Fuzzy-Weighted Sentiment Recognition for

Educational Text-Based Interactions

Christos Troussas

, Christos Papakostas

, Akrivi Krouska

, Phivos Mylonas

and Cleo Sgouropoulou

Department of Informatics and Computer Engineering, University of West Attica, Egaleo, Greece

Keywords: Fuzzy Sentiment Analysis, Educational Natural Language Processing, Affective Computing in Learning,

Interpretability in AI, Student Feedback Mining.

Abstract: In web-based educational environments, students often express complex emotional states – such as confusion,

frustration, or engagement – through reflective texts, forum posts, and peer interactions. Traditional sentiment

analysis tools struggle to capture these subtle, mixed signals due to their reliance on rigid classification

schemes and lack of domain sensitivity. To address this, we propose a fuzzy-weighted sentiment recognition

framework designed specifically for educational text-based interactions. The system combines an augmented

sentiment lexicon, rule-based modifier detection, and semantic similarity using pretrained Sentence-BERT

embeddings to extract nuanced sentiment signals. These inputs are interpreted by a Mamdani-type fuzzy

inference engine, producing a continuous sentiment score and a confidence weight that reflect both the

strength and reliability of the learner’s affective state. The paper details the linguistic pipeline, fuzzy

membership functions, inference rules, and aggregation strategies that enable interpretable and adaptive

sentiment modeling. Evaluation on a corpus of 1125 annotated student texts from a university programming

course shows that the proposed system outperforms both lexicon-based and deep learning baselines in

accuracy, robustness, and interpretability, demonstrating its value for affect-aware educational applications.

1 INTRODUCTION

In the context of distance learning, student

communication is increasingly taking place via

textual media, including discussion boards, reflective

questions, peer review, and open-ended tests. These

text-based discussions provide much insight into

students' thinking, engagement, and affect (Yuvaraj et

al., 2025). Yet, despite their value for instruction,

these texts are not necessarily examined, and

indicators of frustration, satisfaction, confusion, or

motivation may not be noticed (Johansen et al., 2025;

Troussas et al., 2019).

Instructors are frequently unaware of emotional

undercurrents that could signal disengagement,

conceptual difficulty, or misunderstanding –

especially in asynchronous or large-scale settings

https://orcid.org/0000-0002-9604-2015

https://orcid.org/0000-0002-5157-347X

https://orcid.org/0000-0002-8620-5255

https://orcid.org/0000-0002-6916-3129

https://orcid.org/0000-0001-8173-2622

where personal attention is limited.

Sentiment analysis has emerged as a promising

tool for augmenting educational platforms with

affect-sensitive capabilities (Grimalt-Álvaro & Usart,

2024; Kardaras et al., 2024; Tasoulas et al., 2024). By

identifying emotional cues in student language,

sentiment models can help build responsive,

personalized systems that adjust instructional

strategies based on learner affect (Benazer et al.,

2024). However, the majority of existing sentiment

analysis methods rely on categorical labels, such as

positive, negative, or neutral, and apply either

lexicon-based heuristics or supervised classifiers

trained on general-purpose corpora (Alahmadi et al.,

2025). These approaches suffer from several

limitations in the educational domain: they struggle

with the ambiguity and nuance of student discourse,

420

Troussas, C., Papakostas, C., Krouska, A. and Mylonas, P.

Fuzzy-Weighted Sentiment Recognition for Educational Text-Based Interactions.

DOI: 10.5220/0013794200003985

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

In Proceedings of the 21st International Conference on Web Information Systems and Technologies (WEBIST 2025), pages 420-428

ISBN: 978-989-758-772-6; ISSN: 2184-3252

lack interpretability, and often fail to provide

actionable or trustworthy outputs for teachers or

adaptive systems (van der Veen & Bleich, 2025).

Educational texts are not simple declarations of

opinion (Ahmed et al., 2022). A single message may

blend curiosity with uncertainty (“I’m not sure I got

the logic right”), or hesitation with emerging

confidence (“This recursion thing is starting to make

sense”). The emotional expressions are often subtle,

hedged, and domain-specific, particularly in STEM

education where phrases like “I failed the test case”

or “finally compiled successfully” carry implicit

affect. In such contexts, standard sentiment tools tend

to misclassify, overgeneralize, or ignore important

cues (Hafner et al., 2025). Furthermore, educators

require more than just a sentiment label – they need

to know how strong that sentiment is, how reliable the

estimate is, and how to interpret it in light of

instructional goals.

