MFTCXplain: A Multilingual Benchmark Dataset for Evaluating the Moral Reasoning of LLMs through Hate Speech Multi-hop Explanation
Jackson TragerUniversity of Southern California, Francielle VargasUniversity of São PauloDiego AlvesSaarland UniversityMatteo GuidaUniversity of MelbourneMikel K. NgueajioHoward UniversityAmeeta AgrawalPortland State UniversityFlor Plaza-del-ArcoLeiden UniversityYalda DaryanaiUniversity of Southern · 2025
Read MFTCXplain: A Multilingual Benchmark Dataset for Evaluating the Moral Reasoning of LLMs through Hate Speech Multi-hop Explanation by Jackson TragerUniversity of Southern California, Francielle VargasUniversity of São PauloDiego AlvesSaarland UniversityMatteo GuidaUniversity of MelbourneMikel K. NgueajioHoward UniversityAmeeta AgrawalPortland State UniversityFlor Plaza-del-ArcoLeiden UniversityYalda DaryanaiUniversity of Southern (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-19073.txt.
Introduction
Jackson Trager and co-authors release MFTCXplain, a benchmark for testing the moral reasoning of large language models. Their 2025 paper notes two gaps in older tests: labels without stated reasons, and a heavy focus on English. The new dataset uses hate speech posts and asks models to explain their calls step by step.
The set holds three thousand posts in Portuguese, Italian, Persian, and English. Each post carries a hate speech label, moral categories from Moral Foundations Theory, and marked text spans giving the reason. Models detected hate speech well but predicted moral sentiments poorly, scoring under 0.35 on the standard measure. Reason matching was weakest in the less represented languages.
Catalog record
| CREATOR | Jackson TragerUniversity of Southern California, Francielle VargasUniversity of São PauloDiego AlvesSaarland UniversityMatteo GuidaUniversity of MelbourneMikel K. NgueajioHoward UniversityAmeeta AgrawalPortland State UniversityFlor Plaza-del-ArcoLeiden UniversityYalda DaryanaiUniversity of Southern |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 90 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-19073
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
MFTCXplain: A Multilingual Benchmark Dataset for Evaluating the Moral Reasoning of LLMs through Hate Speech Multi-hop Explanation Jackson TragerUniversity of Southern California, Francielle VargasUniversity of São PauloDiego AlvesSaarland UniversityMatteo GuidaUniversity of MelbourneMikel K. NgueajioHoward UniversityAmeeta AgrawalPortland State UniversityFlor Plaza-del-ArcoLeiden UniversityYalda DaryanaiUniversity of Southern CaliforniaFarzan Karimi-MalekabadiUniversity of Southern California Abstract Ensuring the moral reasoning capabilities of Large Language Models (LLMs) is a growing concern as these systems are used in socially sensitive tasks. Nevertheless, current evaluation benchmarks present two major shortcomings: a lack of annotations that justify moral classifications, which limits transparency and interpretability; and a predominant focus on English, which constrains the assessment of moral reasoning across diverse cultural settings. In this paper, we introduce MFTCXplain, a multilingual benchmark dataset for evaluating the moral reasoning of LLMs via hate speech multi-hop explanation using Moral Foundation Theory (MFT). The dataset comprises 3,000 tweets across Portuguese, Italian, Persian, and English, annotated with binary hate speech labels, moral categories, and text span-level rationales. Empirical results highlight a misalignment between LLM outputs and human annotations in moral reasoning tasks. While LLMs perform well in hate speech detection (F1 up to 0.836), their ability to predict moral sentiments is notably weak (F1 < 0.35). Furthermore, rationale alignment remains limited mainly in underrepresented languages. These findings show the limited capacity of current LLMs to internalize and reflect human moral reasoning. 0 0 footnotetext: Warning : This document contains offensive content. 1 Introduction LLMs perform tasks that rely on moral reasoning, from moderating harmful content to providing an ethical advisor ( Dillion et al. 2025 ) . However, recent research has raised concerns about the robustness, explainability, and alignment of LLMs with diverse human values ( Oh and Demberg 2025 ; Huang et al. 2025 ; Agarwal et al. 2024 ; Atari et al. 2023b ; Abdurahman et al. 2024 ) , as they often exhibit inconsistent behavior, lack transparency, and reflect culturally contingent moral frameworks. Figure 1: Hate speech multi-hop explanation. One domain where moral reasoning is both prominent and consequential is in hate speech. People often use moral language around purity, loyalty, and authority, to justify incivility or portray harm to the outside world as righteous ( Kennedy et al. 2023 ; Fiske and Rai 2014 ) . These moral framings of hate trigger moral outrage, increase engagement, and reinforce group loyalty ( Brady et al. 2020 ; Abdurahman et al. 2025 ) . Measuring this type of moral reasoning is crucial, as different moral appeals shape hate in culturally and psychologically distinct ways ( Graham et al. 2013 ; Hoover et al. 