do not explicitly mention unique iden-
tity of entities as defined in the database.
• Terms defining the entity in the document differ
from terms defining the entity in the database.
This is caused by non-standardized represen-
tations like abbreviations, incorrect or missing
punctuation and fused or split words.
• The extraction of content from documents recog-
nized by Optical Character Recognition (called
OCRed documents in this paper) is a challenging
task since we have to deal with OCR errors.
• OCR poorly reproduces the physical and the logi-
cal structures of the document.
• Industrial databases are often voluminous, contain
poor quality of data such as missing fields, typing
errors and non-standardized representation. Also,
they contain redundant records which could be
represented differently but refer to the same real
world entity. Such database is called redundant
database in this paper.
Entities defined in databases can be duplicated
since database can be either dynamic, managed
by different users, or a result of merging mul-
tiple databases. These problems related to the
database make the comparison task with docu-
ment complex. The task of resolving such prob-
lems is called entity resolution. An approach that
benefits from entity resolution to enhance entity
recognition in documents is proposed.
165
Kooli N. and Belaïd A..
Entity Matching in OCRed Documents with Redundant Databases.
DOI: 10.5220/0005177301650172
In Proceedings of the International Conference on Pattern Recognition Applications and Methods (ICPRAM-2015), pages 165-172
ISBN: 978-989-758-076-5
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
In this paper, we address the idiosyncrasies of en-
tity recognition in an OCRed document that contains
noisy data and has a specific structure. We propose
a solution that retrieves entities from OCR blocks
and identifies their structured representation in the
database in spite of the altered content of OCR. Fur-
thermore, we combine the two problems of entity res-
olution in industrial databases and entity recognition
in documents and we show how the entity resolution
task can enhance the entity matching task by improv-
ing the quality of the database.
The remainder of this paper is organized as fol-
lows: Section 2 presents some related research about
entity resolution and entity recognition, Section 3 de-
tails the proposed approach and Section 4 presents the
experiments on real world data.
2 RELATED WORK
Entity Resolution. To make matching decision be-
tween records, some researchers proposed determin-
istic approaches which are based on a preliminary
rule definition (Lee et al., 2000) and probabilistic ap-
proaches that use statistics to assign a probabilistic
weighting to record pairs. Fellegi and Senter pro-
posed in (Fellegi and Sunter, 1969) a statistical model
of linkage between records based on probabilistic def-
initions. This unsupervised machine learning method
views the problem as a task of defining a feature vec-
tor for record pairs into three classes: links, non-
links and possible links (undecided case requiring,
for example, human review). This suggested model
has been, by and large, adopted by subsequent re-
searchers.
Some works, such as (Bilenko et al., 2003), are
based on attribute comparisons in records using sim-
ilarity distances such as edit-based distances (exam-
ple: the edit-distance, Jaro-Winkler), token-based dis-
tances and hybrid ones (example: Soft-tf-idf). The
authors in (Cohen et al., 2003) studied the compari-
son between these different measures on record link-
age results.
In large databases, record pairs comparison is ex-
pensive since it consists of a cartesian product of
records. (Bilenko, 2006) proposed a step of blocking
which consists of separating the database into blocks
that contain approximately similar records. This sep-
aration could be based on a selection of keys such as
regrouping records that share the same zip-code or the
same first three characters of the name.
Entity Recognition. In the literature, several stud-
ies have addressed the problem of entity recognition
in unstructured documents. These studies could be
classified into three categories: context-oriented ap-
proaches, data-intensive approaches and mixed ap-
proaches.
• Context-oriented approaches are based on contex-
tual rules and require an explicit linguistic and
grammatical recognition of text using syntactic
and eventual semantic labeling. We mention,
for example, (Hashemi et al., 2003) that intro-
duced an entity recognition technique for extract-
ing names, titles and their associations. The prob-
lem with these approaches is their non-generality
since they are dependent on the language and the
domain of text. In addition, the task of defining
contextual rules is complex in time and resources.
• For data-intensive approaches, entity recognition
techniques do not require any explicit grammati-
cal knowledge of the document but obtain their in-
formation from an annotated corpus (for example
a knowledge database). These approaches treat
the problem as a classification task and apply ma-
chine learning models to solve it. Different su-
pervised learning methods were used for entity
recognition, such as (Zhang et al., 2010) that uses
Support Vector Machine (SVMs) for recognizing
investigator names in articles.
