An Automatic Method for Building a Taxonomy of Areas of Expertise

Thi Thu Le

, Tuan-Dung Cao

, Xuan Lam Pham

, Duc Trung Pham

and Toan Luu

Department of Research Methodology, Thuongmai University, Hanoi, Vietnam

School of Information and Communication Technology, Hanoi University of Science and Technology, Hanoi, Vietnam

Department of Computer Science, School of Information Technology in Economics,

National Economics University, Hanoi, Vietnam

Move Digital AG, Zurich, Switzerland

Keywords: Topic Taxonomy Construction, Expert Finding System, Expertise Profile.

Abstract: Although a lot of Expert finding systems have been proposed, there is a need for a comprehensive study on

building a knowledge base of areas of expertise. Building an Ontology creates a consistent lexical framework

of a domain for representing information, thus processing the data effectively. This study uses the background

knowledge of machine learning methods and textual data mining techniques to build adaptive clustering, local

embedding, and term ordering modules. By that means, it is possible to construct an Ontology for a domain

via representation language and apply it to the Ontology system of expert information. We proposed a new

method called TaxoGenDRK (Taxonomy Generator using Database about Research Area and Keyword)

based on the method from Chao Zhang et al. (2018)’s research on TaxoGen and an additional module that

uses a database of research areas and keywords retrieved from the internet – the data regarded as an uncertain

knowledge base for learning about taxonomy. DBLP dataset was used for testing, and the topic was “computer

science”. The evaluation of the topic taxonomy using TaxogenDRK was implemented via qualitative and

quantitative methods, producing a relatively good accuracy compared to other existing studies.

1 INTRODUCTION

For searching systems in general and Expert Finding

Systems (EFS) in particular, it is essential to build a

knowledge database of the research areas of experts

because this will enhance the quality of search and

recommendation algorithms (Abramowicz et al.,

2011; Husain et al., 2019). An EFS will contain a lot

of information about experts in different research

fields (Al-Taie et al., 2018). Thus, each area needs a

consistent vocabulary framework to display,

categorize and process the experts’ information

effectively. A sufficient knowledge framework will

help the systems to find experts, with the keywords as

the research areas or related research areas. By that

means, the systems can suggest or recommend related

researchers. The knowledge base will make the

systems work more smartly, producing more accurate

and sufficient search results (Lin et al., 2017).

Currently, there are two primary ways to solve the

problem: manual and automatic taxonomy building.

https://orcid.org/0000-0001-6192-0212

https://orcid.org/0000-0002-3661-9142

Manual building is advantageous because it can

produce a solid taxonomy but requires experts in the

areas as well as considerable time and effort. The

knowledge in a specific field is vast and needs deep

research to have a good taxonomy. Building

automatic taxonomy is supplementary to the manual

method, which can use machine learning to mine the

data structure to create a quality taxonomy (Liu et al.,

2012; Song et al., 2015; Zhang et al., 2018).

Extensive expertise data can be created frequently

due to new research orientations and terminologies.

The directory can change over time, and most of the

data can only be understood by humans, not

computers. This gives rise to the appearance of the

semantics web and Ontology, where data and

knowledge are represented in structures that are

comprehensible to computers (Gomez-Perez &

Corcho, 2002).

This study uses background knowledge of

machine learning methods and textual data mining

techniques to build adaptive clustering, local

Le, T., Cao, T., Pham, L., Pham, T. and Luu, T.

An Automatic Method for Building a Taxonomy of Areas of Expertise.

DOI: 10.5220/0011630500003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 3, pages 169-176

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

169

embedding, and term ordering modules. By that

means, it is possible to construct an Ontology for a

domain via representation language and apply it to the

Ontology system of expert information. We proposed

a new method called TaxoGenDRK (Taxonomy

Generator using Database about Research Area and

Keyword) based on the method from Zhang et al.

(2018)’s research on TaxoGen and an additional

module that uses a database of research areas and

keywords retrieved from the internet – the data

regarded as an uncertain knowledge base for learning

about taxonomy.

