Intelligent Platform Using Natural Language Processing for

Pre-Selection of Personnel Through Professional Values Required

by Private Companies

José L. Silva

, Vicente O. Moya

, Daniel W. Burga

and Carlos A. Tello

Faculty of Engineering, Peruvian University of Applied Sciences, Lima Metropolitana, Lima, Peru

Department of Physics, Paulista State University, São Paulo, Brazil

Keywords: Natural Language Processing (NLP), Professional Values, Cultural Alignment, Preselection Process,

Intelligent Platform.

Abstract: The pre-selection process is essential for companies, as it ensures the recruitment of competent staff for each

position, maintaining a positive working environment crucial to meeting organizational objectives. This

research presents an intelligent platform for the pre-selection of personnel based on professional values. When

the selection process is poorly executed, it can lead to economic and intangible losses, such as delays in project

progress and team demotivation. The platform employs natural language processing (NLP) to analyse

applicant data, making it easier to identify candidates that best suit the needs of the company. The results

indicate that the intelligent platform achieves an 80% accuracy in its recommendations.

1 INTRODUCTION

As of 2022, more than 26 million video interviews

and 5 million candidate evaluations have been

conducted using artificial intelligence (Koutsoumpis

et al.,2024). The COVID-19 pandemic has

accelerated the shift from traditional face-to-face

interviews to digital interviews. This situation has

been crucial to increase the quality of human

resources decisions in a company (Fernandes et al.,

2021).

On the other hand, the use of machine learning

and sentiment analysis in personnel selection open an

opportunity to innovate in human resources

management (Campion, 2024), allowing for more

accurate decision-making even when the data set to

be analysed is very broad (Radonjić, Duarte &

Pereira, 2024). In the same vein, aspects such as data

privacy and personnel selection biases caused by

machine learning models without the sensitive ability

to recognize human soft features have been

https://orcid.org/0009-0002-5103-2881

https://orcid.org/0009-0000-6429-6101

https://orcid.org/0000-0003-0312-727X

https://orcid.org/0000-0002-0369-8999

questioned during early AI integrations within

companies (Delecraz et al., 2022).

However, to support this research, the APA-AVI

study demonstrated that the accuracy of personality

trait assessments significantly improved when

machine learning models were trained using

observer-based reports (Koutsoumpis et al., 2024).

This suggests that defining an organizational ideal of

value—understood as the set of human qualities that

align with the company's culture and make a

candidate more likely to be hired—helps to reduce the

high variance and bias often found in automated

evaluation processes (Wang et al., 2024).

For the development of this analysis, some

complementary points of view will be evaluated. On

the one hand, manual coding, used to classify or

analyse data in a traditional way, is expensive and,

even with extensive training programs, high levels of

reliability are not always achieved (Van Atteveldt et

al. (2021). This highlights the limitations of

traditional approaches to today’s challenges. On the

other hand, it proposes the use of advanced

techniques, such as the KNN method and the

Silva, J. L., Moya, V. O., Burga, D. W. and Tello, C. A.

Intelligent Platform Using Natural Language Processing for Pre-Selection of Personnel Through Professional Values Required by Private Companies.

DOI: 10.5220/0013463000003964

In Proceedings of the 20th International Conference on Software Technologies (ICSOFT 2025), pages 201-206

ISBN: 978-989-758-757-3; ISSN: 2184-2833

201

BERTEC model, to overcome these limitations and

optimize the precision in the analysis (Latifi,

Jannach& Ferraro, 2022). In the same vein, the use of

pre-trained models has proven to be a successful

strategy, one of these being BERT (Gu et al., 2022).

Therefore, it is committed to perform a procedure

based on BERT models, with the help of a data set of

more than ten thousand records, it is expected that the

training will exceed in F1 to 85% of adjustment, this

basing the concept of ideal professional value on

recruiters with years of experience and knowledge in

the field (Cui, et al., 2021).

