MOKA: Moral Knowledge Augmentation for Moral Event Extraction

TL;DR

This work defines a new task, Moral Event Extraction, to recover fine-grained morality in news by modeling events and their participants through moral foundations. It introduces MORAL EVENTS, a dataset of 5,494 morality-bearing event annotations across 474 articles, and MOKA, a retrieval-augmented framework that injects external moral knowledge via a Morality Bank and Moral Scenario Banks. The approach yields consistent improvements over strong baselines on trigger detection and moral inference, and reveals systematic left-right biases in how outlets report moral events. By enabling structure-aware moral reasoning in NLP models, MOKA offers a principled way to study morality in news and its relation to media ideology, with potential applications in media analysis and bias detection.

Abstract

News media often strive to minimize explicit moral language in news articles, yet most articles are dense with moral values as expressed through the reported events themselves. However, values that are reflected in the intricate dynamics among participating entities and moral events are far more challenging for most NLP systems to detect, including LLMs. To study this phenomenon, we annotate a new dataset, MORAL EVENTS, consisting of 5,494 structured event annotations on 474 news articles by diverse US media across the political spectrum. We further propose MOKA, a moral event extraction framework with MOral Knowledge Augmentation, which leverages knowledge derived from moral words and moral scenarios to produce structural representations of morality-bearing events. Experiments show that MOKA outperforms competitive baselines across three moral event understanding tasks. Further analysis shows even ostensibly nonpartisan media engage in the selective reporting of moral events. Our data and codebase are available at https://github.com/launchnlp/MOKA.

Figures (6)

• Figure 1: Sample moral event extractions (MEEs) for a target sentence from our Moral Events dataset. Event participants are annotated per Wikipedia pages if applicable. In each event record, the event trigger is a single word in an event span, and it might embody multiple moralities. Moral event extraction is challenging due to several reasons: implicit participants (e.g. same-sex couples in Event Record 1) may not be mentioned in the target sentence, and understanding the relations among the participants is necessary to correctly infer the morality.
• Figure 2: Overview of MOKA for (downstream) moral event extraction. It highlights the process of retrieving and combining relevant scenarios, and the integration of moral word knowledge through attention-based retrieval. Embeddings in Lexicon are colored in red if moral words embody Harm, or blue if Care. "SCOTUS" is an acronym for "Supreme Court of the United States". $\{\}$ indicates there can be multiple answers.
• Figure 3: Number of moral events in each 100-word segment. Highly partisan media outlets tend to include more moral language than non-partisan ones.
• Figure 4: Employed moral foundation distribution by media outlets of different ideologies.
• Figure A1: Correlation among agent-patient relationships, media outlet ideologies, and Care-/Harm-bearing moral events. Each percentage indicates the proportion of reporting a certain agent-patient interaction, and each column sums up to $100\%$. For example, $6.2\%$ means that, among all Care-bearing events reported by left-leaning media, $6.2\%$ of them are enabled by a Left-leaning entity and affecting a Right-leaning entity.