However, the topic taxonomy for EFS has

received scant attention, and there has been little body

of research on this topic. The algorithm is used to

build a knowledge database on the research areas of

experts to serve the EFS. We experiment with the

DBLP dataset and aim to address two research

questions:

• Question 1. How is the performance of

TaxoGenDRK in creating the topic taxonomy in

Computer Science?

• Question 2. Is the TaxoGenDRK method

suitable for building the knowledge base for EFS?

To realize the research objectives, the study

focuses on presenting the theoretical basis and related

studies related to the problem, including machine

learning and text data processing theories. This is

followed by an overview of the solution adopted and

a detailed description of the modules used in the

method. Afterward, we discuss the tests of the method

performed on the input dataset and evaluate its

effectiveness. The research results draw conclusions

about what has been achieved and the remaining

limitations of the method used. In addition, some

directions for further research and research

recommendations are provided.

2 LITERATURE REVIEW

2.1 Ontology in Expert Finding

Systems

Ontologies, which are clear formal descriptions of the

concepts in the domain and their relationships

(Gruber, 1993), have recently moved from artificial

intelligence laboratories to the desktops of subject-

matter experts. Ontologies on the Web range from

large-scale classification methods for the

classification of websites to products for sale and

their characteristics (Noy & McGuinness, 2001).

Ontologies define a general vocabulary for

researchers who need to exchange information within

a domain (Çelik et al., 2013). Similarly, an ontology-

based method was proposed to find experts (Uddin et

al., 2011) in a certain field, primarily a research topic

in computer science and engineering.

There are a wide variety of methods that

concentrate on EFS, including Fine-Grained (Deng et

al., 2008), hybrid topic and language models (Deng et

al., 2008), and integrated evidence (Bogers et al.,

2008). Additionally, numerous methods involving

academic and social networks have been created for

locating specialists by the quantity of researchers.

However, they are based on text mining and

probabilistic approaches for performing expert

searches. Based on the collection of information,

ontologies used in expert search have been proposed

to improve the knowledge base model for academic

information (Bukowska et al., 2012).

2.2 Methods for Finding Experts

Numerous research has examined methods for

finding experts. According to their primary areas of

focus, the existing techniques can be grouped into

three major categories.

2.2.1 Content-based Method

Content-based methods have received much attention

from different studies. Text REtrieval Conference

(TREC) considers the first type of content-based

methodology to be an information retrieval task.

These methods fall into two categories: profile-

centric method and document-centric method (Balog

et al., 2012; Petkova & Croft, 2006). All documents

or texts related to a candidate are combined into a

single personal profile in the profile-centric method,

and the ranking score for each candidate is then

estimated based on the profile in response to a given

query (Balog et al., 2006). However, instead of

creating a single expertise profile, the document-

centric method analyzes the content of each document

separately (Balog et al., 2006; Wu et al., 2009). To

take advantage of the benefits of both the profile-

centric and document-centric methods, some existing

approaches have combined the two to improve

expert-finding performance (Petkova & Croft, 2008).

By using the term “vectors” with bag-of-words

representation, these studies typically focus on

matching search results to user queries, which differs

from topic-dependent expert finding based on

automatically inferred latent topics.

Topic modeling is the second type of content-

based method. To overcome the limitations of the

earlier topic model, a three-level hierarchical

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

170

Bayesian model called Latent Dirichlet allocation

(LDA) was proposed (Blei et al., 2003) to change the

problems to topic-vector based representation. To

extend the effort of the LDA model, an author-topic

model was introduced (Rosen-Zvi et al., 2004) to

illustrate the connection between the content of

documents and the interests of authors by sharing the

hyperparameter of all publications by the same

authors. As a result, topic models can support in

evaluating the contribution of candidates to an

inferred topic. They do not, however, analyze the

relationship between fields of research that represent

authors’ knowledge to build a network for more

sophisticated identification of experts.