The structure of the research is organized as

follows: Section II presents previous studies on

Natural Language Processing (NLP) techniques

applied to recruitment. Section III details the sources,

variables and methods of data collection, including

criteria for assessing the suitability of candidates for

professional values. Section IV discusses the

implementation of the PLN system for data

processing and profile information extraction, using

pre-trained fine-tuning models based on Wiki Large.

Section V validates results using performance

metrics, such as the MAE. It is necessary to show the

analysis of graphical efficacy of the results (He et al.,

2024). Finally, section VI summarizes the main

findings, limitations and possible future

improvements to the system.

2 RELATED WORKS

Regarding the country of origin of this research, 3 out

of 9 sources come from the Americas region. There

are studies in which methodologies of perception of

personality traits and the relationship to certain

behaviors were evaluated. It is also proposed that the

way candidates express themselves can significantly

affect the outcome of a job interview, especially when

using AI tools (Martín-Raugh et al., 2023). Under the

same study parameters, the impact of correct

personnel selection was evaluated using PLN and

XGBoost techniques, whose central study process

involved the collection of approximately 1.2 million

job reviews (Feng, 2023). It was concluded that a 1%

increase in review ratings is correlated with an

increase from 0.68% to 0.73% in the company's

market value (Feng, 2023).

Regarding research techniques, natural language

processing (NLP) has been applied to tourism through

sentiment analysis, tokenization and lemmatization,

techniques that could be adapted in human resources

to identify professional values in applicants.

In fact, the NLP achieved accuracy rates above

80%, suggesting that, in staff selection, it could

improve the identification of candidates aligned with

organizational culture (Koutsoumpis et al., 2024).

There is also a common approach at the

intersection between PLN technology and human

evaluation (Campion & Campion, 2024; Delecraz et

al., 2022; Koutsoumpis et al., 2024). For example,

regression analyses are used to determine the impact

of vocal and visual characteristics on hiring decisions,

finding a significant correlation with an effect size of

0.40 (Delecraz et al., 2022).

On the other hand, the readability and clarity in

the texts generated by artificial intelligence has been

evaluated, noting that models such as ChatGPT and

Bard achieved almost identical relevance scores, with

averages of 4.93 and 4.92, respectively (Campion &

Campion, 2024). These studies illustrate how

predictive models and artificial intelligence-driven

analysis contribute to decision-making in Human

Resources, sentiment assessment, and performance

forecasting (Campion & Campion, 2024; Delecraz et

al., 2022; Koutsoumpis et al., 2024).

In this sense, the implementation of algorithms

based on machine learning improves accuracy and

equity in identifying suitable candidates for different

positions (Campion & Campion, 2024; Delecraz et

al., 2022; Koutsoumpis et al., 2024).

3 METHOD

The developed platform introduces an innovative

approach to assessing professional values and soft

skills by combining natural language processing

(NLP) and deep learning techniques. Using embeds,

the platform transforms a set of key organizational

values into vector representations, efficiently stored

in pickling files. Similarly, it converts more than

10,000 candidate descriptions into embeds that allow

for high-precision comparisons. Subsequently, the

results that identify the most relevant values per

candidate are consolidated into an Excel file,

optimizing the flow of data to a Fine-Tuning process

with BERT Distil. This trained model integrates

advanced contextual learning and tuning techniques,

providing a detailed and scalable assessment within

the final platform.

3.1 Data Source Collection

The main data source for this investigation comes

from a private company that provided a dataset of

19,555 applicant records, with confidential

ICSOFT 2025 - 20th International Conference on Software Technologies

202

information such as names, contact details and

censored addresses. These data include detailed

curriculum information, descriptions and internal

performance appraisals. The registers cover both

structured data (age, gender, position applied for) and

unstructured data (description of skills, professional

achievements, personal description, cultural

compatibility).All personal information was

anonymised prior to processing, in compliance with

the company’s internal data protection protocols and

applicable privacy regulations.

3.2 Data Collection

The collected data were pre-processed using

tokenization, lemmatization and sentiment analysis

techniques. These processes ensure the

standardization and cleanliness of texts, as

demonstrated in previous research on NLP applied to

Human Resources (Álvarez-Carmona et al., 2022).