To address these challenges, this paper proposes a

fuzzy-weighted sentiment recognition framework

tailored specifically for educational text-based

interactions. Unlike traditional models that make

binary or ternary decisions, our system employs fuzzy

logic to model sentiment as a continuous and

interpretable construct. It assigns each student

utterance a sentiment score on a real-valued scale

[−1,+1] along with a confidence weight [0,1]

reflecting both the polarity and the degree of

certainty. This approach allows the system to capture

the vagueness and variability inherent in learner

expression, while maintaining pedagogical

interpretability and technical robustness.

The model we use is grounded in real data

gathered from a university-level Java programming

course offered through an online learning

environment. The analysis includes students’ forum

posts, their weekly reflections, and feedback gathered

at the end of the course, using a hybrid analytical

pipeline that leverages linguistic preprocessing,

domain-specific sentiment lexicons, contextual

embeddings, and a Mamdani-type fuzzy inference

engine. By combining domain-specific knowledge

with fuzzy reasoning methods, our goal is to bridge

affective computing with real-world applications in

education.

The contributions of this paper are threefold. First,

we introduce a novel fuzzy-weighted sentiment

analysis model designed for the educational domain,

which combines symbolic interpretability with

context-aware computation. Second, we develop a

domain-specific sentiment lexicon enriched with

intensity and confidence metadata, adapted to student

language in technical learning contexts. Third, we

evaluate our approach on a curated dataset of

annotated educational texts, comparing it against both

classical and deep learning sentiment baselines. Our

results show that the proposed system not only

achieves competitive performance but also produces

more nuanced, trustworthy outputs that can support

adaptive learning and instructor awareness.

The remainder of the paper is structured as

follows. Section 2 reviews prior work in sentiment

analysis, particularly in educational contexts. Section

3 outlines the challenges and motivations for

modeling sentiment in student-generated content.

Section 4 presents the architecture and logic of the

fuzzy-weighted sentiment framework. Section 5

describes experimental setup, and analyzes the

results. Finally, Section 6 concludes with reflections

and directions for future research.

2 RELATED WORK

Sentiment analysis assists us in grasping information

on the web and the way individuals utilize language.

It comes in handy in fields such as obtaining customer

opinions, social media monitoring, and websites that

provide suggestions. Sentiment analysis is also

gaining popularity in the field of education for

monitoring students' moods, comprehending their

emotions, and assisting with personal learning.

Nevertheless, technology currently is often not

effective when dealing with complicated matters,

ambiguous meanings, and context in educational

debates.

Initially, sentiment analysis techniques

predominantly employed rules and word lists such as

SentiWordNet, AFINN, and VADER (Hutto &

Gilbert, 2014). These provide fixed scores to words

or phrases. These are robust and require minimal

training data, but they are not contextual, do not

handle negation, and do not handle varying uses of

language in different domains. In addition, methods

that employ lexicon lists tend to miss nuanced or

blended feelings. This is usually observed in school

when students express frustration and improvement

in a single sentence.

With the rise of machine learning, supervised

classifiers such as Naïve Bayes, Support Vector

Machines (SVM), and Random Forests were

introduced for sentiment analysis, offering improved

generalization and adaptability to specific domains

(Pang & Lee, 2008). More recently, deep learning

models – particularly Recurrent Neural Networks

(RNNs), Convolutional Neural Networks (CNNs),

and Transformer-based architectures such as BERT –

Fuzzy-Weighted Sentiment Recognition for Educational Text-Based Interactions

421

have achieved state-of-the-art performance in

sentiment classification tasks across multiple

languages and datasets (Devlin et al., 2019; Liu et al.,

2019). These models learn contextual embeddings

and capture long-range dependencies, allowing them

to outperform traditional methods in complex textual

environments. However, they typically require large

annotated corpora, lack transparency, and offer

limited pedagogical interpretability—factors that

pose challenges for adoption in educational

applications.

In education, sentiment analysis has been applied

to examine students' comments (Altrabsheh et al.,

2014), discussion boards (Yang et al., 2013), and

intelligent tutoring systems (Litman & Forbes-Riley,

2006) to monitor happiness, detect frustration, or

forecast whether students will drop out. For instance,

(Wen et al., 2014) employed sentiment trends to

examine how engaged students are in MOOCs,

whereas (D’mello & Graesser, 2013)examined

emotion detection in self-directed learning with both

vocal and non-vocal cues. Although the findings are

promising, the majority of these approaches

employed general classifiers or universal sentiment

lexicons, which tend to misinterpret some words,

formal terms, or ambiguous expressions that students

employ.