2.2.2 Link Structure-based Method

To describe the direct evidence of candidates’

expertise, link structure-based method using

PageRank (Ding et al., 2009; Page et al., 1999) and

HITS (Kleinberg, 1999) algorithms were applied in

analyzing relationships in the scholarly network to

find authorized experts. AuthorRank, which is a

modification of PageRank with weights among the

co-authorship links, was also introduced to solve the

connection. Citation graph was used to evaluate the

impact of scientific journals and conferences,

publications, and academic authors. Several

modifications of traditional PageRank for

bibliographic networks were presented and achieved

a better result than the standard algorithm (Fiala et al.,

2008). PageRank with damping factors with weighted

algorithms that consider citation, co-authorship

topology and co-citation network was proposed to

measure author impact and rank them (Ding, 2011;

Page et al., 1999). All those efforts aim to improve

the accuracy of classical indicators such as impact

factor, citation count and H-index by applying

different modifications of HITS or PageRank

algorithms. However, they are ineffective for

identifying the top “experts” without focusing on

content features.

2.2.3 Combination of Content-based and

Link Structure-based Methods

Some researchers have employed link analysis to

further improve the ranking results after using

documents or snippet-level content to evaluate topic

relevance for each applicant. Text analysis and

network analysis were used after gathering all emails

particularly related to a topic and examining

connections between every pair of individuals, to sort

individuals and create “expert graph” (Campbell et

al., 2003). Then modified HITS was applied to find

ratings for all senders and recipients related to that

topic. Candidates’ personal profile and publications

were used to estimate an initial expert score for each

applicant and choose the top applicants to build a

subgraph (Zhang et al., 2007). All of them were

accomplished by operating through local

optimization while ranking among a small subset of

applicants. However, a topical factor graph model

was suggested to find representative nodes from

academic networks on a given topic by utilizing the

topic relevance and social ties between linkages

(Tang et al., 2009). In this research, we suggest an

improved method to identify the relationship between

topics for building topic tree-structured hierarchy,

which is called topic taxonomy.

2.3 Topic Taxonomy Construction

Method

A topic taxonomy is a tree-structured hierarchy in

which each node has a group of semantically related

terms that each reflect a particular conceptual topic

(Shang et al., 2020). Additionally, the topic-subtopic

relation should be followed by the parent-child nodes

in the hierarchical tree. For example, if a node has

children S = c

, c

, ..., c

, then each c

should be a sub-

topic of the node and have the same level of precision

as its siblings in S.

It is important to note that a term may be a sub-

topic of multiple conceptual topics and hence appear

in various nodes. For instance, “search engine” may

be a component of both “search engine in information

retrieval” and “image matching search engine”;

“neural networks” may be a component of both

“neural networks in deep learning” and “neural

networks for data mining.”

The main approaches for building taxonomy were

introduced as hyponymy-based methods, clustering-

based methods, network clustering-based methods.

3 METHODOLOGY

3.1 Research Process

From an input corpus about a field, we need to create

a topic taxonomy for that field. It is a data structure

of domain terms, namely a tree-structured hierarchy

of topics and terms within the area, in which each

node of the tree consists of a set of terms representing

a topic, and the child nodes are the sub-topics of the

parent node. The hierarchy is shown in the Figure 1.

This structure meets the requirements for building a

taxonomy of areas of expertise as each topic will

An Automatic Method for Building a Taxonomy of Areas of Expertise

171

usually be described by a group of synonymous terms

or terms related to the same content represented by

the group’s topic.

Figure 1: The topic taxonomy problem.

From the above research objectives and directions,

this study was implemented in the following stages:

1. A new method called TaxoGenDRK

(Taxonomy Generator using Database about

Research area and Keyword) was proposed in

an attempt to improve the topic taxonomy

method Taxogen. This is a method based on

TaxoGen research by Zhang et al. (Zhang et

al., 2018), together with an additional module

that uses a database of research areas and

keywords extracted from the internet – the

data regarded as an uncertain knowledge base

for learning about taxonomy. This dataset is a

data structure about research fields and

keywords built from crawled data from

Google Scholar

2. An experiment was conducted to build a topic

taxonomy, with “Computer Science” as the

main topic for testing, using Python language,

Google Colab Pro, and the DBLP dataset.

