Conference Management System Utilizing an LLM-Based

Recommendation System for the Reviewer Assignment Problem

Vaios Stergiopoulos

1 a

, Michael Vassilakopoulos

1 b

, Eleni Tousidou

1 c

, Spyridon Kavvathas

and

Antonio Corral

2 d

Data Structuring & Eng. Lab., Dept. of Electrical & Computer Engineering, University of Thessaly, Volos, Greece

Dept. of Informatics, University of Almeria, Spain

Keywords:

Conference Management System, Large Language Models, Recommendation System, Paper-Reviewer

Matching, Reviewer Assignment Problem.

Abstract:

One of the most important tasks of a conference organizer is to assign reviewers to papers. The peer review

process of the submitted papers is a crucial step in determining the conference agenda, quality, and success.

However, this is not an easy task; large conferences often assign hundreds of papers to hundreds of reviewers,

making it impossible for a single person to complete the task due to hard time constraints. We propose

a Conference Management System that embodies a Large Language Model (LLM) in its core. The LLM

is utilized as a Recommendation System which applies Content-based Filtering and automates the task of

reviewers-to-papers assignment for a conference. The LLM we select to use is the Bidirectional Encoder

Representations from Transformers (BERT), in two speciﬁc variants, BERT-tiny and BERT-large.

1 INTRODUCTION

Today, the number of papers submitted to academic

conferences is continuously increasing, resulting in

a phenomenon that is extremely intense in sev-

eral prominent ones. For example, the number of

manuscripts received by NeurIPS and AAAI confer-

ences in 2020 was more than 5 times that in 2014

(Zhao and Zhang, 2022).

The assignment of the submitted papers (pro-

posals) to the most scientiﬁcally qualiﬁed reviewers,

known as the Reviewer Assignment Problem (RAP),

remains a very crucial responsibility of conference or-

ganizers to this day, as it is an essential component of

academic integrity and excellence. The quality and

reputation of a conference depend greatly on the re-

ceipt of high-quality reviews. The most prestigious

conferences usually have to assign thousands of pa-

pers to thousands of reviewers, and, with the added

time constraint, this becomes a challenging operation

to be handled by a single person or a small group of

people. Therefore, systems managing the full confer-

https://orcid.org/0000-0003-3458-1491

https://orcid.org/0000-0003-2256-5523

https://orcid.org/0000-0002-1581-0642

https://orcid.org/0000-0002-0069-4642

ence procedures and assigning reviewers to articles

automatically are being increasingly utilized to man-

age the above situation. In addition, such systems re-

ceive a lot of attention from researchers around the

world (Zhao and Zhang, 2022) (Aksoy et al., 2023)

(Ribeiro et al., 2023) to manage the growth in the

number of articles and solve the reviewer assignment

problem.

On the other hand, recent research has put a lot

of focus on Large Language Models (LLMs), or Pre-

trained Language Models (PLMs); they are large-

scale, pretrained language models based on neural

networks that have been trained with a vast amount

of generic textual datasets. Furthermore, recent ad-

vances on transformer-based LLMs (Vaswani et al.,

2017) speciﬁcally have provided state-of-the-art per-

formance to a wide variety of tasks and applications,

ranging from chatbots to summarizing text and trans-

lating. Here, we accomplish to utilize the power of

LLMs in Recommendation Systems. More speciﬁ-

cally, we embody a LLM in a recommendation engine

that deals with the RAP problem inside a web-based

Conference Management System.

In this paper, we present our system in the above-

mentioned framework:

• A web-based application which basically is a full

1004

Stergiopoulos, V., Vassilakopoulos, M., Tousidou, E., Kavvathas, S. and Corral, A.

Conference Management System Utilizing an LLM-Based Recommendation System for the Reviewer Assignment Problem.

DOI: 10.5220/0013482600003929

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 1004-1011

ISBN: 978-989-758-749-8; ISSN: 2184-4992

Conference Management System that can auto-

mate and complete all the sub-tasks needed by the

conference organizers.

• A solution for the classic Reviewer Assignment

Problem (RAP).

• An application of LLMs in Recommendation Sys-

tems (RS) as a solution for RAP.

The remainder of this paper is organized in the fol-

lowing order. Firstly, all related work is presented in

Section 2. Secondly, the Recommendation System is

explained in Section 3. Next, our web Conference

Management System is demonstrated in Section 4.

Lastly, in Section 5 we conclude this work and dis-

cuss related future work directions.