More importantly, few educational sentiment

systems provide confidence-aware outputs or allow

for soft classification of mixed emotional signals.

Instructors and adaptive systems benefit more from

interpretable and graded sentiment indicators than

from rigid class assignments, especially in high-

stakes or sensitive contexts such as student confusion

or demotivation. This highlights the need for

frameworks that not only classify sentiment but also

represent its strength, fuzziness, and reliability.

Fuzzy logic provides a compelling foundation for

addressing these limitations. Rooted in the theory of

approximate reasoning, fuzzy systems allow for the

representation of vague, uncertain, or overlapping

categories – such as “slightly negative” or

“moderately positive” – that align more closely with

human intuition. In sentiment analysis, fuzzy

approaches have been used to assign degrees of

polarity to opinions (Subasic & Huettner, 2001),

model emotional intensities in product reviews

(Taboada et al., 2011), and handle ambiguous

expressions in healthcare forums (Jadhav et al.,

2024). These models typically use fuzzy inference

rules, linguistic variables, and membership functions

to map input features (e.g., word polarity, modifier

strength) to output sentiment values on a continuous

scale.

Several studies have proposed hybrid systems

combining fuzzy logic with lexicon-based or machine

learning methods for increased robustness. For

instance, (Sun et al., 2025) combined fuzzy rules with

SVM for Chinese sentiment classification, while

(Ambreen et al., 2024) developed a fuzzy-based

system for sentiment detection in online news

articles. However, few studies have applied fuzzy

reasoning in the educational domain, where

interpretability, nuance, and domain adaptation are

especially important.

A close related work to our approach includes the

framework of (Anagha et al., 2015), which applies

basic fuzzy rules to movie reviews, and the study by

(Devi et al., 2024), which uses fuzzy sets to model

sentiment confidence in e-commerce data. Neither

system, however, is tailored to educational language,

nor do they integrate semantic similarity or discourse-

based modifiers as features.

This paper seeks to extend the application of fuzzy

logic in sentiment recognition by introducing a fuzzy-

weighted framework specifically designed for

educational text-based interactions. By incorporating

lexical polarity, syntactic modifiers, and contextual

agreement into a unified fuzzy inference model, the

proposed system addresses key gaps in

interpretability, domain sensitivity, and confidence

calibration – thereby advancing the state of the art in

affective computing for education.

3 PROBLEM STATEMENT AND

MOTIVATION

Student-generated text in online learning

environments – discussion forums, peer feedback,

reflective logs – carries implicit emotional cues that

are rarely explicit yet critical to learner modeling.

These cues signal fluctuating engagement, confusion,

motivation, and frustration, which, if detected early

and accurately, can inform timely interventions in

adaptive learning platforms.

Traditional sentiment analysis tools, including

both lexicon-based and neural models, tend to assign

rigid labels (e.g., positive, negative), overlooking the

contextual and affective complexity of educational

discourse. For instance, feedback such as “Not bad,

but I still feel lost with recursion” captures a

simultaneous experience of improvement and

confusion – feelings that are poorly represented by

the bare categorical labels or polarity values alone.

The root problem lies in the existing systems'

failure to validly account for the complex, unclear, or

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

422

overlapping emotional states found in student

writing. This failure is especially problematic in

learning environments, where both teachers and

adaptive learning technologies require unambiguous

and confidence-indicative measurements of student

emotions in order to personalize support and manage

learning trajectories.

Fuzzy logic is a formal framework for dealing

with ambiguity. It allows for expressing sentiment on

a continuum, where the interaction of modifier

intensity, contextually related significance, and

uncertainty produces results that are understandable

to humans and hence suitable for both real-time and

post-facto educational usage.

The research is motivated by the need to fill this

gap. We present a fuzzy-weighted sentiment analysis

approach tailored to interactions with educational

content that excels in producing rich and

understandable sentiment measures. Our goal is to

move beyond mere binary sentiment classification

and create useful affect modeling that enhances

educational decision-making and helps student.

4 METHODOLOGY: THE

FUZZY-WEIGHTED

SENTIMENT ANALYSIS

FRAMEWORK

The proposed fuzzy-weighted sentiment detection

system seeks to analyze text-based interactions

generated by students in an online learning

environment. Such interactions can include

contributions made in forums, peer reviews, weekly

reflective journals, and unstructured feedback

comments, all from a first-year undergraduate Java

programming course conducted through Moodle. The

main goal of the system is to derive interpretable

sentiment measures—using fuzzy logic—that capture

the emotional state of the learners, which can range

from frustration to satisfaction, confusion, or interest.