3. An evaluation of the topic taxonomy using

TaxogenDRK was conducted from qualitative

and quantitative approaches.

The TaxogenDRK method is the focal point of our

research.

3.2 Proposed Method: TaxoGenDRK

The overview model of the TaxoGenDRK method is

shown in Figure 2 below. The initial input and output

are the same as in the TaxoGen method. Two local

terms embedding, and adaptive clustering modules

are used for term clustering to create the output tree.

In addition, the method needs to perform additional

data retrieval from experts’ research areas and

scientific publications before building a dependency

graph from the extracted information, and finally, use

it to arrange terms based on a topic representation

https://scholar.google.com/

level. The two reused modules will not need to be re-

described, but the next section will describe the

further developed term ordering module (Figure 3).

Figure 2: An overview of the TaxoGenDRK method.

The term ordering module includes the tasks shown

in Figure 3, starting with the retrieval of data about

the experts’ research areas and keywords in their

scientific publications to the construction of a dataset

of research areas and related terms. The data set is in

the form of a dependency graph between the research

areas and keywords, all of which will be extracted

from information pages about experts in their fields

of research and scientific publications. Google

Scholar was used because it is the leading scientific

and expert information page used in the research

community. Finally, we combined the classified term

clusters to order the terms. The tasks will be described

in more detail in the following sections.

Figure 3: The stages in the term ordering module.

The next step is to build a dependency graph, a set

of graphs in which each graph has a node as a research

area, the remaining nodes are keywords related to that

research area, connected with the research area by a

weighted edge, and the weight is calculated by the

probability that the keyword is related to the area.

It describes the research fields and the keywords

in the field with a weight w

ra-kw

representing the

probability of the keyword in the field.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

172

𝑤



= 1 𝑜𝑟 𝑤



𝑙𝑒𝑛



𝑟𝑎



(1)

In (1), 𝑙𝑒𝑛



𝑟𝑎



is the number of research areas of the

expert whose scientific publications are crawled to

get the keywords. With two identical keywords

extracted from two different publications added to the

same topic, the weights will be updated by the higher

value between the two keyword weights.

This module has the function of sorting terms

according to a level called topic visibility, making it

easy to select the top terms of a cluster.

In this

method, we will use a score to evaluate the topic

representation of the node, and each node will be

illustrated by the term with the highest topic

representation and the top terms with the highest topic

representation.

The topic representation for each term for a

cluster is calculated using the following formula:

𝑠𝑐𝑜𝑟𝑒



𝑡,𝑆





= 𝑐𝑜𝑠𝑠𝑖𝑚



𝑡,𝑆





∗𝑟



𝑡,𝑆





(2)

In (2), 𝑐𝑜𝑠𝑠𝑖𝑚



𝑡,𝑆





is the similarity degree

cosine of the embedding vector of term t with the

center of cluster S

, and 𝑟



𝑡,𝑆





is the degree of

representation of the term t for the cluster S

(as

included in the parent node of the cluster S

during

clustering of that parent node).

In the step of ordering terms, the clusters

containing the term is the area of research, and the

topic representation score of the research area will be

recalculated based on the association graph above as

follows:

𝑛𝑒𝑤𝑠𝑐𝑜𝑟𝑒



𝑡,𝑆





=𝑠𝑐𝑜𝑟𝑒



𝑡,𝑆





+𝑠𝑐𝑜𝑟𝑒



𝑘𝑤,𝑆





∗𝑤



 ∈ 

(3)

In (3), T represents the set of terms that are keywords

and belong to the dataset of research fields associated

with the keywords built above.