Table 1: Related characteristics.

Characteristics Values

Cultural Compatibilit

0-100%

Model Accurac

Reache

<82%

NLP model BERT, DistilBERT

Number of trainin

records 19,555

Number of values traine

Main skills shown 3

3.3 Modular Procedure

Data processing was carried out using previously

trained NLP models, BERT and DistilBERT,

adjusted to assess professional values and cultural

compatibility. These models were selected for their

effectiveness in reducing biases and their ability to

perform in-depth analyses in complex texts (Gu et al.,

2022; Wang et al., 2024).

Figure 1: Modular structure of the intelligent platform.

The platform operates through five sequential

modules:

 Input Layer:Itreceivestwo inputs—

candidate CVs and recruiterselected values.

 Data Normalization Module: Responsible

for loading, reading, and preprocessing the

input data to ensure consistency.

 DataExtraction and Processing Module:

Text data is tokenized, embeddings are

generated for both CVs and values, and

alignment scores are computed. Results are

saved in a Pickle file.

 NLP and Fine-tuning Module: Uses the

Pickle data to apply a fine-tuned DistilBERT

model for refined matching.

 Alignment Module:Calculates alignment

percentages and ranks candidates by best fit.

The platform was developed in Python, using

Transformers (Hugging Face) for model fine-tuning,

scikit-learn for evaluation, Pandas for data handling,

and PyTorch for embedding generation and similarity

computation.

4 EXPERIMENTAL CONTEXTS

The data was collected by the company over an

estimated period of 2 years and 9 months, from

October 2021 to July 2024, resulting in a total of 421

recruitment campaigns with more than 19,555

candidates for various positions.

Table 2: Number of applicants per year.

Yea

Number of a

licants

2021 3170

2022 8335

2023 6940

2024 1110

Table 3: General basics of training.

Code AABAMA

Professional

Summary

Computer Engineering student at PUCP,

with an interest in video games, data

analysis, business, software, and UX/UI.

Advanced English proficiency, proactive,

and motivated.

Response 1

I want to apply data analysis to make

etter decisions in business.

Response 2

I have learned to balance studies and

personal life, making key decisions

during difficult times.

Response 3

I wish to improve in data analysis to

transform information into valuable

insights.

Response 4

Yes, I want to pursue a master's degree

in data analysis.

Response 5

Is there another way to access Support

enefits if I do not obtain the

osition?

Intelligent Platform Using Natural Language Processing for Pre-Selection of Personnel Through Professional Values Required by Private

Companies

203

Each record stores information about a candidate,

and it is essential that the experimental dataset focus

primarily on columns containing descriptive

information and candidate self-perception data.

4.1 Data Preparation

In this stage of experimentation and construction of

the predictive model, it is important to consider that

the questions posed by the evaluating authority are

treated as predictive elements and have weight in the

adjustment phase. The responses of each applicant

(records) are based on the assessment elements shown

in the following table.

Table 4: Impact of predictive elements on the model.

Predictive elements

Weight

Do you plan to study for a master's degree or

a doctorate?

12.5

Professional summar

12.5

What do you expect from your future career? 12.5

What are your long-term goals? 12.5

What have been its main achievements and

challenges?

12.5

What have been your main interests? 12.5

What topics or types of work do you want to

focus on in the future?

12.5

What questions would you like to ask us? 12.5

4.2 Training Model

This approach proposes a system that generates

vector representations (embeddings) for both values

and applicant descriptions. A model is then fine-tuned

to predict the values that best correspond to a specific

description.

