Sentiment-Enriched AI for Toxic Speech Detection: A Case Study of

Political Discourses in the Valencian Parliament

Antoni Mestre

1 a

, Franccesco Malafarina

2 b

, Joan Fons

1 c

, Manoli Albert

1 d

, Miriam Gil

3 e

and

Vicente Pelechano

1 f

VRAIN Institute, Universitat Polit

ecnica de Val

encia, 46022, Valencia, Spain

Universit

a degli Studi del Sannio, 82100, Benevento, Italy

Departament d’Inform

atica, Universitat de Val

encia, 46100, Burjassot, Spain

{anmesgas, jjfons, malbert,pele}@vrain.upv.es, f.malafarina@studenti.unisannio.it, miriam.gil@uv.es

Keywords:

Toxic Speech Detection, Hate Speech, Text Classiﬁcation, Parliamentary Speeches.

Abstract:

The increasing prevalence of toxic speech across various societal domains has raised signiﬁcant concerns re-

garding its impact on communication and social interactions. In this context, the analysis of toxicity through

AI techniques has gained prominence as a relevant tool for detecting and combating this phenomenon. This

study proposes a novel approach to toxic speech detection by integrating sentiment analysis into binary clas-

siﬁcation models. By establishing a confusion zone for ambiguous probability scores, we direct uncertain

cases to a sentiment analysis module that informs ﬁnal classiﬁcation decisions. Applied to political discourses

in the Valencian Parliament, this sentiment-enriched approach signiﬁcantly improves classiﬁcation accuracy

and reduces misclassiﬁcations compared to traditional methods. These ﬁndings underscore the effectiveness

of incorporating sentiment analysis to enhance the robustness of toxic speech detection in complex political

contexts, paving the way for future research in this relevant area.

1 INTRODUCTION

The growing prevalence of toxic speech across di-

verse societal domains has become a signiﬁcant con-

cern due to its harmful effects on communication, so-

cial cohesion, and trust within communities. Toxic

speech, characterized by offensive, inﬂammatory, or

demeaning language, not only disrupts civil discourse

but also fosters polarization, harassment, and exclu-

sion (Buitrago L

opez et al., ). In an era where digital

platforms amplify such speech and where transcripts

of public interactions are increasingly available, the

urgency of developing robust mechanisms for detect-

ing and mitigating toxic content has escalated.

Artiﬁcial intelligence has emerged as a powerful

tool in this context, enabling automated detection and

analysis of toxic speech (Chhabra and Vishwakarma,

2023). Traditional binary classiﬁcation models have

https://orcid.org/0000-0001-8572-2579

https://orcid.org/0009-0004-4494-6620

https://orcid.org/0000-0002-3718-3096

https://orcid.org/0000-0003-3747-400X

https://orcid.org/0000-0002-2987-1825

https://orcid.org/0000-0003-1090-230X

been widely used for this task, offering a straightfor-

ward approach to categorizing content as toxic or non-

toxic. However, these models often face limitations

when confronted with ambiguous cases—instances

where the likelihood of toxicity is unclear (Islam

et al., 2021). This ambiguity, which typically occurs

when a speech’s toxicity is contextually or sentimen-

tally nuanced, can lead to misclassiﬁcation, hindering

efforts to foster healthier communication, especially

in complex environments such as political discourse.

To overcome these challenges, this study presents

a novel framework that enhances toxic speech de-

tection by incorporating sentiment analysis into con-

ventional binary classiﬁcation models. By deﬁning a

”confusion zone” for ambiguous classiﬁcation prob-

abilities, our approach directs uncertain cases to a

sentiment analysis module. This added layer of con-

textual understanding helps reﬁne classiﬁcation deci-

sions, ensuring greater accuracy in identifying toxic

content.

Our methodology is particularly tailored to polit-

ical discourse, where the impact of toxic language

is profound, inﬂuencing public perception, societal

norms, and behavior. We apply this approach to

Mestre, A., Malafarina, F., Fons, J., Albert, M., Gil, M. and Pelechano, V.

Sentiment-Enriched AI for Toxic Speech Detection: A Case Study of Political Discourses in the Valencian Parliament.

DOI: 10.5220/0013159600003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 555-561

ISBN: 978-989-758-737-5; ISSN: 2184-433X

555

speeches from the Valencian Parliament, a bilingual

legislative body, to evaluate its effectiveness in de-

tecting nuanced toxic speech in a politically charged

environment. By improving accuracy and reducing

misclassiﬁcation, this research contributes to ongoing

efforts to enhance the detection of toxic speech and

promote more respectful and constructive communi-

cation in complex social contexts.