3.3 Experiments

3.3.1 Dataset

The test dataset is the DBLP, a commonly used data

set in some other topic taxonomy studies. Zhang et al.

(2018) published preprocessed DBLP dataset, and

this study used the above dataset for testing and

evaluation. The above dataset contains a corpus with

more than 1.8 million titles of scientific articles in

computer science fields.

Because the topic taxonomy requires a large

corpus, the effectiveness of the experimental setup is

required for its efficiency in optimizing memory

space and processing time. The preprocessing steps

are used as much as possible as they are performed

only once compared to the next repeated steps of the

algorithm. In the preprocessing step, the tool NPC

(Noun Phrase Chunking) is used to extract noun

phrases from scientific article titles to form a term set

after removing duplicate words and selecting the most

common terms to form a collection of more than

thirteen thousand terms. In addition, the next

preprocessing steps include indexing the corpus,

calculating term frequency on each corpus, learning

global embedding vectors, calculating document

lengths, and storing the index of documents where the

term appears.

3.3.2 Parameters

The study is implemented in Python language version

3.6, and is run on Google Colab Pro. The parameters

used for the implementation of the method are as

follows: n_cluster: 5 (the number of clusters

categorized at each level); max_depth: 4 (the

maximum depth of the taxonomy tree);

filter_threshold: 0.25 (the threshold for generic term

removal in adaptive clustering module);

n_cluster_iter: 2 (the maximum number of iterations

of adaptive clustering); n_expand: 100 (the number

of extended terms, counting from the central term

during local embedding); n_include: 10 (the number

of terms close to the cluster center chosen to extract

the secondary corpus during local embedding);

n_top: 10 (the number of terms with the highest

subject representation selected during the term

ordering process).

The study combines the data structure of the

research areas and the keywords crawled from

Google Scholar. To get the dataset, conduct a query

to crawl the data of more than fifty experts in the field

of computer science, with a number of leading

research citations in several research areas, such as

machine learning, computer vision, computer

security, and natural language processing. The data

structure is as follows: Number of research fields:

34 (research fields from the information of experts);

Number of keywords: 213 (crawled keywords from

the publication description pages); Number of links:

759 (the link between research areas and keywords);

Number of reliable links: 113 (link weighted from

0.5 or more); Average number of links: 293 (links

weighted greater than 0.2 and less than 0.5); Number

of less reliable links: 353 (link weighted 0.2 or less).

An Automatic Method for Building a Taxonomy of Areas of Expertise

173

4 RESEARCH RESULT

4.1 Taxonomy Results

The result of building topic classification on the

DBLP dataset, with the main topic being “computer

science,” the topic taxonomy tree is built in 4 levels,

and each parent topic is divided into 5 sub-topics.

Figure 4 shows a part of the topic taxonomy results,

where the sub-topics are represented by the five terms

with the highest topic representation and are labeled

with the words of the highest topic representation.

Figure 4: TaxoGenDRK results for classifying Computer

Science topics in some clusters.

4.2 Evaluation

4.2.1 Qualitative Evaluation

In the results of the topic taxonomy in the field of

“computer science,” not all taxonomies give reliable

results. Since the number of sub-clusters is fixed at 5,

some clusters do not really belong to their parent

cluster, although the labels and top terms are still

quite semantically clear about the same topic. For

instance, the cluster of natural language processing

can be seen in Figure 5, which was founded during

the clustering process of the topic “information

retrieval.” Since we fixed the number of clusters of

the computer science root node to be 5, the terms on

the topic of natural language processing must belong

to one of those 5 clusters, and the cluster whose center

is close to the term set of the most natural language

processing is the information extraction cluster, so it

is classified as a child of the information extraction

node.

In addition, although the process of adaptive

clustering and local embedding helps the clusters to

separate clearly, eliminating common terms in the

sub-clusters, it is not completely effective for the

whole result. In some clusters, after being split, one

of the sub-clusters is almost still at the center of the

parent cluster, as can be seen in the “search results”

cluster at level two in where there is a subcluster that

seems to be the center of the parent cluster, and they

all represent the same topic. When splitting the above

cluster into sub-clusters, there is a cluster located at

the center of the parent cluster.

