Amala3/iRONNIE_EN_evalset

Dataset Card for the iRONNIE data collection

This dataset was used to assess how well generative models can produce irony explanations for English and Dutch. As opposed to other work on irony and sarcasm explanation, this work does not make use of paraphrases but instead aims to produce explanations that are grounded in world knowledge (social, factual or any other) that is needed to understand the text but that may not be explicitly mentioned. The collection contains both fine-tuned models as well as the train and test data used in the experiments. Dataset access is based on manual requests to guard the data from being collected in large-scale scraping. The data should be used for concious and aware training and evaluation, since the goal is to carefully assess what models are capable of or not.

Collection Overview

Datasets

• iRONNIE_train: includes the gold standard explanations of the train set for both English and Dutch.
• iRONNIE Knowledge: includes the gold standard irony explanations for English and Dutch, as well as the gold standard knowledge items.
• iRONNIE_EN_evalset: includes the gold standard irony explanations for English and the AI-generated irony explanations on the test set.
• iRONNIE_NL_evalset: includes the gold standard irony explanations for Dutch and the AI-generated irony explanations on the test set.

Models

• IronyExplainer_mixed: Llama 3 model fine-tuned for irony explanation on both English and Dutch (scrambled dataset)
• IronyExplainer_Llama3_EN: Llama3 model fine-tuned for irony explanation on only English.
• IronyExplainer_NL_monolingual: Llama3 model fine-tuned for irony explanation on only Dutch.
• IronyExplainer_ENthenNL: Llama 3 model that is sequentially fine-tuned for irony explanation, first on English and then on Dutch (scrambled dataset).

Source Data

The tweet texts of this dataset are collected by Van Hee et al., 2018 for the task of irony detection. The English tweets in the dataset were used for the SemEval-2018 Task 3. Both the English and the Dutch tweets are also used for fine-grained (confidence-aware) annotation and trigger word annotation as part of TRIC: a confidence-aware multilingual dataset for irony detection and rationales in English, Dutch and Italian.

Annotation Procedure

The guidelines and details about the annotation procedure are documented as a technical report

Publications:

For the English materials: Maladry et al. (forthcoming, accepted at CAS) Understanding Irony through Explanations and Background. (link to be added upon publication) Knowledge.

For the Dutch materials: Maladry et al. (forthcoming, accepted at LREC2026) Exploring the Transfer of Irony Explanation Generation from English to Dutch. (link to be added upon publication)

Contact:

If you would like access to additional models or data. Feel free to reach out to aaron.maladry@gmail.com or aaron.maladry@ugent.be