2 RELATED WORK

In this section, we ﬁrst review related work on LLMs

and Recommendation Systems and we next review

work on Conference Management Systems and the re-

viewer assignment problem.

2.1 Large Language Models and

Recommendation Systems

Vaswani et al. (Vaswani et al., 2017) propose a spe-

ciﬁc kind of deep neural network called Transformer,

based solely on attention mechanisms, dispensing

with recurrence and convolutions entirely. Their work

was initially suggested for translation tasks, but later

proved to be a huge advancement for LLMs in gen-

eral. Their suggested network architecture comprises

encoder and decoder stacks (each one of them con-

sisting of 6 identical layers) and is the base for all

state-of-the-art LLMs today.

The BERT LLM, proposed by Devlin et al. (De-

vlin et al., 2019), which stands for Bidirectional

Encoder Representations from Transformers, is de-

signed to pre-train deep bidirectional representations

by jointly conditioning on both left and right con-

text in all layers (process all tokens before and after

at each tokenized word). The BERT model architec-

ture is a multilayer bidirectional transformer encoder

which computes vector-space representations of natu-

ral language that are suitable for use in deep learning

models. BERT models are usually pre-trained on a

large corpus of text, then ﬁne-tuned for speciﬁc tasks.

It has been successful in a variety of tasks in NLP

(natural language processing). Also, the pre-trained

BERT representations can be ﬁne-tuned with just one

additional output layer to create state-of-the-art mod-

els for a wide range of tasks, such as question answer-

ing and language inference, without substantial task-

speciﬁc architecture modiﬁcations.

Moreover, Sun et al. (Sun et al., 2019) intro-

duce a new sequential recommendation model, which

adopts BERT to a new task, sequential recommenda-

tion. Thus, the BERT LLM is embodied in a Recom-

mendation System. Their system, called BERT4Rec,

employs the deep bidirectional self-attention to model

user behavior sequences and predicts the random

masked items in the sequence by jointly condition-

ing on their left and right context. They train a bidi-

rectional representation model to make recommenda-

tions by allowing each item in user historical behav-

iors to fuse information from both left and right sides.

Wu et al. (Wu et al., 2021) present their efforts to

strengthen a news Recommendation System by uti-

lizing pre-trained language models (PLMs). Results

from ofﬂine experiments on datasets for monolingual

and multilingual news recommendation demonstrate

that using PLMs for news modeling can signiﬁcantly

enhance news recommendation performance. After

being implemented on the Microsoft News platform,

the PLM-powered news recommendation models saw

notable increases in clicks and pageviews in both En-

glish and international markets. So, in their frame-

work, they instantiate a news encoder PLM to cap-

ture the deep contexts in news texts. They denote an

input news text with a number of tokens. Next, the

PLM converts each token into its embedding, and then

learns the hidden representations of words through

several Transformer layers; this way a hidden token

representation sequence is produced. Moreover, they

use an attention network to summarize the hidden to-

ken representations into a uniﬁed news embedding

which is further used for user modeling and candidate

matching.

2.2 Conference Management Systems

and the Reviewer Assignment

Problem

The work of Charlin & Zemel (Charlin and

Zemel, 2013), the "Toronto Paper Matching System"

(TPMS), is one of the ﬁrst high-performance attempts

to create a reviewers assignment system that would

automate this task for a prominent conference. They

use Latent Dirichlet Allocation (LDA), an unsuper-

vised probabilistic method used to model documents,

and a matching score that predicts a reviewer’s score

as the dot product between a reviewer’s proﬁle rep-

resentation (archive data of her/his papers) and each

submission.

Furthermore, Maleszka et al. (Maleszka et al.,

2020) present the idea of an overall modular system

Conference Management System Utilizing an LLM-Based Recommendation System for the Reviewer Assignment Problem

1005

for determining a grouping of reviewers, as well as

three modules for such a system: a keyword-based

module, a social graph module, and a linguistic mod-

ule. They start from a single reviewer and look for a

diverse group of other possible candidates that would

complement the ﬁrst one in order to cover multiple

areas of the review.

Additionally, Stelmakh et al. (Stelmakh et al.,

2023) provide publicly a gold-standard dataset that is

needed to perform reproducible research on the pro-

cess of assigning submissions to reviewers. Their

dataset consists of 477 self-reported expertise scores

provided by 58 researchers (reviewer candidates) who

evaluated their expertise in reviewing papers they had

previously read. Moreover, they use this data to com-

pare several popular algorithms currently employed in

conferences and present the respective performance

metrics. This currently expanding dataset could be

the standard dataset for algorithmic comparison in the

area of RAP task.