Figure 5: TaxoGenDRK results for the cluster “information

retrieval”.

4.2.2 Quantitative Evaluation

The quantitative evaluation of a topic taxonomy tree

is not an easy task to implement. The researchers

made references to related studies, as in (J. Shang,

2020; Zhang et al., 2018), to learn about how to build

an assessment of the topic taxonomy tree and to

provide quantitative measures to evaluate the results.

In general, the related studies all constructed a set of

evaluation indicators, then used humans as the main

agent to evaluate each selected example, and finally

arrived at a quantitative evaluation. From the

evaluation indicators, we can compare the results

with each other even at a relative level because, in

different studies, the main agent to evaluate (humans)

are different. The indicators used in this study are:

Relation accuracy: the accuracy between

relationships that measures the accuracy of the

relationship between parent and child nodes in a

classification tree; Exclusive sibling: the semantic

separation between sibling nodes in the same parent

topic; Term coherence: the coherence of terms that

measures how well the top terms represent the same

topic

In all the above criteria, the data points of the

results are given to a team of three engineers working

in the field of computer science, who will learn about

unknown terms and use their knowledge of known

terms to evaluate each result. The majority’s result

was selected as the final result for each data point. In

machine learning taxonomy, the commonly used

metrics are Precision, Recall, and F-score. However,

in the topic taxonomy problem, each data point will

only have True Positive and False Positive, so only

the Precision measure will be used for the three

evaluation indicators above. The specific formula for

the measurements is as follows:

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

174

Precision:

(4)

In (4), FP is the number of False Positives (number of

false predictions), TP is the number of True Positives

(number of correct predictions), PP is the number of

Predicted Positives (total number of predictions), and

RP is the number of Real Positives (the total number

of correct labels).

The specific results are shown in Table 1. The

first column is the name of the methods, and the next

columns are respectively the evaluation indicators for

the methods, and the blank boxes are the results not

evaluated. The methods are: TaxoGenDRK: the

proposed research method; TaxoGen: the method

that does not use the term sorting module but only

uses the original representative points to generate the

top terms; NoAC: the method does not use an

adaptive clustering module, but uses algorithms to

cluster without removing generic terms; NoLE: the

method does not use local embedding modules but

uses global embedding vectors in the clustering steps;

HLDA: the non-parametric topic hierarchical model

tested to build a topic taxonomy tree; Hclus: building

clustering using global embedding vectors and using

only K-means spherical clustering algorithm to build

a hierarchical taxonomy tree (Zhang et al., 2018).

Table 1: Results of the topic taxonomy tree evaluation.

Method

Relation

Accuracy

Exclusive

Sibling

Term

Coherence

TaxoGenDR

0.79 0.76 0.82

TaxoGen 0.77 0.73 0.84

HClus 0.44 0.47 0.62

NoLE 0.64 0.70 -

HLDA 0.27 0.44 -

NoAC 0.56 0.35 -

These results show the effectiveness of the model

in building the taxonomy tree. All three modules of

adaptive clustering, local embedding, and term

sorting provide good indicators, and the results are

consistent with the qualitative evaluation in the

previous section. The last module is to order the

terms, which seems to show only a small effect

because it does not affect the process of creating

clusters, but when the terms have been hierarchically

clustered, it orders them to get the position of the top

terms. The results also showed the efficiency of

TaxoGenDRK compared to TaxoGen in the

evaluation index of the relationships between parent

and child nodes as well as the separation between

sibling clusters.

5 CONCLUSION

The research process on building an automatic

ontology of expertise has brought insights into how to

represent knowledge in the current web systems, the

approaches, and deployment of machine learning

methods, including adaptive clustering module, local

embedding module, and term ordering module to

solve the problem of building topic taxonomy. The

proposed method helped construct a topic taxonomy

in the field of “computer science” with a relatively

good degree of accuracy compared with the existing

studies. Thus, research question 1 has been resolved.