• Mixed entity recognition methods try to combine
the advantages of the machine learning and the
rule based approaches. For example, (Laishram
and Kaur, 2013) combines Conditional Random
Field (CRF) with rule definition that helps to iden-
tify features for some particular entities to im-
prove the classification by the CRF.
In the context of data-intensive approach, the corpus
could be a structured database that represents the en-
tities. For matching entities in documents with their
representations in a database, (Chakaravarthy et al.,
2006) proposed an algorithm, called EROCS, that
identifies entities embedded in document segments.
It uses a score, defined for an entity with respect to
a segment, that considers the frequency of the com-
mon terms in the segment and their importance in the
database. This work, concerns electronic textual doc-
uments. It has deficiencies in the case of OCRed doc-
uments caused by their altered structure and content.
Indeed, it treats strict comparison between terms and
considers the text as a sequence of lines.
The task of disambiguation in the case of more
than one record returned for an entity referenced in
the document has been treated by (Wu et al., 2007)
that exploited relationships between candidate entities
for different fragments (a fragment is defined as entity
features extracted from the document using seman-
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods
166
Figure 1: Global schema of the proposed system.
tic patterns) of the same document for identifying the
identity of fragments. In contrast, our approach does
not remedy ambiguity cases but it avoids them from
the beginning.
Some works treat the problem of Information Re-
trieval (IR) in noisy text in the context of OCR.
(Taghva et al., 2006) proposes studies that prove
the degradation of the information extraction process
caused by altered characters in OCR text. For OCR
error correction, (Pereda and Taghva, 2011) uses ap-
proximate string matching to remedy some character
misrecognition.
3 PROPOSED APPROACH
Figure 1 presents the global schema of the proposed
approach. It is a succession of two main modules. The
entity resolution module takes as input a database,
proposes to link its contained records and constructs
an entity model for records referring the same real
word entity. It then uses this model to produce as out-
put a cleaned database composed of entities by merg-
ing linked records. The entity recognition module
proposes to label data in the document. These labels
are used to filter segments in the document and enti-
ties in the database. Then, this module tries to find the
entities in the document segments.
3.1 Entity Resolution
The module of entity resolution is composed of two
main sub-tasks: record linkage and record merge de-
scribed in the following.
3.1.1 Record Linkage
Let E be a structured database that contains n records
{r
i
} and m columns {c
j
}. Each record r
i
is composed
of m attributes corresponding to each column. The at-
tribute of r
i
that corresponds to the column c
j
is noted
r
i
.c
j
. We propose to group the records in E into q
blocks {B
k
} using regrouping keys. Each block con-
tains records that may refer the same entity. The num-
ber of records in B
k
is defined as |B
k
|.
Let dist(a,b) be the similarity distance between
two field values a and b. These latter are decided to
be similar if their distance exceeds a threshold T 1.
For the linkage decision, the statistical model given by
(Fellegi and Sunter, 1969) is used. It estimates match-
ing M(p) and unmatching probabilities U(p) for each
field of index p. We propose to compute a ratio (ratio)
between each pair of records and decide to link them
if the value of ratio exceeds the threshold T2 (see Al-
gorithm 1). T 1 and T 2 are empirically defined.
The output of the record linkage process is repre-
sented as an entity model where each entity is repre-
sented by a tree of three levels: the root node repre-
sents the entity reference, the nodes of second level
represent the entity fields where each field node is re-
lated to nodes of the third level that correspond to dif-
ferent field values of the linked records.
3.1.2 Record Merge
Record merge considers linked records in the entity
model. It returns a new record that provides better
information.
For identical records, this step of fusion will
remove duplicates. For complementary records
where each record provides additional attributes,
we propose to gather all values in a merge resulting
record. For example, the following linked records r1
and r2 are merged into r3.
r1 = [name: Xerox, zip-code: 92202]
r2 = [name: Xerox, phone: 0825082081]
r3 = [name: Xerox, zip-code: 92202, phone:
0825082081]
EntityMatchinginOCRedDocumentswithRedundantDatabases
167
input : E // the database
T 1, T 2: int // the thresholds
output: E
0
// linked database
for k ← 1 to q do
for i ← 1 to |B
k
|- 1 do
for j ← i + 1 to |B
k
| do
Ratio = 0;
for p ← 1 to m do
d = dist(r
i
.c
p
,r
j
.c
p
);
if d ≥ T 1 then
Ratio+ = log(
M(p)
U(p)
) ∗ d;
else
Ratio+ = log(
1−M(p)
1−U(p)
);
end
end
if Ratio ≥ T 2 then
E
0
=link(E, r
i
,r
j
)
end
end
end
end
Algorithm 1: The record linkage algorithm.
