EHLD: a general data framework for harmful language detection and evaluation

Purpose Research into harmful language detection is hindered by fragmented data sets and incompatible label schemas, which significantly limit evaluation. This paper aims to introduce evaluation of harmful language detection (EHLD): a general, extensible data framework supporting the integration, enrichment and evaluation of diverse harmful language resources, with a particular focus on Italian. Design/Methodology/Approach EHLD defines a unified schema with essential detection attributes (e. g. harmfulₒrₙot and harmₛubcategory), optional contextual dimensions (e. g. target, category and intersectionality), evaluation-oriented text complexity features and metadata for provenance. The authors’ instantiate EHLD by integrating three data sources: curated data sets from the literature, large-scale large language model (LLM) -based annotation from the Mappa dell’Intolleranza and LLM-generated data for inclusive language detection. They further introduce labelconfidence to explicitly encode label reliability. Model selection follows an iterative cycle combining distributional analysis, linguistic diversity assessment and performance evaluation. Findings The resulting data set contains 237, 956 instances with 18 features and exhibits substantial linguistic variety (Self-bilingual evaluation understudy drops from 0. 759 at 1 gram to 0. 107 at 4 grams). After iterative, data-centric refinement, generative pretrained transformer-5-nano is selected as the reference model and achieves an F1-score of 0. 813 on binary harmful language detection, outperforming reported baselines such as bidirectional encoder representations from transformers (BERT) -based multitask models and other LLMs on comparable Italian settings. Research limitations/implications Due to differences in data sets, domains, label definitions and evaluation protocols, comparisons across studies are not fully controlled. The framework partly relies on automatically generated labels, which, despite confidence-aware handling, may introduce residual noise. Practical implications By combining standardized labels, provenance tracking and complexity-oriented evaluation features, EHLD supports the construction of reproducible data sets and richer model auditing, enabling a more transparent deployment of harmful language detection systems. Originality/value EHLD contributes a reusable, confidence-aware integration framework that connects different harmful language resources and allows iterative, data-driven model selection and evaluation, with a focus on Italian.