The proposed TaxoGenDRK model, however, has

given relatively reliable results compared with

existing studies on the problem of automatically

building topic taxonomies. However, it still has some

limitations that must be acknowledged and improved.

Some parameters used seem to significantly influence

the results, such as the number of clusters in each

topic, the threshold to remove generic terms, and so

on.

In research question 2, It can be seen that the

TaxoGenDRK method is quite suitable for building

the knowledge base for EFS. However, the extracted

data about the information of research fields and

keywords in expert publications are only uncertain

knowledge bases. As a result, sometimes, the details

learned are not accurate. Further studies can focus on

parameters suitable for different steps or different

inputs. In addition, the input corpus also dramatically

affects the results of classification construction, so it

is necessary to do more research on building a

generalized input for a topic with complete and

relevant contexts for building a taxonomy. Another

approach is to create a data set on the research field

and keywords in the area with higher reliability.

Currently, this study only focuses on the field of

computer science with DBLP datasets. Further

research can be extended to other fields to build

knowledge representations for scientists in diverse

research areas.

ACKNOWLEDGEMENTS

The research team would like to thank the Innovation

Fund (VinIF), Big Data Research Institute (VinBDI)

for financial support to carry out this study.

An Automatic Method for Building a Taxonomy of Areas of Expertise

175

REFERENCES

Abramowicz, W., Bukowska, E., Dzikowski, J.,

Filipowska, A., & Kaczmarek, M. (2011). Semantically

Enabled Experts Finding System - Ontologies,

Reasoning Approach and Web Interface Design.

ADBIS.

Al-Taie, M. Z., Kadry, S., & Obasa, A. (2018).

Understanding expert finding systems: domains and

techniques. Social Network Analysis and Mining, 8(1).

Balog, K., Azzopardi, L., & Rijke, M. d. (2006). Formal

models for expert finding in enterprise corpora

Proceedings of the 29th annual international ACM

SIGIR, Seattle, Washington, USA.

Balog, K., Fang, Y., Rijke, M. d., Serdyukov, P., & Si, L.

(2012). Expertise Retrieval. Found. Trends Inf. Retr.,

6(2–3), 127–256.

Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent

dirichlet allocation. 3(J. Mach. Learn. Res.), 993–1022.

Bogers, T., Kox, K., & van den Bosch, A. (2008). Using

Citation Analysis for Finding Experts in Workgroups.

Bukowska, E., Kaczmarek, M., Stolarski, P., &

Abramowicz, W. (2012). Ontology-Based Retrieval of

Experts – The Issue of Efficiency and Scalability within

the eXtraSpec System. Multidisciplinary Research and

Practice for Information Systems, Berlin, Heidelberg.

Campbell, C. S., Maglio, P. P., Cozzi, A., & Dom, B.

(2003). Expertise identification using email

communications Proceedings of the twelfth

international conference on Information and knowledge

management, New Orleans, LA, USA.

Çelik, D., Karakas, A., Bal, G., Gultunca, C., Elçi, A.,

Buluz, B., & Alevli, M. C. (2013). Towards an

Information Extraction System Based on Ontology to

Match Resumes and Jobs. 2013 IEEE 37th Annual

Computer Software and Applications Conference

Workshops, 333-338.

Deng, H., King, I., & Lyu, M. R. (2008). Formal Models

for Expert Finding on DBLP Bibliography Data. 2008

Eighth IEEE International Conference on Data Mining,

163-172.

Ding, Y. (2011). Topic-based PageRank on author

cocitation networks. JASIST, 62(3), 449-466.

Ding, Y., Yan, E., Frazho, A., & Caverlee, J. (2009).

PageRank for ranking authors in co-citation networks.

Journal of the American Society for Information

Science and Technology, 60(11), 2229-2243.