2 RELATED WORK

The detection of toxic speech—deﬁned as language

that is rude, disrespectful, or likely to disrupt conver-

sations and alienate users (Arab and D

ıaz, 2015)—has

become increasingly important in the digital era,

where online discourse is widespread and often

volatile. Toxic speech detection has evolved signiﬁ-

cantly, starting with basic rule-based systems and key-

word ﬁltering approaches. These early methods strug-

gled to capture the nuances of language, often leading

to high rates of false positives and negatives, as they

were unable to understand context or infer meaning

from subtle linguistic cues (Bonetti et al., 2023).

The introduction of machine learning represented

a major leap forward, with models like Support Vec-

tor Machines and Naive Bayes improving detection

accuracy by leveraging features extracted from text,

such as word frequencies and n-grams (Subramanian

et al., 2023). While these models offered better gen-

eralization than rule-based approaches, they still en-

countered difﬁculties when dealing with more com-

plex, context-dependent cases of toxicity.

The rise of deep learning further revolutionized

the ﬁeld, offering more sophisticated techniques for

detecting toxic speech. Convolutional Neural Net-

worksand Recurrent Neural Networks became popu-

lar for their ability to automatically learn and repre-

sent features from raw text data (Garg et al., 2023).

Models like Long Short-Term Memory networks and

Transformer-based models such as Bidirectional En-

coder Representations from Transformers (BERT)

have proven especially effective by capturing intricate

semantic relationships and understanding the sub-

tleties of language (Malik et al., 2021). However, de-

spite their improved performance, these models are

resource-intensive, requiring large annotated datasets

and signiﬁcant computational power to train and ﬁne-

tune (Subramanian et al., 2023).

One of the most pressing challenges in toxic

speech detection is the multilingual and culturally nu-

anced nature of global communication. Toxicity is

contextually dependent not only on language but also

on cultural values and norms, making cross-linguistic

detection particularly complex (Leite et al., 2020).

Recent studies have shown that sentiment analysis can

be an effective tool for addressing these challenges,

especially in multilingual contexts. For example, re-

search by Proksch et al. (Proksch et al., 2019) demon-

strated how sentiment analysis applied to legislative

debates across different languages could capture un-

derlying political conﬂicts and nuances, suggesting

that sentiment-aware models could enhance toxicity

detection by providing additional contextual informa-

tion.

Despite these advances, handling ambiguous

cases in toxicity detection remains a signiﬁcant chal-

lenge. Ambiguity often arises when language is nei-

ther clearly toxic nor benign, complicating classiﬁca-

tion efforts. Sheth et al. (Sheth et al., 2022) high-

lighted the difﬁculty of deﬁning the boundaries be-

tween toxic and non-toxic speech, while Subramanian

et al. (Subramanian et al., 2023) discussed the limita-

tions of current models in addressing these borderline

cases. Approaches such as multi-channel CNNs have

been proposed to address this issue, but accurately

classifying ambiguous content continues to pose dif-

ﬁculties.

Building on these advancements, our proposed

methodology integrates sentiment analysis with tradi-

tional binary classiﬁcation models to address the chal-

lenge of ambiguous cases. By introducing a confu-

sion interval for uncertain toxicity scores, our model

redirects these cases to a sentiment analysis module,

which enriches the classiﬁcation process with addi-

tional emotional and contextual information. This

dual-layered approach improves the robustness and

accuracy of toxic speech detection, particularly in

complex, politically charged environments where nu-

ance and sentiment play a critical role.

3 PROPOSAL ARCHITECTURE

To ensure robust and accurate detection of toxic

speech, we propose a comprehensive, modular frame-

work that integrates multiple components to enhance

the precision, adaptability, and reliability of toxicity

inference. This architecture is designed to handle the

challenges posed by complex, multilingual, and nu-

anced content, making it well-suited for both digital

platforms and politically charged environments. The

following section details the methodology and system

architecture, which together form the backbone of our

proposed solution.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

556

3.1 Data Acquisition Module

The ﬁrst step in the process is the Data Acquisition

Module, which is responsible for the automatic re-

trieval of textual data from diverse sources, includ-

ing digital platforms (social media, online forums)

and transcriptions of spoken interactions (e.g., parlia-

mentary debates). This module is designed to pro-

cess various types of input data, ensuring scalability

and adaptability to different environments and lan-

guages. By encompassing multiple sources, it allows

the framework to operate on both real-time streams

and historical archives, making it applicable to a wide

range of use cases, from online content moderation to

retrospective analysis of political speeches.