The study of Bouanane et al. (Bouanane et al.,

2024) introduces the Balanced and Fair Reviewer As-

signment Problem (BFRAP), which aims to maximize

the overall similarity score (efﬁciency) and the mini-

mum paper score (fairness) subject to coverage, load

balance, and fairness constraints. They conduct a the-

oretical investigation into the threshold conditions for

the problem’s feasibility and optimality. To facili-

tate this investigation, they establish a connection be-

tween BFRAP, deﬁned over m reviewers, and the Eq-

uitable m-Coloring Problem. Building on this theoret-

ical foundation, they propose FairColor, an algorithm

designed to retrieve fair and efﬁcient assignments.

Leyton-brown et al. (Leyton-Brown et al., 2024)

in their paper introduce Large Conference Matching

(LCM), a novel reviewer–paper matching approach

that was recently deployed in the 35th AAAI Con-

ference on Artiﬁcial Intelligence (AAAI 2021), and

has since been adopted by other conferences. LCM

consists of three primary components:

1. Gathering and managing input information to pin-

point issues with matches and produce scores for

reviewers and papers;

2. Creating and resolving an optimization challenge

to identify optimal reviewer-paper pairings; and

3. A two-phase evaluation process that reallocates

review resources from submissions that are prob-

able to be declined and toward documents closer

to the decision limit.

Zhang et al. (Zhang et al., 2024) propose a uni-

ﬁed model for paper-reviewer matching that jointly

considers semantic, topic, and citation factors. To

be speciﬁc, during training, they instruction-tune a

contextualized language model shared across all fac-

tors to capture their commonalities and characteris-

tics; during inference, they chain the three factors to

enable step-by-step, coarse-to-ﬁne search for quali-

ﬁed reviewers given a submission.

Latypova (Latypova, 2023) proposes a method of

reviewer assignment decision support in an academic

journal based on a joint use of multicriteria assess-

ment and text mining. Calculation of an integral indi-

cator with the use of additive folding of weighted re-

viewer’s indicators is at the core of the method. Text

mining of manuscripts and reviewer’s papers is uti-

lized to determine value of one of signiﬁcant indica-

tors. The proposed method allows to assess reviewers

not only by authority and expertise, but also allows to

take into account their work in the role of a reviewer,

deciding how good they are in this role.

3 THE RECOMMENDATION

ENGINE

In this section, we describe the functional architec-

ture of our recommendation engine as the core part of

the full Conference Management System. The recom-

mendation engine works smoothly with the rest of the

components of the system, while various procedures

are executed, e.g. conference setup, reviewers input,

paper submission, etc.

The recommendation engine periodically checks

the system’s database for new conferences. As soon

as a new conference is created and the user (confer-

ence organizer) characterizes it as ﬁnalized, meaning

that the submitted papers, as well as the list of the re-

viewers, have been inserted into it, the recommenda-

tion engine starts processing the speciﬁc conference.

To begin with, the Reviewer Module retrieves

from the database the reviewers’ data, i.e. full name,

email and university/organization. In case a reviewer

already has a proﬁle in the database, created at a

previous conference, the module just loads this pro-

ﬁle. If a reviewer does not have a proﬁle in our

database, the module creates a new scientiﬁc proﬁle

by collecting text from all the papers she/he has co-

authored, using data from the AMiner Citation Net-

work Dataset. We have pre-processed the AMiner Ci-

tation Network Dataset in order to easily retrieve an

author and her/his relevant papers. Moreover, all pa-

pers’ text has been cleaned and pre-processed, to ac-

celerate procedures and make them more efﬁcient.

Additionally, the Reviewer Module utilizes the

"edit distance measure", also known as the general-

ized Levenshtein distance, presented in Li et al. (Yu-

jian and Bo, 2007), to calculate the differences be-

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

1006

tween sequences and patterns. The Levenshtein dis-

tance, d(x,y), computes the minimal cost of trans-

forming string x to string y. The transformation of a

string is carried out using a sequence of the following

operators: delete a character, insert a character, and

substitute one character for another. We use the Lev-

enshtein distance in order to compare the reviewers’

names with the authors’ names of the AMiner Cita-

tion Network Dataset and retrieve the right authors’

proﬁles. Finally, the reviewers module provides the

BERT LLM part with the reviewers’ scientiﬁc pro-

ﬁles.