A record is said dominated by another when the
second record contains more information that enables
it to give a higher quality description of the underlying
entity. In the case of domination between records, the
merge step will keep the record that better describes
the entity. In the previous example, we keep r1 which
is dominated by r3 since it describes the entity with
more attributes. For records with differences in some
attributes caused by various descriptions of the entity
or changes of its characteristics, different representa-
tions are retained in the resulting record. For exam-
ple, for the linked records r3 and r4, different phone
numbers of the entity are retained in the record r5 for
future use.
r4 = [name: Xerox, zip-code: 92202, phone:
0825082082]
r5 = [name: Xerox, zip-code: 92202, phone:
<0825082081; 0825082082>]
3.2 Entity Recognition
Entity recognition consists of matching entities in the
document with referring to the database. It includes
three main steps : data labeling, segment and entity
filtering and the matching described in the following.
3.2.1 Data Labeling
Data labeling consists of identifying some entity com-
ponents in the document in order to localize their con-
taining segments.
Dictionaries and regular expressions, defined from
instances in the database, are used for this labeling.
3.2.2 Segment and Entity Filtering
Comparing all terms in each segment with all terms
of each entity is a costly task. We propose, then,
to localize segments that may refer an entity in the
database and limit the search to only these segments.
This consists of filtering segments in the document
focusing only on segments that contain some labeled
data. Furthermore, the labeled data is used to filter
entities in the database and keep only the records that
have at least one field value that corresponds to one
label value of the same type. This step of filtering
will improve the efficiency of the matching.
3.2.3 Entity-document Matching Model
For entity matching in the document with the
database, we are inspired by EROCS algorithm
(Chakaravarthy et al., 2006) which is described in the
following. It perceives the document d as a set of
segments d = {s
i
} where each segment s
i
is a collec-
tion of terms s
i
= {t
j
}. Each record of the structured
database E is considered as an entity e, having its own
terms e = {t
l
} contained in its attributes. If e is an en-
tity that matches the segment s
l
, then each term t
p
in
s
l
corresponds to some term t
q
of the entity e.
The weight of a term t is defined as:
w(t) =
log((N/n(t)) if n(t) > 0
0 otherwise
(1)
where N is the total number of distinct entities in the
database, and n(t) is the number of distinct entities
that contain t.
Let T (e,s) be the set of terms that appear in the
segment s and contained as well in the entity e. The
score of an entity e with respect to a segment s is de-
fined by (Chakaravarthy et al., 2006) as:
score(e,s) =
∑
t∈T (e,s)
TF(t, s).w(t) (2)
For OCR error tolerance, this approach proposes
an adapted score that employs the edit distance as:
score(e,s) =
∑
t
1
∈close(θ,T (s),T (e))
TF(t
1
,s).w(t
2
) (3)
where close(θ,T (s), T (e)) is the set of terms t
1
in
s such that there is some t
2
in e with dist(t1,t2) ≤ θ.
dist(t1,t2) is defined as the edit distance between t
1
and t
2
. It computes the minimum number of edit char-
acters (insert, delete, substitute) required to convert
the string t
1
to the string t
2
with |t
1
| ≤ |t
2
|.
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods
168
Figure 2: (a) An example of image document. (b) The OCR result of the block surrounded in red color.
EROCS proposes to match each segment s in the doc-
ument with an entity e in the database where:
e
max
= arg max
e∈E
score(e,s)
To limit false positive (FP) matching cases, this
approach defines a threshold of rejection T as: if
score(e
max
,s) > T then we decide to match the seg-
ment s with the entity e
max
.
3.2.4 OCRed Documents Model for Matching
An OCRed document is represented by a hierarchy
of blocks containing lines and lines containing words.