The acquisition process includes mechanisms to

handle noise in the data, such as ﬁltering non-relevant

content and pre-processing steps like tokenization and

removal of stop words. This ensures that only perti-

nent text data are fed into subsequent modules, im-

proving the overall efﬁciency and accuracy of the sys-

tem.

3.2 Language Uniﬁcation Module

Given the multilingual nature of modern discourse,

particularly in political settings such as the Valencian

Parliament, the Language Uniﬁcation Module plays

a crucial role in standardizing text data before fur-

ther processing. This module automatically detects

the language of incoming text and translates it into a

uniﬁed language format (in this case, Spanish), en-

abling consistent and reliable analysis.

The challenge of multilingual text is twofold:

lexical variations across languages and the differing

cultural connotations of certain words. The Lan-

guage Uniﬁcation Module addresses both by lever-

aging state-of-the-art machine translation models that

are sensitive to context. This ensures that translations

maintain the semantic integrity of the original speech,

preserving subtle nuances that are essential for accu-

rate toxicity detection. By addressing linguistic dis-

crepancies at this early stage, the framework is able

to avoid errors that might arise from handling multi-

ple languages simultaneously, ensuring a more coher-

ent and reliable toxicity assessment downstream.

3.3 Toxic Detection Module

The Toxic Detection Module serves as the core com-

ponent of the framework, responsible for evaluating

the toxicity of the standardized text. This module

utilizes advanced AI techniques, speciﬁcally binary

classiﬁcation models, to determine whether a given

piece of text is likely to be toxic. In this study, we

employed the pre-trained Detoxify model (Hanu and

Unitary, 2023), which has demonstrated strong per-

formance in various toxic speech detection tasks.

Detoxify assigns a probability score to each text

sample, representing the likelihood that the content is

toxic. A score close to 0 indicates non-toxic content,

while a score closer to 1 suggests toxic speech. How-

ever, despite its effectiveness, the module can some-

times produce ambiguous results, especially when the

probability score falls within a certain range where the

classiﬁcation is not deﬁnitively clear. To address this,

we introduce a predeﬁned ”confusion zone,” typically

ranging between 42% and 58%. Texts with scores

within this zone are neither clearly toxic nor clearly

non-toxic, indicating a need for further analysis to

reach a conﬁdent conclusion.

This confusion zone is particularly relevant in the

context of political discourse, where language is of-

ten nuanced and may involve sarcasm, rhetorical de-

vices, or indirect speech. Such complexities can make

it difﬁcult for the model to assign a clear classiﬁca-

tion, resulting in borderline cases. To handle these,

the system ﬂags these texts for additional processing

by the Sentiment Analysis Module, which provides

further contextual understanding to improve classiﬁ-

cation accuracy.

3.4 Sentiment Analysis Module

To handle these uncertain cases, we introduce a sec-

ond layer of analysis: the Sentiment Analysis Module.

Once a text is ﬂagged by the Toxic Detection Module

as ambiguous, it is redirected to this module for fur-

ther evaluation. The Sentiment Analysis Module per-

forms a deeper analysis of the emotional tone, provid-

ing additional context to aid in the ﬁnal classiﬁcation.

By assessing the emotional valence of the text, the

sentiment analysis adds a nuanced layer of interpre-

tation that binary classiﬁers typically overlook. For

instance, a politically charged statement with a highly

negative sentiment score is more likely to be toxic,

even if the initial classiﬁer was uncertain. On the

other hand, a text with low sentiment intensity may

indicate sarcasm or rhetorical neutrality, reducing the

likelihood of it being classiﬁed as toxic.

This dual-layered approach—combining toxicity

classiﬁcation with sentiment analysis—enables the

system to handle complex and ambiguous language

more effectively. Texts with strong negative sentiment

are reclassiﬁed as toxic, while those with more neu-

tral or positive sentiment are deemed non-toxic, thus

signiﬁcantly reducing false positives and negatives in

the overall classiﬁcation process.

Sentiment-Enriched AI for Toxic Speech Detection: A Case Study of Political Discourses in the Valencian Parliament

557

Figure 1: Architecture diagram of the proposed toxic speech detection framework.

3.5 System Integration and Scalability

The integration of the aforementioned modules forms

a robust system capable of accurately detecting toxic

speech across a wide variety of linguistic and con-

textual scenarios. The modular design allows for

ﬂexibility, as each component can be updated or re-

placed as new models and techniques become avail-

able. For example, the Sentiment Analysis Module

can be easily swapped with more advanced models

like transformer-based sentiment classiﬁers as they

evolve.

