Irony and sarcasm represent complex pragmatic phenomena wherein the intended meaning diverges from the literal interpretation, often relying on contextual and cultural cues [7]. Their detection poses unique challenges to natural language processing (NLP) systems, particularly in short texts like news headlines, where linguistic economy intensifies ambiguity [2][3]. Early approaches to sarcasm detection used lexicon-based or rule-based systems, often relying on punctuation, emoticons, or polarity shifts [1][10]. While useful for exploratory purposes, such systems typically failed to generalize across domains or capture deeper semantic contradictions, prompting a shift toward machine learning (ML) and deep learning (DL) methods.

Recent years have seen a surge in the use of deep learning architectures for detecting irony and sarcasm, particularly with the advent of large-scale pre-trained transformer models such as BERT, RoBERTa, and XLNet [11]-[13]. These models, trained on massive text corpora, exhibit an improved ability to capture context, co-reference, and syntactic nuance. Studies using RoBERTa fine-tuned on sarcasm corpora have demonstrated F1-score improvements over traditional classifiers like SVMs and LSTMs [9]. Furthermore, attention mechanisms in transformer-based models allow for explainability, offering insights into which words contribute most to classification, a valuable feature when dealing with inherently ambiguous content like ironic headlines [14].

Irony and sarcasm are not merely stylistic features; they serve epistemic and ideological functions, especially in politically loaded discourse. Research shows that sarcastic news headlines can obscure factual clarity and foster misinterpretation, effectively serving as a conduit for misinformation [8]. Unlike blatant disinformation, sarcastic phrasing is often immune to conventional fact-checking, making it a subtler but potent threat to information integrity [15]. Moreover, ironic framing may allow content producers to evade accountability while maintaining plausible deniability, amplifying the challenges for automated misinformation detection systems [16].

Publicly available datasets such as the SemEval-2018 Task 3 corpus, the Sarcasm Corpus V2, and the iSarcasm dataset have facilitated benchmarking in this domain [17][18]. These corpora typically feature social media posts annotated for various forms of irony and sarcasm, with recent iterations incorporating multilingual and multimodal features. Nonetheless, few corpora explicitly focus on news headlines, a gap this study addresses by constructing a dataset specifically tailored to media discourse. The scarcity of sarcasm-labeled headline corpora continues to constrain generalizability, further reinforcing the need for domain-adapted models.

Sarcasm detection is often linked to sentiment analysis, given that ironic statements frequently exhibit polarity reversal where the surface sentiment contradicts the underlying intent [19]. Traditional sentiment analysis tools such as VADER [20], NRCLex [21], and SenticNet [22] have been applied to this task with limited success due to their context-agnostic nature. Nevertheless, combining sentiment lexicons with transformer embeddings has shown promise in identifying emotional incongruities that signal sarcasm [23]. Hybrid approaches incorporating both affective signals and contextual semantics are increasingly considered state-of-the-art.

Latent Dirichlet Allocation (LDA) and related topic modeling techniques have been used to uncover thematic structures in news content, including sarcasm-laced misinformation [24]. Such methods allow researchers to identify recurring frames or narratives embedded in ironic headlines, particularly those that reinforce ideological bias. Coupled with irony detection, topic modeling offers a dual lens through which to view both tone and content, revealing how sarcasm may strategically obscure intent or deflect critique [25].

While prior work has laid substantial groundwork in sarcasm detection, sentiment analysis, and misinformation studies, few integrate these threads within the context of news headlines a domain where irony often intersects with political messaging. By leveraging transformer-based architectures and sentiment-aware embeddings, our study builds on and advances this research tradition, offering methodological contributions and practical implications for combating misinformation through enhanced content moderation and automated fact-checking systems.

For both training and evaluation purposes, we employed a composite dataset merging the iSarcasm collection with an independently compiled set of 47,000 English news headlines. Although an initial pool of approximately 1.2 million headlines was screened during preliminary data harvesting, only the 22,097 carefully validated and labeled headlines were used in model training, evaluation, and all results reported in this study. These headlines were annotated for irony and sarcasm through a hybrid approach combining expert judgments and consensus-driven crowdsourcing, with each entry labeled to indicate the presence or absence of these rhetorical devices. To mitigate class imbalance (where ironic instances accounted for 34.7% and non-ironic ones for 65.3%) we implemented balancing techniques.

Unlike social media corpora such as SemEval-2018 Task 3 [17], our compilation exclusively targeted journalistic headlines from verified media sources, ensuring closer relevance to misinformation propagated via editorial content. Ambiguous cases were systematically removed by enforcing an inter-rater reliability threshold (Cohen’s κ > 0.72) to maintain annotation consistency.

The iSarcasm dataset itself, comprising roughly 10,000 tweets marked for deliberate sarcasm and enriched with contextual replies, contributed social media context. Despite originating from tweets, its annotation scheme (emphasizing speaker intent and sarcasm targets) proved adaptable for analyzing headline-level text, especially those exhibiting subjective or ironic nuances. Complementing this, we assembled a collection of 12,346 headlines spanning political, economic, and health topics from 40 global digital outlets, annotated through a rigorous two-stage process yielding a strong Fleiss’ κ of 0.79. This subset included 4232 ironic or sarcastic headlines and 8114 neutral ones.

Covering the period from January 2020 through December 2022, this dataset was curated to focus on areas prone to misinformative irony [25], with purely factual or non-subjective headlines filtered out. To broaden the semantic range and improve class variability without compromising coherence, we applied paraphrasing augmentation techniques leveraging PEGASUS [26].