Figure 2 (a) presents an example of a document where
blocks obtained from OCR are surrounded. Figure 2
(b) presents the OCR result of its block surrounded in
red color in the document.
A segment in the matching model detailed in Sec-
tion 3.2.3 corresponds, firstly, to a block in the OCR.
However, an entity can be split into more than one
segment of the OCR as shown in Figure 3. Indeed,
the entity surrounded of green color is split into four
segments. We propose then to increasingly construct
a segment by merging consecutive blocks.
Furthermore, OCR does not restitute the physical
and the logical structure of the original image doc-
ument. An entity can therefore be split into non-
consecutive blocks. For example, in Figure 3, the en-
tity is split into blocks 4,6,8 and 10. We use the block
coordinates in the OCR to merge contiguous blocks
into one segment.
For matching, Algorithm 2 details the proposed al-
gorithm. This algorithm merges increasingly contigu-
ous segments based on an iterative score computation.
Figure 3: An example of entity structure altered by OCR.
It identifies the set of segments that contain some la-
beled data (segSet). For each identified segment (s), it
enlarges the search area by adding each time the clos-
est segment while the matching score is increased dur-
ing λ consecutive merges. Once the merge is stopped,
the last λ added segments are removed from the re-
sulting segment. λ is experimentally set to 2. The
entity that corresponds to the resulting segment is re-
tained if its score exceeds a threshold T .
4 EXPERIMENTS
For the experiments, we use a database containing a
table of suppliers composed of 229345 records. It car-
ries information about suppliers such as their names,
addresses and phone numbers. The supplier is the en-
tity of interest. We consider a dataset of 500 image
documents and their OCR results. These documents
represent industrial invoices or purchase orders. For
evaluation, we use a ground truth data containing a ta-
ble that relates each document with its contained en-
tity identifiers in the database. A document can be
EntityMatchinginOCRedDocumentswithRedundantDatabases
169
input : E // the database
D // the document
output: matchE =
/
0 // matched entities
D
0
= dataLabeling (E,D);
segSet = segFilter (D
0
);
foreach s in segSet do
E
0
= entityFilter (E,D
0
,s);
score = max
e∈E
0
score(e,s);
emax = argmax
{e∈E
0
}
score(e,s);
i=0;
while i < λ do
s= addClosestSeg (s);
scoreNew = max
e∈E
0
score(e,s);
emax = argmax
{e∈E
0
}
score(e,s);
i + +;
if scoreNew ≥ score then
i = 0;
score = scoreNew;
end
end
s= removeClosestSeg (s,λ);
if score ≥ T then
matchE =
addEntity(matchE,emax,score);
end
end
Algorithm 2: The matching algorithm.
related to one supplier’s identifier or more (in the case
of an entity that can reference a client in the document
and is present in the DB as a supplier of an other doc-
ument). This table was manually prepared by an in-
dustrial expert.
4.1 Entity Resolution
For computing similarity between field attribute pairs,
we propose to use Soft-tf-idf measure combined with
Jaro-Winkler measure since we have multiterm at-
tributes with possible typing errors. The choice
is based on comparative study of similarity mea-
sures proposed in (Cohen et al., 2003) with a fixed
threshold = 0.8 for Jaro-Winkler distance.
The matching probability is defined as:
Mprobability = 1 − error rate where error rate
is defined as the percentage that a pair of matched
records do not agree on the field. It is estimated by a
preliminary review of labeled data. The unmatching
probability for each field is approximated with the
frequency of its distinct values. It is defined as:
Uprobability = 1/#distinct values. Figure 4 presents
the evolution of unmatched record pairs frequency
varying the value of coupling ratio. This Figure
Table 1: Record linkage for varying key values in blocking.
Blocks
number
Compared
pairs
number
Linked
pairs
number
Without block — 2643850 11750
Key1 953 48772 11599
Key2 498 104182 11746
Key3 1114 92411 11745
shows the three zones of the probabilistic model:
the non-matching zone, the matching zone and the
undecided zone. One may note the presence of FP
(False Positives) pairs in the zone of matching and
FN (False Negatives) in the zone of non-matching.