Fig. 1 shows a general schematic of the proposed

solution, instantiated with the use case of the Valen-

cian Parliament. The proposed architecture not only

addresses the challenges of ambiguous toxicity infer-

ence but also paves the way for future advancements

in the ﬁeld of toxic speech analysis.

4 CASE STUDY: THE

VALENCIAN PARLIAMENT

The Valencian Parliament, known as Les Corts Va-

lencianes, provides an ideal case study for analyz-

ing toxic speech detection, primarily due to its dual-

language environment (Spanish and Catalan) and the

increasing focus on monitoring political discourse.

As a legislative body where formal and respectful dis-

course is generally expected, parliamentary debates

often reﬂect broader societal trends, including polar-

ized speech and contentious rhetoric. This makes Les

Corts a valuable setting for studying the dynamics of

toxic speech and its impact on public perception and

societal attitudes.

Les Corts Valencianes serves as the legislative

body of the Valencian Community in Spain, where

both Spanish and Catalan hold ofﬁcial language sta-

tus. This bilingual context introduces unique chal-

lenges for toxic speech detection, as language-speciﬁc

nuances can affect how speech is perceived. Conse-

quently, robust language processing capabilities are

required to accurately assess toxic speech in such a

complex setting.

In this study, we applied a combination of the pre-

trained Detoxify model (Hanu and Unitary, 2023) and

the sentimentR library (Rinker, 2016) to enhance toxi-

city detection. The Detoxify model was chosen for its

strong performance in toxic speech detection, and we

integrated sentiment analysis to handle the complex-

ities inherent in political discourse. The combination

allowed us to create a more reﬁned system capable

of distinguishing between overtly toxic language and

subtler, context-dependent forms of toxicity common

in political settings.

Our experiment involved analyzing ten plenary

sessions from the Valencian Parliament. The speeches

were segmented into individual sentences, which

were then used as samples for toxicity analysis. The

initial toxicity inference using Detoxify revealed that

the average level of toxicity across these sessions

ranged from 0.50% to 0.89%, indicating a generally

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

558

(a) Samples distribution. (b) Sentiment analysis results.

Figure 2: Distribution of samples and sentiment analysis results.

low but present occurrence of toxic speech in this for-

mal setting.

For this study, we selected 870 samples, half of

which (435) had a toxicity score greater than 50%, in-

dicating a higher likelihood of toxicity. The other half

had a toxicity score below 50%. These samples were

used to evaluate the performance of our sentiment-

enriched approach. The distribution of these samples

is shown in Fig. 2a, while the results of the sentiment

analysis are depicted in Fig. 2b.

From these 870 samples, we chose a subset of 80

for manual validation. The remaining samples were

divided into intervals based on their toxicity infer-

ence, with one group having scores below 0.25 and

another above 0.75. Fig. 2b shows two distinct clus-

ters in the sentiment analysis: one centered near a sen-

timent score of 0 (indicating neutral or non-toxic con-

tent) and the other near -0.33 (indicating toxic con-

tent).

For validation, we deﬁned a confusion zone for

toxicity inference between 42% and 58%, represent-

ing ambiguous cases where classiﬁcation was uncer-

tain. We compared the performance of the Detox-

ify model alone against our sentiment-enriched model

using this confusion zone. A group of three experts

(two in philology and one in sociology) manually la-

beled the samples for further validation.

The results indicated that without the confusion

zone, our model achieved an accuracy of 80.35%.

However, when the confusion zone was incorpo-

rated, along with sentiment reassignment, accuracy

improved signiﬁcantly to 87.89%. This demonstrates

the effectiveness of adding a sentiment layer to re-

ﬁne toxic speech detection, particularly in politically

charged and linguistically complex contexts such as

the Valencian Parliament.

To further benchmark our model, we compared

its performance against several other state-of-the-art

models, including Logistic Regression, BERT, and

HateBERT (Caselli et al., 2021). Table 1 provides

a comparison of key performance metrics, including

accuracy, precision, recall, and F1-score.

As shown in Table 1, our sentiment-enriched

model outperformed baseline models in all metrics,

achieving an accuracy of 88%, a precision of 86%, a

recall of 87%, and an F1-score of 86%. The inclu-

sion of the confusion zone signiﬁcantly improved the

overall performance, demonstrating the advantage of

using sentiment analysis to handle ambiguous cases.