4.2.1 Embedding Generation for Values and

Descriptions

The embedding generation process converts both

professional values and applicant descriptions into n-

dimensional vectors in a vector space. This allows for

measuring similarity based on their proximity in that

space. For a given value v

and description d, the

embedding is defined as:

𝑒



= 𝑒𝑚𝑏𝑒𝑑𝑑𝑖𝑛𝑔



𝑣





𝜖

ℝ



)

4.2.2 Measuring Similarity Between

Descriptions and Values

To determine which values are most aligned with the

applicant's description, we calculate the cosine

similarity between the embeddings of the description

and each value. Cosine similarity is defined as:

𝑠𝑖𝑚



𝑒



, 𝑒





𝑒



·e



𝑒



·e



)

This formula measures the angle between the two

vectors e

and e

, where a value close to 1 indicates a

higher similarity between the description and the

value, while a value close to -1 indicates they are

completely different.

4.2.3 Model Fine-Tuning

The next step is to fine-tune a pre-trained model to

learn how to predict values more accurately, using

tagged data to correct their errors. The optimized loss

function during fine-tuning is the cross entropy,

defined as:

𝐿



𝑦, 𝑦



= −𝑦



log (𝑦



)







)

In this case, y

is the probability that the model

predicts for value V

, and y

is the true label indicating

if the value is relevant to the applicant. The loss

function is minimized using gradient descent,

adjusting the model weights θ as follows:

𝜃←𝜃−𝜂∇



𝐿(

𝑓



(

𝑒



)

, 𝑦)

)

4.2.4 Predicting Relevant Values

Once adjusted, the model can predict the most

relevant values for a new description of the applicant.

To achieve this, compare the similarity between

embedding the new description and embedding all

values, selecting the closest matches.

𝑣 = 𝑎𝑟𝑔𝑚𝑎𝑥





𝑠𝑖𝑚(𝑒





, 𝑒





)

This approach converts values and descriptions

into comparable numerical representations,

optimizing the model with Fine-Tuning to predict

values more accurately. The use of F1 scoring has

been shown to be highly effective, especially for

predicting multiple relevant values at once. The

effectiveness of the model will be observed after the

combination of both techniques.

5 RESULTS

Table 4 shows that as the number of training samples

increases, both ROC AUC and F1 Score improve

consistently, reaching up to 0.9661 and 0.9482

respectively in the final iteration. Training time

ICSOFT 2025 - 20th International Conference on Software Technologies

204

Table 5: Summary of fit and accuracy metrics.

Step Roc Auc F1 Runtime

Samples

per

secon

800 0.88163 0.816044 26.4994 222 005

1600 0.929502 0.889046 26.6936 220.39

2400 0.933901 0.901786 26,2108 224.45

3200 0.947302 0.916705 26.6532 220,724

4000 0.950816 0.920298 26.3827 222,987

4800 0.954999 0.928422 26.5894 221,253

5600 0.956941 0.930526 27.9548 210.447

6400 0.961154 0.937032 26.8393 219 193

7200 0.961192 0.934783 27.5994 213.157

8000 0.961487 0.939702 26.9416 218.361

8800 0.964093 0.941079 26.7634 219.815

9600 0.959429 0.935578 26.9127 218,571

10400 0.964106 0.942293 26.9424 218.354

12000 0.965494 0.94506 26.8664 218,972

19555 0.966073 0.948213 27.4084 214,642

remains stable (~27 seconds), confirming processing

efficiency. Notably, gains start to stabilise after

around 8800 samples, suggesting diminishing returns

beyond that point.

Overall, DistilBERT fine-tuning outperformed

the previous RoBERTa-based model (96% vs. 84%

accuracy), offering greater precision in identifying

candidates aligned with organisational values.

Figure 2: Metrics evaluation and comparison.

This figure illustrates the comparative

performance of the proposed DistilBERT model

across four key evaluation metrics: Eval Loss, ROC

AUC, Hamming Loss, and F1 Score. The results

show that the model achieves a low evaluation loss

and Hamming Loss, both under 0.1, while obtaining

high ROC AUC and F1 Score values, both exceeding

0.94. This balance indicates strong predictive

capacity with minimal error, confirming that the

model not only makes accurate predictions but also

maintains consistency across multiple classification

labels. The values were normalised between 0 and 1

for comparative clarity.