These measures can then be used to improve

personalized feedback, deploy adaptive interventions,

and guide learning analytics.

Unlike traditional sentiment classification models

that produce discrete labels, the new model produces

a continuous sentiment score σ∈[−1,+1] accompanied

by a confidence weight γ∈[0,1]. This score reflects

both the polarity and the strength of sentiment

expressed in a student utterance, while the confidence

weight indicates the degree of certainty in the

inference, based on linguistic and contextual cues.

Each input text is first processed through a

domain-sensitive linguistic pipeline. Tokenization

and lemmatization are performed using spaCy,

preserving key features of educational language. A

rule-based module identifies negators, intensifiers

(e.g., “extremely,” “barely”), hedging expressions

(e.g., “I think,” “sort of”), and discourse markers.

Syntactic parsing is used to highlight subject-verb-

object relations and dependency chains, which are

often crucial in interpreting affect in educational text.

Additionally, we generate dense semantic

representations using Sentence-BERT embeddings,

which serve to measure the contextual alignment

between a student's utterance and prototypical

examples of positive, neutral, or negative sentiment.

The Sentence-BERT embeddings were used

without additional fine-tuning on the educational

dataset, relying on the pretrained all-mpnet-base-v2

model. While domain-specific fine-tuning may

improve semantic similarity estimation, our primary

goal was to preserve generalization and

interpretability.

Thresholds for modifier strength and polarity

adjustment were determined via grid search on a

validation subset (20% of the corpus), optimizing for

interpretability-consistent agreement with expert

annotations.

From this processing, a feature vector x=[S,M,A]

is constructed for each sentence, where S denotes the

lexical sentiment score, M represents modifier

intensity, and A captures contextual agreement. The

sentiment score SSS is computed as a weighted mean

of the scores of matched terms from an augmented

sentiment lexicon. This lexicon combines general-

purpose entries from SentiWordNet and VADER

with education-specific terms (e.g., “debugging,”

“recursion,” “compile”) manually annotated by three

expert raters. Each term has a polarity score

∈[−1,+1], a confidence weight C

∈[0,1], and a

context tag (e.g., “evaluation”, “effort”, “difficulty”).

Modifier intensity M is calculated as a normalized

sum of the impact of linguistic intensifiers,

diminishers, and negations detected in the sentence.

Contextual agreement A is the cosine similarity

between the input sentence’s embedding and seed

vector centroids for each sentiment category.

These features are fuzzified using piecewise linear

membership functions. The lexical sentiment score

SSS is mapped to five fuzzy categories:

LowNegative, MediumNegative, Neutral,

MediumPositive, and HighPositive. Modifier

intensity and contextual agreement are similarly

mapped to fuzzy sets: Weak, Medium, Strong and

Low, Medium, High respectively. Below is the

Fuzzy-Weighted Sentiment Recognition for Educational Text-Based Interactions

423

complete definition of the membership functions for

lexical polarity:

 LowNegative

𝜇





𝑆





1, 𝑆≤−0.8

−0.4 − 𝑆

0.4

,−0.8≤𝑆≤−0.4

0, 𝑆>−0.4

(1)

 MediumNegative

𝜇





𝑆



⎩

⎪

⎨

⎪

⎧

0, 𝑆≤−0.8 𝑜𝑟 𝑆≥0

𝑆+0.8

0.4

,−0.8≤𝑆≤−0.4

−𝑆

0.4

,0.4<𝑆<0

(2)

 Neutral

𝜇





𝑆



=

0, |𝑆|≥0.6

1−

𝑆

0.6

,|𝑆|<0.6

(3)

 MediumPositive

𝜇





𝑆



⎩

⎪

⎨

⎪

⎧

0, 𝑆≤0 𝑜𝑟 𝑆≥0.8

𝑆

0.4

, 0<𝑆≤0.4

0.8 − 𝑆

0.4

, 0.4<𝑆<0.8

(4)

 HighPositive

𝜇





𝑆





0, 𝑆≤0.4

𝑆−0.4

0.4

, 0.4<𝑆≤0.8

1, 𝑆>0.8

(5)

The fuzzy inference engine applies a set of 27

expert-defined rules over these inputs. Each rule has

the general structure:

IF Lexical Polarity is X AND Modifier Intensity

is Y AND Contextual Agreement is Z THEN

Sentiment Output is C

For instance:

 IF Lexical Polarity is MediumNegative AND

Modifier Intensity is Strong AND Contextual

Agreement is High THEN Output is

NegativeStrong

• IF Lexical Polarity is Neutral AND Modifier

Intensity is Weak AND

Contextual Agreement

is Medium THEN Output is Neutral

Each rule produces a fuzzy output set (e.g.,

NegativeStrong, PositiveWeak) with an associated

degree of membership. Using the minimum operator

for rule activation and centroid defuzzification, the

final sentiment score σ is computed as:

𝜎=

∑

𝜇



∙𝑦







∑

𝜇







(6)

where μ

is the activation degree of the i-th rule

and y

is its associated output score (e.g., -0.8 for

NegativeStrong, +0.6 for PositiveWeak). The system

confidence in its prediction is taken as:

𝛾=max



𝜇



(7)

For longer textual entries such as forum posts

containing multiple sentences, sentence-level

sentiment predictions are aggregated using a

weighted average:

𝛴=

∑





∙







∑









, 𝛤=





∑

𝛾







(8)

where w

is a sentence weight derived from TF-

IDF scores and discourse role (e.g., conclusion,

elaboration). This produces a final post-level

sentiment profile (Σ,Γ), interpretable as “moderately

positive with medium confidence” or similar

qualitative labels.

This hybrid fuzzy system offers three advantages

in the educational context: (1) it handles ambiguity

and mixed affect naturally, (2) it avoids the opacity of

deep learning classifiers, and (3) it enables human-

readable outputs that educators and adaptive systems

can interpret and act upon. The model is not trained

via backpropagation but is manually calibrated using

a development set of annotated educational texts,

allowing it to generalize across similar learning

environments without requiring large-scale

supervised data.

5 EVALUATION RESULTS AND

DISCUSSION

To assess the performance and practical viability of

the proposed fuzzy-weighted sentiment recognition

framework, we conducted a comprehensive empirical

evaluation using authentic data collected from a

university-level Java programming course. The

course was delivered over a 12-week semester using

a Moodle-based web platform and included weekly

assignments, peer discussion activities, and reflective

tasks. In total, 96 undergraduate students participated

in the course, generating 2,350 textual messages.

These comprised 1,470 posts and replies in

asynchronous discussion forums, 580 short reflective

responses to weekly prompts, and 300 comments

submitted as part of the final course feedback.

Each message was anonymized and manually

annotated for sentiment polarity by three trained

human raters. Annotations included both a categorical

sentiment label (positive, neutral, or negative) and a

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

424

confidence score on a 3-point scale (low, medium,

high). To ensure annotation quality, inter-rater

agreement was measured using Krippendorff’s alpha,

which yielded a coefficient of 0.81—indicating

substantial agreement. A stratified subset of 800

messages was reserved as the evaluation set. These

messages were balanced across sentiment classes and

served as the gold standard for testing all models.

In this evaluation, our proposed fuzzy-weighted

model was compared with five representative

baseline systems. These include: (1) a simplified

fuzzy logic implementation (FuzzyLex), which uses

only polarity scores and fixed modifier weights; (2)

VADER, a rule-based model optimized for social

media text; (3) TextBlob, a naive Bayes classifier

with a general-purpose sentiment lexicon; (4) a BERT

model fine-tuned on the Stanford Sentiment Treebank

(SST-2) and lightly adapted with domain-specific

examples; and (5) an LSTM-based neural model

trained on 1,000 manually labeled student texts from

the same course domain. All models were evaluated

under the same conditions and tested on the same

evaluation set to ensure consistency and fairness.

To capture different aspects of model

performance, we employed four evaluation metrics.

First, we computed the macro-averaged F1 score to

evaluate classification accuracy across the three

sentiment classes. Second, we calculated the mean

absolute error (MAE) between predicted sentiment

scores and the human-annotated confidence-weighted

ground truth. Third, we assessed interpretability, a

critical factor in educational applications, by asking

three experienced educators to rate the clarity and

pedagogical value of each model’s output on a 5-

point Likert scale. Fourth, we measured robustness to

paraphrasing by evaluating each model on a curated

subset of 100 sentiment-preserving paraphrases

derived from the original texts.

The quantitative results are summarized in Table

1. It presents each model’s performance across all

four metrics. As shown, the proposed fuzzy-weighted

system achieved an F1 score of 0.81, closely

approaching the 0.84 of the fine-tuned BERT model,

and outperforming all other baselines. The fuzzy

model also yielded a low MAE of 0.11, nearly

matching BERT’s 0.10, and significantly

outperforming VADER (0.23) and TextBlob (0.25).

These results indicate that the fuzzy system offers

competitive classification performance while

maintaining lower prediction error.

Table 1: Evaluation Results.