Furthermore, there is the Papers Module which

processes the submitted papers and their data. It re-

trieves all submitted papers from every single confer-

ence in the database and executes the necessary text

pre-processing and cleaning, implementing the meth-

ods described by Stergiopoulos et al. (Stergiopoulos

et al., 2022). Next, the papers module outputs the pre-

processed text (title, abstract) of the papers into the

BERT LLM part.

In the core of our recommendation engine (Fig-

ure 1) there is the BERT LLM (Devlin et al., 2019)

which computes the vector-space representations for

both the reviewers (text from the researcher’s proﬁle)

and the submitted papers (text from title and abstract).

To be more speciﬁc, we run two sets of experi-

ments with two BERT variants:

1. BERT-tiny: a BERT model with 2 layers, 128 hid-

den unit size, and 2 attention heads (L-2, H-128,

A-2), pre-trained on uncased text.

2. BERT-large: a BERT model with 24 layers, 1024

hidden unit size, and 16 attention heads (L-24, H-

1024, A-16), pre-trained on uncased text.

The BERT LLM (Figure 1) usually comprises

two processing layers: the preprocessor and the

transformer-encoder layer.

The preprocessor layer receives as input, simple

natural language text. It tokenizes, formats and packs

input sentences. The result of the preprocessing layer

is a batch of ﬁxed-length input sequences of tokens

for the transformer encoder. An input token-sequence

starts with one start-of-sequence token, followed by

the tokenized segments, each terminated by one end-

of-segment token. Also, it provides the token dictio-

nary to the next layer.

Next, the output of the preprocessor layer is in-

serted into the transformer-encoder layer. It doesn’t

just take the tokenized strings as input, but it also ex-

pects these to be packed into a particular format. The

transformer-encoder layer creates and outputs the ac-

tual vector-space representations for both the review-

ers and the submitted papers, so that the Reviewers

Figure 1: The functional architecture of our Recommenda-

tion Engine.

Assignment Module - RAM (Figure 1) can use them

to calculate their similarity.

The RAM in Figure 2, incorporates a number of

assumptions and constraints as listed below:

• Each paper must be reviewed by a speciﬁc number

of reviewers; the reviewers are assigned a maxi-

mum number of papers, usually three, and the pa-

pers are assigned to three reviewers, too. This is a

variable that could change depending on the user’s

needs.

• Reviewers have a limit on the number of papers

they can review.

• The authors of submitted papers should not be re-

viewers at the same moment; a mechanism that

increases the integrity of the peer review proce-

dure, as the reviewers that are authors of a paper

in a conference are removed from the reviewers

list of the speciﬁc conference.

• Assign the most suitable reviewers, depending on

their expertise, for each paper while meeting the

aforementioned constraints.

Our recommendation engine actually implements

Content-based Filtering (CBF), i.e. it calculates the

similarity between a reviewer and a paper, trying to

Conference Management System Utilizing an LLM-Based Recommendation System for the Reviewer Assignment Problem

1007

Figure 2: The BERT LLM in the core of our recommenda-

tion Engine.

predict the expertise of a reviewer for a paper. RAM

uses the embeddings of both reviewers and papers and

calculates the similarity between them. Actually, co-

sine similarity is being used as it performs really well

in text similarity tasks. Cosine similarity is a measure

of similarity between two vectors in an inner product

space. It determines the degree to which two vectors

are pointing in the same direction by calculating the

cosine of the angle between them. Cosine similarity

is commonly used in text analysis to measure the sim-

ilarity between documents based on the frequency of

words or phrases they contain. So, if this distance is

small, there will be a high degree of similarity, but

when the distance is large, there will be a low degree

of similarity. Our aim is, via cosine similarity, to de-

termine (predict) a reviewer’s expertise on a paper.

Later on, the RAM recommends (assigns) review-

ers to papers in a high-degree-of-similarity-goes-ﬁrst

logic. Speciﬁcally, the ﬁrst paper (that is being pro-

cessed) is assigned to the three reviewers who are

found to have the highest similarity to it, then the

second paper, and so on. Finally, it outputs the re-

sulting reviewer-to-paper recommendations (assign-

ments) into the database of the system, so that the user

(conference organizer) can check it, make any adjust-

ments or modiﬁcations and approve it. As soon as the

reviewers recommendation is approved, the review-

ers are being informed of the papers that have been

assigned to them and they can start the peer review

process.