A matching pair is defined as a pair that refers the
same entity in reality. Precision and Recall are then
defined as:
Recall =
#correctly linked pairs
#matching pairs
Precision =
#correctly linked pairs
#linked pairs
Figure 5 presents the evolution of Precision, Re-
call and F-measure for varying the coupling ratio
threshold. The curves show that setting the threshold
value at 13 maximizes the value of F-measure. This
value is retained in the rest of the study.
For blocking evaluations, we define three different
keys (key1: the supplier’s name, key2: the first term of
the name and key3: concatenation of the first term of
the name with the zip-code ). Table 1 presents the
results for different key values obtained by entity res-
olution in a snippet of 2300 records of the Suppliers
database.
This Table shows that the blocking produces a sig-
nificant decrease in the number of compared pairs of
records compared to that of record linkage without
blocking. However, restricting the choice could in-
fluence the quality of results as for key1 where one
may notice that the low number of comparisons rela-
tively to other keys causes a reduction in the number
of linked pairs. One may notice that key3 decreases
the number of compared records compared to that of
key2 while retaining nearly the same number of linked
records. key3 is considered as the key value in the rest
of the study.
After a merging step, results show a decrease of
about 30% in the number of records in Suppliers table.
4.2 Entity Recognition
OCRed documents are parsed to extract segments
with their contained words. Also, a vector of terms
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods
170
Figure 4: Ratio coupling.
Figure 5: Precision, Recall and F-measure of record linkage
for varying threshold of ratio.
is constructed for each entity in the database and its
corresponding weight as defined in eq (1).
A relevant entity for a document is defined as an
entity present in the document and that refers a record
in the database. Precision and Recall are defined as:
Recall =
# relevant matched entities
# relevant entities
Precision =
# relevant matched entities
# matched entities
Figure 6 shows the evolution of Precision, Recall
and F-measure of entity matching in documents with
varying the threshold T defined in Section 3.2.3. It
shows that setting the threshold value at 17 maximize
the value of F-measure. This value is retained to eval-
uate the matching results in the rest of the study.
Entity recognition process is evaluated for the
main proposed improvement on the original version
of EROCS. M EROCS
1
is defined as the first mod-
ified version with OCR error tolerance that consists
of integrating edit distance in the matching score as
shown in eq (3). M EROCS
2
concerns the integra-
tion of the filtering step detailed in Section 3.2.2.
M EROCS
3
concerns the application of the Entity
Resolution process.
Table 2 presents the obtained results. It shows
that the Recall of matching increased from 67.58%
for EROCS to 71.05% for M EROCS
1
but the Run-
Figure 6: Precision, Recall and F-measure of entity match-
ing for varying threshold of score.
time increased slightly due to the edit distance com-
parisons. Also, it shows a significant decrease in the
run time for M EROCS
2
(from 70.8 to 6.2 sec per
document). Furthermore, the Recall and the Preci-
sion were improved respectively with 2.31 and 14.81
points for M EROCS
3
. The Runtime was also re-
duced by about 32% which is in relation to the re-
duction of about 30% of record number.
The increase of matching rates for M EROCS
3
is due to the step of merging records that completes
missing information in the matched entity and the
increase of some term weights in this entity thanks
to redundancy reduction. In addition, entity resolu-
tion solves some ambiguity cases due to the reduc-
tion of several records, referring the same entity and
maximizing the score, into only one record. It also
reduces the algorithm complexity and the execution
time thanks to the reduction of the number of com-
pared entities in database for each segment in the doc-
ument.
The error cases of matching are explained. In
about 5% of cases, failure is caused by the fact that
entity is not well represented in the document. For
example, the supplier of document could be repre-
sented only by a logo. In about 5% of cases, failure is
caused by the bad quality of scanned documents that
produces grave OCR errors. In about 4% of cases, it
is caused by non-standardized representations of en-
tity terms in the document and the database. In about
7% of cases, it is caused by incomplete entities in
database where some fields are missing even after en-
tity resolution step. Finally, in about 6%, failure is
caused by non contiguity of entity components.
5 CONCLUSION AND FUTURE
WORK
We present a method, called M EROCS, for entity
matching in the database with their representations
EntityMatchinginOCRedDocumentswithRedundantDatabases
171
Table 2: Entity matching rates.