Method F1

Scor

Interpretabili

ty Score

Robustne

(Accurac

Fuzzy-

Weighted

(Propose

)

0.81 0.11 4.7 0.83

FuzzyLe

Baseline

0.72 0.18 4.2 0.75

VADER 0.66 0.23 2.0 0.61

TextBlob 0.62 0.25 2.3 0.57

BERT

Fine-

tune

0.84 0.10 1.2 0.79

LSTM

(domain-

tuned

)

0.79 0.13 2.0 0.76

While the performance differences in F1 and

MAE are noteworthy, perhaps more significant are

the results concerning interpretability and robustness

– two criteria of particular importance in educational

systems. As shown in Figure 1, the fuzzy-weighted

model received the highest interpretability rating

(4.7/5) from domain experts, far exceeding the ratings

of black-box models such as BERT (1.2) and the

LSTM variant (2.0). Educators noted that the fuzzy

model’s scalar sentiment score, coupled with its

confidence output and linguistic justification (e.g.,

influence of modifiers), allowed them to better

understand and act upon the sentiment output.

Figure 1: Interpretability Ratings for All Models here.

Robustness to paraphrasing, depicted in Figure 2,

further demonstrates the reliability of the fuzzy

approach. On the paraphrased subset, where students

expressed similar sentiment using alternate phrasing,

the fuzzy model maintained a robust accuracy of 0.83,

outperforming VADER (0.61), TextBlob (0.57), and

even slightly surpassing BERT (0.79). This indicates

that the rule-guided fuzzy inference engine, though

Fuzzy-Weighted Sentiment Recognition for Educational Text-Based Interactions

425

not pretrained on massive corpora, exhibits strong

generalization capacity in the face of surface

linguistic variation.

Figure 2: Accuracy on Paraphrased Inputs.

To visualize the overall classification

performance, Figure 3 presents a bar plot of macro F1

scores across all models. The fuzzy-weighted system,

while slightly behind BERT in raw accuracy, clearly

outperforms both classical and rule-based baselines

and provides substantially more explainable output.

This balance of performance and interpretability

suggests that fuzzy reasoning is particularly well-

suited for affective modeling in education, where

transparency and trust are necessary for practical

deployment.

Figure 3: Comparison of Macro F1 Scores.

The results of the evaluation reveal that the fuzzy-

weighted sentiment recognition model is an effective

and pedagogically appropriate method to perform

affective analysis in text-based communication for

educational purposes. Having attained a high degree

of accuracy in classification and with minimal

prediction errors, shown to be robust against

linguistic heterogeneity, and offering significant

interpretability to educators, this model is considered

especially appropriate for implementation in adaptive

learning systems, feedback dashboards, or real-time

engagement monitoring tools.

The findings not only demonstrate the

computational efficiency of the fuzzy-weighted

sentiment recognition system but also provide insight

into its widespread applicability and limitations in

terms of online learning environments. Specifically,

the system's ability to identify varied levels of

sentiment and provide interpretable results speaks to

the educational need for transparency and actionable

information within analytics. In contrast to deep

learning models that might achieve superior raw

accuracy but not interpretability, the fuzzy approach

offers linguistically grounded explanations for each

classification outcome. This feature is particularly

salutary for uses requiring human engagement, like

educator dashboards and formative assessment tools.

At the same time, the analysis identifies several

limitations. The rule base and lexicon are currently

designed to effectively treat programming-specific

language; however, while promising, continual

refinement might be necessary in order to address a

broader range of domains or more casual

communicative contexts. In addition, while the fuzzy

model effectively handles the vagueness of sentiment,

it does not at present allow temporality or the

changing states of learners over time. These findings

suggest that future research would be improved by

combining fuzzy reasoning with context-sensitive

neural designs or sequence-based approaches to

sentiment monitoring, thereby enabling more

adaptive and longitudinal models of affect.

6 CONCLUSIONS AND FUTURE

WORK

The present work presented a fuzzy-weighted

sentiment analysis model that combines lexical

polarity, linguistic modifiers, and contextual

agreement in a Mamdani-type fuzzy inference

system. The proposed framework produces

interpretable and fine-grained sentiment ratings. An

empirical study using data from an online discussion

forum for a university-level computer programming

course showed that the proposed approach is

competitive under the F1 measure and mean absolute

error, well outperforming baseline models under

interpretability and robustness. These results confirm

the effectiveness of fuzzy logic in capturing the

emotional nuances present in pedagogical discourse,

which often contains features such as subtlety,

ambivalence, or context-dependent affective terms.