4 THE WEB SYSTEM

4.1 Technologies

In this section, the design of the system is presented

in order to understand in detail how the system oper-

ates. As shown in Figure 3 the system clearly consists

of four different entities. Firstly the front-end, which

includes the user interface, the client side routing and

the integration with the external service; secondly the

back-end, which includes the routing, the middleware

and the data handling; thirdly the database, which

provides and receives data from the back-end; last but

not least, the Recommendation System, which com-

municates directly with the database to retrieve papers

and reviewers data, or send the recommendations (as-

signments).

The choice of web application type was meticu-

lously determined by prioritizing factors such as user

experience, application speed, and overall ﬂuidity.

Crucial considerations were given to seamless inte-

gration between front-end, back-end, and database

components. After comprehensive evaluation, it be-

came evident that the Single Page Application (SPA)

model best aligned with the objectives of our system.

Leveraging the React.js framework for front-end im-

plementation further reinforced our commitment to

delivering an intuitive and responsive user interface.

The term back-end refers to the server-side of a

web application responsible for handling data pro-

cessing, logic, and database management. It acts

as the backbone of the system, facilitating seam-

less communication between the front-end and the

database. Here, for the web-based Recommenda-

tion System for scientiﬁc conferences, we employed

a combination of powerful technologies to ensure ro-

bustness, scalability, and efﬁciency. The technologies

utilized in the back-end include Node.js, Express.js,

SQL, and MySQL.

4.2 Requirements Analysis

It is crucial to outline the various capabilities (system

requirements) that users have when using this web-

based Recommendation System for scientiﬁc confer-

ences. Below we present an analytical list of the sys-

tem requirements:

1. Users can create a new account using the registra-

tion page.

2. Users can log in to the application using the cre-

dentials provided during the registration process.

3. Users have the option to log out of the application

at any time after registering or logging in.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

1008

Figure 3: The architecture of the system.

4. Once logged in, users can navigate through the ap-

plication’s features and functionalities.

5. Users can create new scientiﬁc conferences, pro-

viding all the necessary information required for

a scientiﬁc conference.

6. Users can view the conferences that have been

created.

7. Users have the ability to add scientiﬁc papers in

multiple ways:

• Add papers one by one, providing only the title

and abstract.

• Upload papers individually, where the system

extracts the title and abstract automatically.

• Import papers via Excel or CSV ﬁles, allowing

multiple papers to be added simultaneously.

8. Users also have the ability to add reviewers for

their conferences using various methods:

• Add reviewers one by one, providing their

email and name.

• Upload reviewers individually, with the system

extracting details from the provided informa-

tion.

• Import reviewers through Excel or CSV ﬁles,

facilitating the addition of multiple reviewers at

once.

9. Users can automatically create the reviewers-to-

papers-assignment.

4.3 Graphical User Interface

Following, we present the Graphical User Interface

(GUI) of this system.

Initially, the potential user has to create an account

(register) in the Home page, as shown in Figure 4, to

be able to use the system.

After successful registration and login, the user

can observe the options available within the applica-

tion (Figure 5). So now, she/he can initiate a new

conference by clicking the Create button, review and

update conferences she/he has already created via the

View button, and participate in other conferences by

selecting the Join button (e.g. if the user is a reviewer

for one or more conferences).

The management of conferences (i.e. create a new

conference or update an existing one) can be accom-

plished via the options in the Conference page (Fig-

ure 6). There, the user can insert proposals (submitted

papers) or reviewers; these procedures can be done ei-

ther one by one (single paper or reviewer) or via bulk

input using csv ﬁles.

Finally, the system provides the assignments of re-

viewers to papers, as can be seen in Figure 7. After re-

viewing the recommendations (assignments), the user

has the option to export them, either as a CSV ﬁle or

in XLSX format for further analysis or sharing.

Figure 4: The Home page of the web-system.

Conference Management System Utilizing an LLM-Based Recommendation System for the Reviewer Assignment Problem

1009

Figure 5: The basic user page of the system with all the available options.

Figure 6: The conference page of the system with all the available options.

Figure 7: The reviewers-to-papers assignment page of the system.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

1010

5 CONCLUSION AND FUTURE

PLANS

In conclusion, the process of assigning reviewers to

papers is, without question, the most time-consuming

task for the organizers of a conference. Therefore,

it is absolutely necessary to use an automated sys-

tem that completes this task and achieves high per-

formance and quality assignments. We strongly be-

lieve that a web Conference Management System that

solves RAP, is required in order to upgrade a confer-

ence quality, by upgrading the peer-review process.

In this work, we propose such a system which ex-

ploits the power of LLMs by incorporating them into

Recommendation Systems. Regarding our plans for

future work:

• We will put our system to online access, in the

near future.