5 CONCLUSION AND FUTURE

WORK

In this preliminary study, we introduced a novel ap-

proach to toxic speech detection by integrating sen-

timent analysis into traditional binary classiﬁcation

models. The proposed sentiment-enriched framework

signiﬁcantly enhances the accuracy and robustness of

toxicity inference, particularly in handling ambigu-

ous cases where traditional models often struggle. By

adding a sentiment analysis layer, our approach of-

fers a more nuanced understanding of the emotional

and contextual elements of speech, which is critical

in complex and politically charged environments like

the Valencian Parliament.

The results from this study underscore the poten-

tial of combining toxicity detection with sentiment

analysis, especially in domains where language car-

ries signiﬁcant emotional weight and subtlety, such

Sentiment-Enriched AI for Toxic Speech Detection: A Case Study of Political Discourses in the Valencian Parliament

559

Table 1: Comparison of performance metrics across different models. The table compares accuracy, precision, recall, and

F1-score for various models, showing the effectiveness of our sentiment-enriched model with and without the confusion zone.

Model Accuracy Precision Recall F1-Score

Logistic Regression 0.75 0.72 0.68 0.70

BERT 0.84 0.82 0.83 0.82

HateBERT 0.87 0.85 0.86 0.85

Our Model (without confusion zone) 0.80 0.78 0.79 0.78

Our Model (with confusion zone) 0.88 0.85 0.87 0.86

as political discourse. Our method not only improves

detection accuracy but also reduces the likelihood of

misclassifying borderline cases that often fall within

the “confusion zone” of probability scores.

One of the key areas for future improvement is

the sentiment analysis component. While the current

study utilized the sentimentR library, more sophisti-

cated techniques, such as transformer-based models

(e.g., BERT or GPT-like models) speciﬁcally ﬁne-

tuned for sentiment analysis, could offer signiﬁcant

improvements. Transformer models have shown ex-

ceptional capabilities in capturing the intricacies of

language, including context, irony, and sentiment po-

larity, making them ideal candidates for reﬁning this

aspect of the framework. Fine-tuning such models on

domain-speciﬁc datasets, such as political speeches or

social media conversations, could further boost accu-

racy and adaptability.

Another promising direction for future work in-

volves the integration of additional contextual fea-

tures that go beyond the sentiment of the text itself.

For instance, speaker intent, tone, and audience re-

action could provide valuable cues for determining

toxicity. Analyzing speaker history (e.g., previous

inﬂammatory remarks or speech patterns) and audi-

ence engagement (e.g., applause, boos, or online sen-

timent in response to a speech) could offer deeper in-

sights into whether a statement is likely to be toxic or

merely provocative. Leveraging these contextual fea-

tures could enhance the model’s ability to distinguish

between different forms of aggressive or inﬂamma-

tory speech.

A further potential area of research is exploring

cross-modal toxicity detection, where textual data is

combined with other modalities such as audio, video,

or even social media interactions (Maity et al., 2024).

In many cases, tone of voice, body language, or visual

cues may reveal toxicity that is not explicitly present

in the text itself. For example, integrating acoustic

analysis from speeches or debates could capture vari-

ations in tone that indicate sarcasm, anger, or pas-

sive aggression—key indicators of underlying toxic-

ity. Video analysis of facial expressions or gestures

could similarly add depth to toxicity inference, par-

ticularly in live debates or interviews.

In parallel with advancing technical capabilities,

ensuring that these models are explainable and ethi-

cally sound is crucial. As toxic speech detection sys-

tems are increasingly used in sensitive domains, such

as political discourse or online moderation, it is im-

portant to make sure that the decision-making process

is transparent and justiﬁable (Mahajan et al., 2021).

Further research should explore methods for making

these models interpretable, allowing users or moder-

ators to understand why a particular piece of content

was ﬂagged as toxic. Incorporating fairness measures

to prevent biases related to gender, race, or political

ideology will be vital for building trust and ensuring

that the models do not inadvertently perpetuate dis-

crimination.

In conclusion, this preliminary study presents a

promising approach for improving toxic speech de-

tection by integrating sentiment analysis with binary

classiﬁcation models. While the current implemen-

tation has demonstrated success in a political context,

particularly within the Valencian Parliament, there are

ample opportunities for reﬁnement and broader appli-

cation. As toxic speech continues to be a pressing

issue in both political and digital spaces, advancing

this methodology will contribute to fostering healthier

communication environments. Through further val-

idation, the integration of advanced sentiment mod-

els, contextual features, and multi-modal data, this ap-

proach has the potential to lead the way in addressing

the challenges of toxic speech detection and promot-

ing constructive dialogue in complex social settings.