The explainable reasoning of the model and its

educational value make it an appropriate candidate

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

426

for potential deployment in adaptive learning

systems, reflective feedback mechanisms, and

monitoring via learner dashboards.

The fuzzy-weighted approach offers significant

advantages in interpretability and flexibility, but

many directions for future work are still to be

pursued. Foremost, the expansion of the domain-

specific sentiment lexicon to include a wider variety

of academic disciplines and communication

modalities (e.g., chat posts or transcribed voice

communications) would enhance its applicability.

Second, the use of hybrid approaches that combine

fuzzy reasoning with transformer-based contextual

embeddings may improve handling of complex

semantic structures and figurative language while still

allowing for some level of explainability. Third,

conducting longitudinal studies of affect over time—

by following emotional trajectories rather than

examining only discrete posts in isolation—may

provide deeper insight into engagement patterns and

learning trajectories. Finally, incorporation of this

approach into intelligent tutoring systems or massive

open online courses (MOOCs), along with user-

centered validation, would enable empirical testing of

the model's performance in real-time educational

interventions and decision-making applications.

REFERENCES

Ahmed, Y., Kent, S., Cirino, P. T., & Keller-Margulis, M.

(2022). The not-so-simple view of writing in struggling

readers/writers. Reading & Writing Quarterly, 38(3),

272–296. https://doi.org/10.1080/10573569.2021.

1948374

Alahmadi, K., Alharbi, S., Chen, J., & Wang, X. (2025).

Generalizing sentiment analysis: A review of progress,

challenges, and emerging directions. Social Network

Analysis and Mining, 15(1), 45. https://doi.org/10.

1007/s13278-025-01461-8

Altrabsheh, N., Cocea, M., & Fallahkhair, S. (2014).

Sentiment analysis: Towards a tool for analysing real-

time students’ feedback. 2014 IEEE 26th International

Conference on Tools with Artificial Intelligence

(ICTAI), 419–423. https://doi.org/10.1109/

ICTAI.2014.70

Ambreen, S., Iqbal, M., Asghar, M. Z., Mazhar, T., Khattak,

U. F., Khan, M. A., & Hamam, H. (2024). Predicting

customer sentiment: The fusion of deep learning and a

fuzzy system for sentiment analysis of Arabic text. Social

Network Analysis and Mining, 14(1), 206.

https://doi.org/10.1007/s13278-024-01356-0

Anagha, M., Kumar, R. R., Sreetha, K., & Reghu Raj, P. C.

(2015). Fuzzy logic-based hybrid approach for

sentiment analysis of Malayalam movie reviews. 2015

IEEE International Conference on Signal Processing,

Informatics, Communication and Energy Systems

(SPICES), 1–4. https://doi.org/10.1109/SPICES.

2015.7091512

Benazer, S. S., Thangamani, A. K. S., & Prabu, S. (2024).

Integrating sentiment analysis with learning analytics

for improved student engagement. In Advancement of

IoT in Blockchain Technology and Its Applications (pp.

1262–1284). Zenodo. https://doi.org/10.5281/

zenodo.13752749

Devi, R., Prabu, S., & Legapriyadharshini, N. (2024).

Sentiment analysis for the improved user experience in

the e-commerce platform with the fuzzy model. Journal

of Computer Allied Intelligence, 2, 28–40.

https://doi.org/10.69996/jcai.2024013

Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019).

BERT: Pre-training of deep bidirectional transformers

for language understanding. In Proceedings of the 2019

Conference of the North American Chapter of the

Association for Computational Linguistics: Human

Language Technologies (NAACL-HLT) (pp. 4171–

4186). https://doi.org/10.48550/arXiv.1810.04805

D’Mello, S., & Graesser, A. (2013). AutoTutor and

affective AutoTutor: Learning by talking with

cognitively and emotionally intelligent computers that

talk back. ACM Transactions on Interactive Intelligent

Systems, 2(4), Article 23. https://doi.org/10.

1145/2395123.2395128

Grimalt-Álvaro, C., & Usart, M. (2024). Sentiment analysis

for formative assessment in higher education: A

systematic literature review. Journal of Computing in

Higher Education, 36(3), 647–682. https://doi.org/

10.1007/s12528-023-09370-5

Hafner, F. S., Hafner, L., & Corizzo, R. (2025). ‘Slightly

disappointing’ vs. ‘worst sh** ever’: Tackling cultural

differences in negative sentiment expressions in AI-

based sentiment analysis. Journal of Computational

Social Science, 8(3), 57. https://doi.org/10.