• We currently work on the front-end to improve

UX-UI issues.

• We will make our system available to all academic

personnel organizing a conference.

• We plan to continue running experiments on the

performance and efﬁciency of the reviewer as-

signment module using different datasets.

• We are going to perform an extensive evaluation

and calculate a variety of performance metrics for

the reviewer assignment module.

ACKNOWLEDGEMENTS

The work of M. Vassilakopoulos and A. Cor-

ral was funded by the Spanish Government

MCIN/AEI/10.13039/501100011033 and the "ERDF

A way of making Europe" under the Grant PID2021-

124124OB-I00.

REFERENCES

Aksoy, M., Yanik, S., and Amasyali, M. F. (2023). Re-

viewer assignment problem: A systematic review of

the literature.

Bouanane, K., Medakene, A. N., Benbelghit, A., and Bel-

haouari, S. B. (2024). Faircolor: An efﬁcient algo-

rithm for the balanced and fair reviewer assignment

problem. Inf. Proc. & Management, 61(6):103865.

Charlin, L. and Zemel, R. S. (2013). The toronto paper

matching system: An automated paper-reviewer as-

signment system. In ICML workshop on peer review-

ing and publishing models.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K.

(2019). BERT: Pre-training of deep bidirectional

transformers for language understanding. In Burstein,

J., Doran, C., and Solorio, T., editors, Proc. 2019

Conf. of the North American Chapter of the ACL: Hu-

man Language Technologies, Volume 1 (Long & Short

Papers), pages 4171–4186, Minneapolis, MN. ACL.

Latypova, V. (2023). Reviewer assignment decision support

in an academic journal based on multicriteria assess-

ment and text mining. In 2023 IX Int. Conf. on Inf.

Technology and Nanotechnology (ITNT), pages 1–4.

Leyton-Brown, K., Mausam, Nandwani, Y., Zarkoob, H.,

Cameron, C., Newman, N., and Raghu, D. (2024).

Matching papers and reviewers at large conferences.

Artiﬁcial Intelligence, 331:104119.

Maleszka, M., Maleszka, B., Król, D., Hernes, M., Mar-

tins, D. M. L., Homann, L., and Vossen, G. (2020). A

Modular Diversity Based Reviewer Recommendation

System, page 550–561. Springer Singapore.

Ribeiro, A. C., Sizo, A., and Reis, L. P. (2023). Investigat-

ing the reviewer assignment problem: A systematic

literature review. Journal of Information Science.

Stelmakh, I., Wieting, J., Neubig, G., and Shah, N. B.

(2023). A gold standard dataset for the reviewer as-

signment problem.

Stergiopoulos, V., Tsianaka, T., and Tousidou, E. (2022).

Aminer citation-data preprocessing for recommender

systems on scientiﬁc publications. In Proc. 25th Pan-

Hellenic Conf. on Informatics, PCI ’21, page 23–27,

New York, NY. ACM.

Sun, F., Liu, J., Wu, J., Pei, C., Lin, X., Ou, W.,

and Jiang, P. (2019). Bert4rec: Sequential recom-

mendation with bidirectional encoder representations

from transformer. In Proc. 28th ACM Int. Conf. on

Inf. and Knowledge Management, CIKM ’19, page

1441–1450, New York, NY. ACM.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L. u., and Polosukhin,

I. (2017). Attention is all you need. In Guyon,

I., Luxburg, U. V., Bengio, S., Wallach, H., Fer-

gus, R., Vishwanathan, S., and Garnett, R., editors,

Advances in Neural Information Processing Systems,

volume 30. Curran Associates, Inc.

Wu, C., Wu, F., Qi, T., and Huang, Y. (2021). Empower-

ing news recommendation with pre-trained language

models. In Proc. 44th Int. ACM SIGIR Conf. on Re-

search and Development in Inf. Retrieval, SIGIR ’21,

page 1652–1656, New York, NY. ACM.

Yujian, L. and Bo, L. (2007). A normalized levenshtein dis-

tance metric. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 29(6):1091–1095.

Zhang, Y., Shen, Y., Kang, S., Chen, X., Jin, B., and Han, J.

(2024). Chain-of-factors paper-reviewer matching.

Zhao, X. and Zhang, Y. (2022). Reviewer assignment al-

gorithms for peer review automation: A survey. Inf.

Proc. & Management, 59(5):103028.

Conference Management System Utilizing an LLM-Based Recommendation System for the Reviewer Assignment Problem

1011