2021 ) . Despite advances in hate speech detection in terms of linguistic and social bias ( Davani et al. 2023 ; Plaza-del arco et al. 2023 ; Vargas et al. 2021 ; Davidson et al. 2019 ) , models often remain “black-box”, producing classifications without revealing the underlying reasoning ( Kennedy et al. 2020 ) . In addition, most existing benchmarks are English-only, constraining the development of models capable of culturally grounded moral reasoning ( Buscemi et al. 2025 ) . To address these challenges, we introduce MFTCXplain 1 1 1 The dataset and code, and results are available at https://github.com/franciellevargas/MFTCXplain , ( M oral F oundations T witter C orpus with E X planations), the first multilingual benchmark dataset for evaluating moral reasoning in LLMs using expert-annotated rationales based on Moral Foundations Theory ( Graham et al. 2013 ) . The dataset comprises 3,000 tweets in four languages—Portuguese, Persian, Italian, and English—annotated for hate speech, fine-grained moral categories, and text spans (rationales) that justify each moral label. Notably, the Persian and Brazilian Portuguese portions represent the Global South, low-resource contexts often neglected in NLP research, while Italian, though not from the Global South, remains understudied. Annotations were conducted by native speakers with diverse cultural backgrounds, and detailed annotator metadata is provided to support bias analysis ( Prabhakaran et al. 2021 ; Davani et al. 2024 ) . While multi-hop explanation has traditionally been studied in question answering ( Jhamtani and Clark 2020 ; Valentino et al. 2021 ; Thayaparan et al. ) and fact verification ( Ma et al. 2024 ) , we draw inspiration from this paradigm to frame moral explanation as a multi-hop reasoning process, as shown in Figure 1 . This requires LLMs to connect expressions of hate with moral principles through intermediate inferences, such as identifying hate speech (offensive content against targeted groups) - hop 1; recognizing and mapping moral violations - hop 2; and their rationales - hop 3; thus requiring multi-level moral reasoning and compositional inference. To assess current model capabilities, we benchmark state-of-the-art LLMs including GPT-4o and LLaMA 3.3 70B using zero-shot, few-shot, and Chain-of-Thought (CoT) prompting. Our results reveal substantial divergence between human and model-generated explanations and low macro F1 scores (often below 0.35) when predicting moral sentiment, highlighting the challenges LLMs face in aligning with culturally human moral reasoning. The dataset, including annotation agreements/disagreements, model outputs, and code, is publicly available to facilitate future research. Our contributions are summarized as follows: (i) We provide the first expert-annotated multilingual dataset of hate speech multi-hop explanation for evaluating LLMs’ moral reasoning; (ii) We propose a novel annotation schema based on Moral Foundations Theory to enhance the interpretability of hate speech explanations by multi-hop explanation; (iii) We conduct the first multilingual study on explainable moral reasoning in LLMs, revealing insights across cultures. 2 Related Work Authors Datasets Languages # Instances Platforms Models F-score Johnson and Goldwasser 2018a Non-Name EN 2,050 Twitter SVM 0.72 Hoover et al. 2020a MFTC EN 35,108 Twitter SVM, LSTM 0.80 Trager et al. 2022 MFRC EN 16,123 Reddit SVM, BERT 0.76 Beiró et al. 2023 Non-Name EN 4,811 Facebook LSTM, Regression 0.82 Table 1: Overview of annotated datasets with Moral Foundation Theory (MFT). Authors Datasets Languages # Instances Platforms Models F-score Mathew et al. 2021 HateXplain EN 20,148 Gab and Twitter CNN-GRU, BiRNN, BERT 0.69 Pavlopoulos et al. 2021 Non-Name EN 10,629 Civil Comments dataset BERT 0.71 Ravikiran and Annamalai 2021 DOSA TA-EN 4786 YouTube BERT 0.40 KN-EN 1097 Hoang et al. 2023 ViHOS VI 11,056 Facebook, YouTube BiLSTM, BERT 0.78 Delbari et al. 2024a PHATE FA 7,000 Twitter BERT, GPT 0.78 Salles et al. 2025 HateBRXplain PT 7,000 Instagram BERT, DistilBERT, PTT5 0.91 Table 2: Overview of datasets with human rationales annotation for explainable hate speech detection. 2.1 Moral Foundation Theory Annotated Datasets Previous works have created datasets annotated with MFT (See Table 1 ). The Moral Foundations Twitter Corpus (MFTC) Hoover et al. 2020b includes over 35,000 tweets annotated with the five MFT pairs proposed by Graham et al. 2013 : care/harm, fairness/cheating, loyalty/betrayal, authority/subversion, and purity/degradation. Similarly, the Moral Foundations Reddit Corpus (MFRC) Trager et al. 2022 contains approximately 16,000 Reddit comments annotated across eight moral dimensions based on the updated MFT proposed by Atari et al. 2023a . Johnson and Goldwasser 2018b introduced a dataset of over 2,000 politician tweets labeled for the original five MFT pairs. Additional work during the COVID-19 pandemic extended moral analysis Beiró et al. 2023 , examining 500,000 Facebook posts through the lens of Liberty, Care, and Authority, while Pacheco et al. 2022 analyzed 750 annotated tweets linking vaccine attitudes to