1007/s42001-025-00382-y

Hutto, C., & Gilbert, E. (2014). VADER: A parsimonious

rule-based model for sentiment analysis of social media

text. Proceedings of the International AAAI Conference

on Web and Social Media, 8(1), 216–225.

https://doi.org/10.1609/icwsm.v8i1.14550

Jadhav, S., Shiragapur, B., Purohit, R., & others. (2024). A

systematic review on the role of sentiment analysis in

healthcare. Authorea. https://doi.org/10.22541/au.

172516418.81416769/v1

Johansen, M. O., Eliassen, S., & Jeno, L. M. (2025).

Autonomy need satisfaction and frustration during a

learning session affect perceived value, interest, and

vitality among higher education students. Scandinavian

Journal of Educational Research, 69(4), 757–771.

https://doi.org/10.1080/00313831.2024.2348457

Kardaras, D. K., Troussas, C., Barbounaki, S. G., Tselenti,

P., & Armyras, K. (2024). A fuzzy synthetic evaluation

approach to assess usefulness of tourism reviews by

considering bias identified in sentiments and articulacy.

Information, 15(4), 236. https://doi.org/10.3390/info

15040236

Fuzzy-Weighted Sentiment Recognition for Educational Text-Based Interactions

427

Litman, D. J., & Forbes-Riley, K. (2006). Recognizing

student emotions and attitudes on the basis of utterances

in spoken tutoring dialogues with both human and

computer tutors. Speech Communication, 48(5), 559–

590. https://doi.org/10.1016/j.specom.2005.09.008

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V.

(2019). RoBERTa: A robustly optimized BERT

pretraining approach. arXiv preprint.

https://doi.org/10.48550/arXiv.1907.11692

Pang, B., & Lee, L. (2008). Opinion mining and sentiment

analysis. Foundations and Trends in Information

Retrieval, 2(1–2), 1–135. https://doi.org/10.1561/

1500000011

Subasic, P., & Huettner, A. (2001). Affect analysis of text

using fuzzy semantic typing. IEEE Transactions on

Fuzzy Systems, 9(4), 483–496. https://doi.org/10.

1109/91.940962

Sun, Y., Yu, Z., Sun, Y., Xu, Y., & Song, B. (2025). A

novel approach for multiclass sentiment analysis on

Chinese social media with ERNIE-MCBMA. Scientific

Reports, 15(1), 18675. https://doi.org/10.1038/s41598-

025-03875-y

Taboada, M., Brooke, J., Tofiloski, M., Voll, K., & Stede,

M. (2011). Lexicon-based methods for sentiment

analysis. Computational Linguistics, 37(2), 267–307.

https://doi.org/10.1162/COLI_a_00049

Tasoulas, T., Troussas, C., Mylonas, P., & Sgouropoulou,

C. (2024). Affective computing in intelligent tutoring

systems: Exploring insights and innovations. 2024

SEEDA-CECNSM Conference Proceedings,

1–6. https://doi.org/10.1109/SEEDA-CECNSM63478.

2024.00025

Troussas, C., Krouska, A., & Virvou, M. (2019). Trends on

sentiment analysis over social networks: Pre-processing

ramifications, stand-alone classifiers and ensemble

averaging. In G. A. Tsihrintzis, D. N. Sotiropoulos, &

J. L. C. de Carvalho (Eds.), Machine learning

paradigms: Advances in data analytics (pp. 161–186).

Springer. https://doi.org/10.1007/978-3-319-94030-

4_7

van der Veen, A. M., & Bleich, E. (2025). The advantages

of lexicon-based sentiment analysis in an age of

machine learning. PLOS ONE, 20(1), e0313092.

https://doi.org/10.1371/journal.pone.0313092

Wen, M., Yang, D., & Rose, C. P. (2014). Sentiment

analysis in MOOC discussion forums: What does it tell

us? Proceedings of the 7th International Conference on

Educational Data Mining (EDM 2014), 1–8.

http://www.cs.cmu.edu/~mwen/papers/edm2014-

camera-ready.pdf

Yang, D., Sinha, T., Adamson, D., & Rosé, C. P. (2013).

Turn on, tune in, drop out: Anticipating student

dropouts in massive open online courses. Proceedings

of the 2013 NIPS Data-Driven Education Workshop,

11–14.

Yuvaraj, R., Mittal, R., Prince, A. A., & Huang, J. S. (2025).

Affective computing for learning in education: A

systematic review and bibliometric analysis. Education

Sciences, 15(1), 65. https://doi.org/10.3390/

educsci15010065.

WEBIST 2025 - 21st International Conference on Web Information Systems and Technologies

428