As for the impact on the pre-selection process, the

implementation of DistilBERT led to a 12% reduction

in selection errors, optimizing the recruitment process

by minimizing incorrect decisions. This reduction

represents a significant improvement in platform

accuracy.

Table 6: Validation and Comparison with Terman Software.

Terman Intelli

entPlatform

ID Vr Nr Me Rm Ca Ex Pd In Int Ta (%)

1 75 70 65 72 68 85 88 78 88 84.6

2 60 68 72 65 70 75 65 75 80 75

3 55 50 60 55 60 78 82 85 82 76.3

4 88 85 90 85 88 80 78 76 84 80.6

5 60 55 58 65 68 88 90 82 90 84.4

6 45 40 50 48 52 72 68 60 70 69.8

7 78 75 80 77 80 82 85 85 86 81

8 50 55 60 58 65 78 82 85 84 81.6

9 82 80 85 80 82 85 84 80 85 81.4

10 65 60 62 67 68 80 85 78 88 83

Table 7: Meaning of abbreviations.

Abbreviation Meanin

Verbal Reasonin

Numerical Reasoning

Me Short-term memory or retention

acit

Rm Mechanical Reasoning

Ca Abstract Comprehension or

Anal

tical Ca

acit

Ex Excellence

Professional Development

In Innovation

Int Integrit

(

)

Total ali

nment %

This table presents a comparison between Terman

Test results and the intelligent platform’s inferred

professional values. While the Terman dimensions

focus on cognitive abilities, the platform evaluates

alignment with organisational values such as

Excellence, Integrity, and Innovation. The alignment

percentage reflects the consistency between both

approaches, with most candidates scoring above 75%.

This suggests that the platform is capable of

approximating traditional human assessment criteria

through automated value analysis, offering a scalable

alternative for early-stage candidate evaluation.

Training time with DistilBERT was reduced by

72%—from 6 hours to just 1 hour and 40 minutes—

without compromising performance. This efficiency,

combined with high accuracy, enhances the

Intelligent Platform Using Natural Language Processing for Pre-Selection of Personnel Through Professional Values Required by Private

Companies

205

platform’s capacity to accelerate recruitment while

ensuring alignment with company values and

supporting more objective decision-making in HR.

6 CONCLUSIONS

The developed platform applies NLP and deep

learning models to analyse and select candidates

aligned with organizational values, achieving 96%

accuracy and a 12% reduction in selection errors. This

system not only optimizes the recruitment process by

identifying more compatible profiles, but also

improves cultural integration and objectivity in hiring

decisions, allowing the company to build more

cohesive teams aligned with strategic objectives.

REFERENCES

Álvarez-Carmona, M., Aranda, R., Rodríguez-Gonzalez, A.

Y., Fajardo-Delgado, D., Sánchez, M. G., Pérez-

Espinosa, H., Martínez-Miranda, J., Guerrero-

Rodríguez, R., Bustio-Martínez, L., & Díaz-Pacheco,

Á. (2022). Natural language processing applied to

tourism research: A systematic review and future

research directions. Journal of King Saud University -

Computer and Information Sciences, 34(10), 10125–

10144. https://doi.org/10.1016/j.jksuci.2022.10.010

Campion, E. D., & Campion, M. A. (2024). Impact of

machine learning on personnel selection.

Organizational Dynamics, 53(1), 101035. https://

doi.org/10.1016/J.ORGDYN.2024.101035

Cui, Y., Che, W., Liu, T., Qin, B., & Yang, Z. (2021). Pre-

Training with Whole Word Masking for Chinese

BERT. IEEE/ACM Transactions on Audio Speech and

Language Processing, 29, 3504–3514.

https://doi.org/10.1109/TASLP.2021.3124365

Delecraz, S., Eltarr, L., Becuwe, M., Bouxin, H., Boutin,

N., &Oullier, O. (2022). Responsible Artificial

Intelligence in Human Resources Technology: An

innovative inclusive and fair design matching algorithm

for job recruitment purposes. Journal of Responsible

Technology, 11. https://doi.org/10.1016/j.jrt.2022.1

00041

Feng, S. (2023). Job satisfaction, management sentiment, and

financial performance: Text analysis with job reviews

from indeed.com. International Journal of Information

Management Data Insights, 3(1). https://doi.org/10.1

016/j.jjimei.2023.100155

Fernandes, T., São, H., Pereira, J., & Pereira, V. (2023).