Recall (%) Precision (%) Fmeasure (%) Runtime (sec/doc)
EROCS (Chakaravarthy et al., 2006) 67.58 54,09 60,09 69.5
M EROCS
1
(+OCR error tolerance) 71,05 53,89 61,29 70.8
M EROCS
2
(+filtering) 71,05 54,77 61,86 6.2
M EROCS
3
(+entity resolution) 73,36 69,58 71,43 4,4
by segments in OCRed document. The extensions on
term matching and segment restructure of EROCS are
proven effective for OCRed documents which have
altered content and structure. A filtering step based
on data labeling reduces the runtime from 70.8 sec
to 6.2 sec per document. The pre-processing step of
entity resolution on the database improves the match-
ing rates with 2.31 points for the recall and 14.81
points for the precision and it decreases the runtime
with about 1.8 seconds by document. The results on a
dataset of 500 documents are promising and achieve
about 73% for recall and about 70% for precision.
The future work is to solve the problem of non-
contiguity of elements composing an entity. In case
of incomplete entity, we will choose from distant la-
beled elements those they complete correctly the en-
tity. The choice will be focused on the elements in-
creasing the matching score. Furthermore, we will
plan the use of other datasets, limited in this study to
supplier entities, in order to enlarge the field search
to all elements (close and distant) with more complex
spatial relations. The idea is to integrate the physical
and logical structures of the document and to exploit
them in the element searching. Another prospect is to
apply other methods for OCR matching and correc-
tion. A dictionary that maintains spell variations of
fields, such as abbreviations and character variations,
will be used.
ACKNOWLEDGEMENT
We would like to thank our collaborator ITESOFT for
providing real word data (images, OCRed documents
and database) for test.
REFERENCES
Bilenko, M. (2006). Adaptive blocking: Learning to scale
up record linkage. In Proceedings of the 6th IEEE
International Conference on Data Mining, pages 87–
96.
Bilenko, M., Mooney, R., Cohen, W., Ravikumar, P., and
Fienberg, S. (2003). Adaptive name matching in
information integration. IEEE Intelligent Systems,
18(5):16–23.
Chakaravarthy, V. T., Gupta, H., Roy, P., and Mohania, M.
(2006). Efficiently linking text documents with rele-
vant structured information. In International Confer-
ence on Very Large Data Bases, pages 667–678.
Cohen, W. W., Ravikumar, P., and Fienberg, S. E. (2003).
A comparison of string distance metrics for name-
matching tasks. In Proceedings of IJCAI-03 Workshop
on Information Integration, pages 73–78.
Fellegi, I. P. and Sunter, A. B. (1969). A theory for record
linkage. Journal of the American Statistical Associa-
tion, 64(328):1183–1210.
Hashemi, R. R., Ford, C., Bansal, A., Sieloff, S. D., and Tal-
burt, J. R. (2003). Building semantic-rich patterns for
extracting features from events of an on-line newspa-
per. In Proceedings of the IADIS International Con-
ference WWW/Internet, pages 627–634.
Laishram, J. and Kaur, D. (2013). Named entity recognition
in Manipuri: a hybrid approach. In The International
Conference of the German Society for Computational
Linguistics and Language Technology, volume 8105,
pages 104–110.
Lee, M.-L., Ling, T. W., and Low, W. L. (2000). Intelli-
clean: a knowledge-based intelligent data cleaner. In
ACM SIGKDD Conference on Knowledge Discovery
and Data Mining, pages 290–294.
Pereda, R. and Taghva, K. (2011). Fuzzy information ex-
traction on OCR text. In ITNG, pages 543–546.
Taghva, K., Beckley, R., and Coombs, J. S. (2006). The
effects of OCR error on the extraction of private infor-
mation. In Document Analysis Systems, pages 348–
357.
Wu, N., Talburt, J., Heien, C., Pippenger, N., Chiang, C.-
C., Pierce, E., Gulley, E., and Moore, J. (2007). A
method for entity identification in open source docu-
ments with partially redacted attributes. J. Comput.
Small Coll., 22(5):138–144.
Zhang, X., Zou, J., Le, D. X., and Thoma, G. R. (2010).
Investigator name recognition from medical journal
articles: a comparative study of svm and structural
svm. In Document Analysis Systems, ACM Interna-
tional Conference Proceeding Series, pages 121–128.
ICPRAM2015-InternationalConferenceonPatternRecognitionApplicationsandMethods
172