For classical machine learning algorithms, we converted the text into TF-IDF feature vectors using the Scikit-learn library, which produces a sparse representation based on weighted n-gram occurrences. In contrast, deep learning models utilized pretrained tokenizers to generate contextual token embeddings. Specifically, BERT employed WordPiece tokenization [11], RoBERTa used Byte-Pair Encoding [12], and XLNet relied on SentencePiece tokenization [13]. All tokenizers were accessed through the HuggingFace Transformers repository, enabling seamless integration with the respective models.

We developed and evaluated six distinct classifiers to detect sarcasm and irony in a binary setting. The traditional machine learning techniques included logistic regression, support vector machines, and random forests. Alongside these, three transformer-based architectures (BERT, RoBERTa, and XLNet) were fine-tuned for the same task. Although RoBERTa performed competitively, XLNet consistently outperformed all models across all folds; hence, the analyses presented in the Results and Discussion focus on XLNet. To address class imbalance, we applied stratified sampling and supplemented the minority class using SMOTE. Model performance was primarily assessed through the ROC-AUC metric, calculated via scikit-learn’s evaluation tools.

Following [27], we used ROC-AUC as our primary evaluation metric, calculated using Equations (1) through (3) below:

We present mean ROC-AUC scores across folds in Table 1.

Table 1. Mean ROC-AUC scores across classifiers for sarcasm detection.

As shown, XLNet achieved the highest accuracy, particularly in detecting ironic headlines, outperforming other models by a significant margin (Price et al., 2020). Its performance aligns with prior findings on XLNet’s autoregressive context modeling superiority in irony detection [13].

To uncover the thematic patterns within sarcastic and non-sarcastic news headlines, we applied Latent Dirichlet Allocation (LDA) via the Gensim library. Prior to modeling, terms occurring in fewer than 40 headlines or in over 90% of the documents were excluded to enhance topic quality. We then fine-tuned the model’s hyperparameters (including the number of topics (k) and the Dirichlet priors (α and η)) using a grid search aimed at maximizing Cv coherence scores. Distinct LDA models were developed separately for sarcastic and non-sarcastic headline sets. For the sarcastic subset, we identified 10 coherent topics, with examples of the top three presented in Table 2.

Table 2. Identified coherent topics.

These sarcastic framings commonly mirror or mock conspiracy theories, often embedding rhetorical distortion that obfuscates the factual grounding of the headline [8].

Tools and Software

Deep Learning: HuggingFace Transformers v4.32, PyTorch v2.0Preprocessing: NLTK v3.8, spaCy v3.5Visualization: Matplotlib v3.6, Seaborn v0.12Topic Modeling: Gensim v4.3Sentiment Analysis: NRCLex v3.0

All experiments were executed on NVIDIA RTX A6000 GPUs under Ubuntu 22.04 LTS. Code and datasets are archived on OSF for transparency.

This data pipeline integrates neural and symbolic methods to detect irony in the context of news-based misinformation. Results from model evaluation and topic decomposition lay the groundwork for practical applications in content moderation, editorial verification, and digital literacy tools.

Our best-performing deep learning model achieved an overall accuracy of 82.47% in distinguishing ironic or sarcastic headlines from neutral or straightforward ones, surpassing baseline lexicon-based methods such as VADER and TextBlob, which scored 65.32% and 61.75%, respectively. Within the dataset, 29.1% of headlines were labelled as containing sarcasm, while 21.7% exhibited irony without overt sarcasm. The remaining 49.2% were classified as neutral or literal.

Table 3 summarizes the precision, recall, and F1-scores for the XLNet-based architecture compared to baseline models across irony and sarcasm categories.

Table 3. Performance metrics of deep learning and baseline models on irony and sarcasm detection.

Sarcasm detection proved more challenging, as indicated by a lower recall rate (0.76) relative to irony (0.82). This aligns with the subtlety and contextual dependency inherent in sarcastic expression [28].

The analysis revealed four recurring sarcasm archetypes:

1) Hyperbolic Praise—exaggerated positivity directed toward a clearly negative situation;

2) Contradictory Juxtaposition—pairing mutually incompatible claims to implicitly expose absurdity;

3) Deadpan Literalism—presenting an obviously false statement in a flat, factual tone;

4) Ironic Reversal—stating the opposite of the intended meaning to highlight hypocrisy or failure;

5) These archetypes were derived through manual inspection of high-confidence XLNet predictions and were consistently observed across major news categories.

We further analyzed temporal patterns in the occurrence of ironic and sarcastic headlines over a six-month period encompassing major political and social events. Table 2 shows a pronounced increase in sarcasm-related headlines coinciding with politically charged episodes, such as elections and legislative controversies, suggesting heightened use of sarcasm to frame contentious issues.

Topic modeling via Latent Dirichlet Allocation (LDA) revealed thematic clusters strongly associated with ironic and sarcastic headlines. Sarcastic headlines often centered around political hypocrisy, media bias, and celebrity scandals, while ironic headlines frequently addressed unexpected or paradoxical developments in international affairs and economics (see Table 4). These clusters support the notion that irony and sarcasm serve distinct discursive functions in news framing [1].

Table 4. Topical clusters identified in ironic and sarcastic headlines.

To complement detection results, we generated sentiment profiles for ironic and sarcastic headlines using the NRC Emotion Lexicon [21]. Sarcastic headlines exhibited elevated anticipation, disgust, and surprise compared to irony and neutral categories, which aligns with sarcasm’s function as a form of social critique [29]. Irony showed a more balanced emotional profile with moderate joy and sadness.

Our deep learning models also highlighted specific syntactic markers and discourse cues as significant predictors of sarcasm, including the presence of intensifiers (e.g., “totally”, “absolutely”), unexpected conjunctions (“but”, “yet”), and punctuation cues such as exclamation points or quotation marks. Additionally, lexical items like “yeah”, “right”, and interjections were frequently detected in sarcastic headlines, consistent with prior research on pragmatic markers [30].