Artificial intelligence applied to potential assessment

and talent identification in an organizational context.

Information Sciences, 2. https://doi.org/10.1016/

j.heliyon.2023.e14694

Gu, Y., Tinn, R., Cheng, H., Lucas, M., Usuyama, N., Liu,

X., Naumann, T., Gao, J., & Poon, H. (2022). Domain-

Specific Language Model Pretraining for Biomedical

Natural Language Processing. ACM Transactions on

Computing for Healthcare, 3(1). https://doi.org/

10.1145/3458754

He, W., Farrahi, K., Chen, B., Peng, B., & Villavicencio, A.

(2024). Representation transfer and data cleaning in

multi-views for text simplification. Pattern Recognition

Letters, 177, 40–46. https://doi.org/10.1016/J.PA

TREC.2023.11.011

Huang, X., Yang, F., Zheng, J., Feng, C., & Zhang, L.

(2023). Personalized human resource management via

HR analytics and artificial intelligence: Theory and

implications. Asia Pacific Management Review, 28(4),

598–610. https://doi.org/10.1016/j.apmrv.2023.04.004

Koutsoumpis, A., Ghassemi, S., Oostrom, J. K., Holtrop,

D., van Breda, W., Zhang, T., & de Vries, R. E. (2024).

Beyond traditional interviews: Psychometric analysis

of asynchronous video interviews for personality and

interview performance evaluation using machine

learning. Computers in Human Behavior, 154.

https://doi.org/10.1016/j.chb.2023.108128

Latifi, S., Jannach, D., & Ferraro, A. (2022). Sequential

recommendation: A study on transformers, nearest

neighbors and sampled metrics. Information Sciences,

609, 660–678. https://doi.org/10.1016/j.ins.2022.07.079

Latifi, S., Mauro, N., &Jannach, D. (2021). Session-aware

recommendation: A surprising quest for the state-of-

the-art. Information Sciences, 573, 291–315.

https://doi.org/10.1016/j.ins.2021.05.048

Martin-Raugh, M. P., Leong, C. W., Roohr, K. C., & Chen,

X. (2023). Perceived conscientiousness and openness to

experience through the relationship between vocal and

visual features and hiring decision in the interview.

Computers in Human Behavior Reports, 10.

https://doi.org/10.1016/j.chbr.2023.100272

Raman, R., Venugopalan, M., & Kamal, A. (2024).

Evaluating human resources management literacy: A

performance analysis of ChatGPT and bard. Heliyon,

10(5). https://doi.org/10.1016/j.heliyon.2024.e27026

Radonjić, A., Duarte, H., & Pereira, N. (2024). Artificial

intelligence and HRM: HR managers’ perspective on

decisiveness and challenges. European Management

Journal, 42(1), 57–66. https://doi.org/10.101

6/J.EMJ.2022.07.001

van Atteveldt, W., van der Velden, M. A. C. G., &Boukes,

M. (2021). The Validity of Sentiment Analysis:

Comparing Manual Annotation, Crowd-Coding, Dictio-

nary Approaches, and Machine Learning Algorithms.

Communication Methods and Measures, 15(2), 121-140.

https://doi.org/10.1080/19312458.2020.1869198

Wang, L., Zhou, Y., Sanders, K., Marler, J. H., & Zou, Y.

(2024). Determinants of effective HR analytics

Implementation: An In-Depth review and a dynamic

framework for future research. Journal of Business

Research, 170, 114312. https://doi.org/10.1016/J.

JBUSRES.2023.114312.

ICSOFT 2025 - 20th International Conference on Software Technologies

206