Fiala, D., Rousselot, F., & Ježek, K. (2008). PageRank for

bibliographic networks. Scientometrics, 76(1), 135-

158.

Gomez-Perez, A., & Corcho, O. (2002). Ontology

languages for the Semantic Web. Intelligent Systems,

IEEE, 17, 54-60.

Gruber, T. R. (1993). A translation approach to portable

ontology specifications. Knowledge Acquisition, 5,

199-220.

Husain, O., Salim, N., Alias, R. A., Abdelsalam, S., &

Hassan, A. (2019). Expert Finding Systems: A

Systematic Review. Applied Sciences, 9(20).

J. Shang, X. Z., L. Liu and S. Li. (2020). NetTaxo:

Automated Topic Taxonomy Construction from Text-

Rich Network. Proceedings of The Web Conference.

Kleinberg, J. M. (1999). Authoritative sources in a

hyperlinked environment. J. ACM, 46, 604-632.

Lin, S., Hong, W., Wang, D., & Li, T. (2017). A survey on

expert finding techniques. J. Intell. Inf. Syst., 49(2),

255–279.

Liu, X., Song, Y., Liu, S., & Wang, H. (2012). Automatic

taxonomy construction from keywords. Knowledge

Discovery and Data Mining.

Noy, N., & McGuinness, D. (2001). Ontology

Development 101: A Guide to Creating Your First

Ontology. Knowledge Systems Laboratory, 32.

Page, L., Brin, S., Motwani, R., & Winograd, T. (1999).

The PageRank Citation Ranking : Bringing Order to the

Web. The Web Conference.

Petkova, D., & Croft, W. (2008). Hierarchical Language

Models for Expert Finding in Enterprise Corpora.

International Journal on Artificial Intelligence Tools,

17, 5-18.

Petkova, D., & Croft, W. B. (2006). Hierarchical Language

Models for Expert Finding in Enterprise Corpora.

ICTAI'06,

Rosen-Zvi, M., Griffiths, T., Steyvers, M., & Smyth, P.

(2004). The author-topic model for authors and

documents Proceedings of the 20th conference on

Uncertainty in artificial intelligence, Banff, Canada.

Shang, J., Zhang, X., Liu, L., Li, S., & Han, J. (2020).

NetTaxo: Automated Topic Taxonomy Construction

from Text-Rich Network Proceedings of The Web

Conference 2020, Taipei, Taiwan.

Song, Y., Liu, S., Liu, X., & Wang, H. (2015). Automatic

Taxonomy Construction from Keywords via Scalable

Bayesian Rose Trees. IEEE Transactions on

Knowledge and Data Engineering, 27(7), 1861-1874.

Tang, J., Sun, J., Wang, C., & Yang, Z. (2009). Social

influence analysis in large-scale networks Proceedings

of the 15th ACM SIGKDD international conference on

Knowledge discovery and data mining, Paris, France.

Uddin, M. N., Duong, T. H., Oh, K.-j., & Jo, G.-S. (2011).

An Ontology Based Model for Experts Search and

Ranking. Intelligent Information and Database

Systems, Berlin, Heidelberg.

Wu, H., Pei, Y., & Yu, J. (2009). Hidden Topic Analysis

Based Formal Framework for Finding Experts in

Metadata Corpus. 2009 Eighth IEEE/ACIS

International Conference on Computer and Information

Science.

Zhang, C., Tao, F., Chen, X., Shen, J., Jiang, M., Sadler, B.

M., Vanni, M. T., & Han, J. (2018). TaxoGen:

Unsupervised Topic Taxonomy Construction by

Adaptive Term Embedding and Clustering. Proceedings

of the 24th ACM SIGKDD International Conference on

Knowledge Discovery & Data Mining, London, UK.

Zhang, J., Tang, J., & Li, J. (2007). Expert Finding in a

Social Network. In R. Kotagiri, P. R. Krishna, M.

Mohania, & E. Nantajeewarawat, Advances in

Databases: Concepts, Systems and Applications.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

176