[Paper Review] A Survey in Automatic Irony Processing: Linguistic, Cognitive, and Multi-X Perspectives

Qingcheng Zeng, Anran Li|arXiv (Cornell University)|Sep 10, 2022

Topic Modeling4 citations

TL;DR

This paper presents a comprehensive survey of automatic irony processing, integrating linguistic theories, cognitive science insights, and recent advances in NLP, including deep learning and multimodal approaches. It identifies key challenges in irony detection, proposes new tasks inspired by sarcasm research, and advocates for explainable, multi-X, and domain-adaptive frameworks to advance computational irony understanding beyond current detection-focused models.

ABSTRACT

Irony is a ubiquitous figurative language in daily communication. Previously, many researchers have approached irony from linguistic, cognitive science, and computational aspects. Recently, some progress have been witnessed in automatic irony processing due to the rapid development in deep neural models in natural language processing (NLP). In this paper, we will provide a comprehensive overview of computational irony, insights from linguistic theory and cognitive science, as well as its interactions with downstream NLP tasks and newly proposed multi-X irony processing perspectives.

Motivation & Objective

• To provide a systematic review of computational irony processing, spanning linguistic theories, cognitive science, and NLP advancements.
• To address the imbalance in figurative language research by highlighting underexplored irony research compared to sarcasm.
• To identify gaps in current irony detection systems, particularly the lack of theoretical grounding and cognitive-informed design.
• To propose new research directions, including explainable irony processing, multiagent dialogue systems, and multimodal irony understanding.
• To encourage interdisciplinary research between linguistics and human language technology through unified frameworks for figurative language processing.

Proposed method

• Surveying theoretical foundations of irony from relevance theory, echoic use, and psychological reversal processes.
• Reviewing irony and sarcasm datasets across world languages, analyzing annotation schemes and their limitations.
• Tracing the evolution of irony detection from traditional machine learning and RNNs to modern pre-trained language models (PLMs).
• Integrating insights from sarcasm research—such as target identification, intended vs. perceived irony, and explanation tasks—into irony processing.
• Proposing multiagent irony processing for dialogue systems, where robots understand and generate irony in human-like interactions.
• Applying explainable AI techniques (SHAP, LIME) to interpret irony detection decisions, focusing on linguistic cues like punctuation and strong words.

Experimental results

Research questions

• RQ1 How do linguistic and cognitive theories of irony differ from one another, and how can they inform computational models?
• RQ2 Why has irony research remained underexplored compared to sarcasm in NLP, and what are the consequences for figurative language understanding?
• RQ3 To what extent do pre-trained language models improve irony detection, and what are their limitations in capturing contextual incongruities?
• RQ4 How can sarcasm-specific tasks—such as target identification and intended vs. perceived irony—be adapted to advance irony processing?
• RQ5 What role can explainable AI and multi-X perspectives (e.g., multimodal, multiagent) play in improving the robustness and interpretability of irony detection systems?

Key findings

• Pre-trained language models (PLMs) show improved performance on irony detection but still fall short of human-level understanding, particularly in capturing subtle psychological and contextual cues.
• Punctuation and emotionally strong words are significant predictors in irony detection, as revealed by SHAP and LIME explainability methods.
• The distinction between intended and perceived irony—previously studied in sarcasm—has not been adequately addressed in irony research, indicating a critical gap.
• Sarcasm detection datasets and tasks, such as target identification and explanation, offer promising templates for advancing irony processing, especially in multimodal and dialogue settings.
• Current irony detection systems remain largely isolated from theoretical frameworks in linguistics and cognitive science, limiting their interpretability and generalization.
• Multiagent irony processing in dialogue systems enhances user experience, suggesting that irony generation and understanding should be integrated into interactive AI systems.