ACKNOWLEDGEMENTS

We acknowledge the support and assistance of Alba

Gasc

on, Julia Palomares and Paula Gabarda in the

validation phase of this study.

This work has been developed with the ﬁnancial

support of the Generalitat Valenciana under project

GV/2021/072 and by the MINECO under project

PRODIGIOUS PID2023-146224OB-I00.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

560

REFERENCES

Arab, L. E. and D

ıaz, G. A. (2015). Impact of social

networks and internet in adolescence: strengths and

weaknesses. Revista M

edica Cl

ınica Las Condes,

26(1):7–13.

Bonetti, A., Mart

ınez-Sober, M., Torres, J. C., Vega, J. M.,

Pellerin, S., and Vila-Franc

es, J. (2023). Comparison

between Machine Learning and Deep Learning Ap-

proaches for the Detection of Toxic Comments on So-

cial Networks. Applied Sciences, 13(10):6038.

Buitrago L

opez, A., Pastor-Galindo, J., and Ruip

erez-

Valiente, J. A. How toxic is the Spanish information

environment? Exploring the sentimentalism and hate

speech in online news and public reactions.

Caselli, T., Basile, V., Mitrovi

c, J., and Granitzer, M.

(2021). HateBERT: Retraining BERT for abusive lan-

guage detection in English. In Proceedings of the 5th

Workshop on Online Abuse and Harms (WOAH 2021),

pages 17–25, Online. Association for Computational

Linguistics.

Chhabra, A. and Vishwakarma, D. K. (2023). A literature

survey on multimodal and multilingual automatic hate

speech identiﬁcation. 29(3):1203–1230.

Garg, T., Masud, S., Suresh, T., and Chakraborty, T. (2023).

Handling Bias in Toxic Speech Detection: A Survey.

ACM Comput. Surv., 55(13s):264:1–264:32.

Hanu, L. and Unitary, t. (2023). Detoxify.

Islam, M. H., Farzana, K., Khalil, I., Ara, S., Shazid, M. A.,

and Kabir Mehedi, M. H. (2021). Unmasking toxicity:

A comprehensive analysis of hate speech detection in

banglish. In 2024 6th International Conference on

Electrical Engineering and Information & Communi-

cation Technology (ICEEICT), pages 963–968. ISSN:

2769-5700.

Leite, J. A., Silva, D. F., Bontcheva, K., and Scarton, C.

(2020). Toxic Language Detection in Social Media for

Brazilian Portuguese: New Dataset and Multilingual

Analysis.

Mahajan, A., Shah, D., and Jafar, G. (2021). Explainable

AI approach towards toxic comment classiﬁcation. In

Hassanien, A. E., Bhattacharyya, S., Chakrabati, S.,

Bhattacharya, A., and Dutta, S., editors, Emerging

Technologies in Data Mining and Information Secu-

rity, pages 849–858. Springer Nature.

Maity, K., Poornash, A. S., Saha, S., and Bhattacharyya,

P. (2024). ToxVidLM: A multimodal framework for

toxicity detection in code-mixed videos.

Malik, P., Aggrawal, A., and Vishwakarma, D. K. (2021).

Toxic Speech Detection using Traditional Machine

Learning Models and BERT and fastText Embedding

with Deep Neural Networks. In 2021 5th Interna-

tional Conference on Computing Methodologies and

Communication (ICCMC), pages 1254–1259.

Proksch, S.-O., Lowe, W., W

ackerle, J., and Soroka, S.

(2019). Multilingual Sentiment Analysis: A New Ap-

proach to Measuring Conﬂict in Legislative Speeches.

Legislative Studies Quarterly, 44(1):97–131.

Rinker, T. (2016). sentimentr: Calculate Text Polarity Sen-

timent.

Sheth, A., Shalin, V. L., and Kursuncu, U. (2022). Deﬁning

and detecting toxicity on social media: context and

knowledge are key. Neurocomputing, 490:312–318.

Subramanian, M., Easwaramoorthy Sathiskumar, V.,

Deepalakshmi, G., Cho, J., and Manikandan, G.

(2023). A survey on hate speech detection and senti-

ment analysis using machine learning and deep learn-

ing models. Alexandria Engineering Journal, 80:110–

121.

Sentiment-Enriched AI for Toxic Speech Detection: A Case Study of Political Discourses in the Valencian Parliament

561