Fine Grained and Multi Target Offensive Language Detection

124 concepts117 papers1 proposal

Early attempts at automated content moderation treated offensive language detection as a simple binary task: a text is either toxic or safe. However, real-world social media discourse is rarely so black-and-white. An insult directed at an individual operates differently from systemic hate speech targeting a marginalized group, and profane language used casually differs fundamentally from targeted harassment. To build machine learning models that truly understand online abuse, researchers have shifted toward fine-grained, multi-target, and hierarchical annotation frameworks.

Hierarchical Taxonomies and Target Identification

Instead of collapsing all problematic text into a single category, modern frameworks break the problem down into structured layers. A foundational example of this approach is the Offensive Language Identification Dataset (OLID) and its semi-supervised extension, SOLID. As explored in Fermi at SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media using Sentence Embeddings by Vijayasaradhi Indurthi et al. and TECHSSN at SemEval-2020 Task 12: Offensive Language Detection Using BERT Embeddings by Rajalakshmi Sivanaiah et al., this taxonomy structures the task into three progressive sub-tasks:

  1. Detection: Is the text offensive or not?
  2. Categorization: Is the offense targeted at a specific entity, or is it untargeted profanity?
  3. Target Identification: Is the target an individual, a group, or another entity?

By decoupling the presence of offensive language from its intent and target, models can learn more nuanced representations. For instance, a model can recognize that a tweet containing a swear word might be untargeted venting rather than a malicious attack. This hierarchical strategy has been adapted globally to capture complex intersections of hate speech, such as in Portuguese (A Hierarchically-Labeled Portuguese Hate Speech Dataset by Paula Fortuna et al.) and Arabic (AraTar: A Corpus to Support the Fine-grained Detection of Hate Speech Targets in the Arabic Language by Seham Alghamdi et al.), allowing for fine-grained multi-label classification of both the type of hate and its specific victims.

Mapping the Spectrum of Intensity and Subtlety

Offensiveness is not merely categorical; it is also a matter of degree and subtlety. Some language is explicitly vulgar, while other forms of harm are deeply implicit or cloaked in polite phrasing.

To capture the continuous nature of abuse, researchers have integrated intensity ratings into datasets. For example, the authors of At the Lower End of Language—Exploring the Vulgar and Obscene Side of German (Elisabeth Eder et al.) developed the VULGER lexicon, which replaces discrete labels with continuous obscenity ratings derived from human assessments. Similarly, in low-resource settings like Amharic, Abinew Ali Ayele et al. introduced graded intensity scales in Exploring Boundaries and Intensities in Offensive and Hate Speech: Unveiling the Complex Spectrum of Social Media Discourse to map the transition from mild offense to severe hate speech.

Beyond explicit vulgarity, models must also detect subtle harms like Patronizing and Condescending Language (PCL). As highlighted in Don't Patronize Me! An Annotated Dataset with Patronizing and Condescending Language towards Vulnerable Communities by Carla Perez-Almendros et al., PCL often targets vulnerable communities under the guise of pity or superiority. Because PCL lacks obvious toxic keywords, detecting it requires deep contextual and world knowledge, prompting specialized datasets to train transformers on these highly nuanced linguistic structures.

Embracing Annotator Disagreement and Sensitivity

A major bottleneck in training toxic language classifiers is that human annotators frequently disagree. Traditional machine learning pipelines treat disagreement as noise, resolving it via majority voting to create a single "ground truth." However, recent work suggests that annotator disagreement is a valuable signal reflecting real-world opinion polarization.

To address this, the authors of Polarized Opinion Detection Improves the Detection of Toxic Language (John Pavlopoulos et al.) proposed "unpolarized learning," introducing a metric called nDFU to quantify opinion polarization directly from raw annotations. By treating highly disputed comments as a distinct class rather than forcing a consensus, models achieve better generalization and robustness.

Furthermore, the data collection process itself introduces subtle biases. In Order Effects in Annotation Tasks: Further Evidence of Annotation Sensitivity, Jacob Beck et al. demonstrated the existence of "order effects," showing that human annotators become less likely to label a tweet as offensive if they encounter it later in a long annotation session. Recognizing these sensitivities is crucial for designing cleaner benchmarks and understanding model uncertainty.

The Multilingual and Cultural Imperative

Offensive language is deeply tied to local idioms, political climates, and regional vocabulary. A model trained purely on English data fails to capture the unique dynamics of global social media discourse.

Consequently, recent research focuses on building localized, high-fidelity datasets. For example, Ayme Arango Monnar et al. developed a manually annotated dataset capturing regional Chilean vocabulary in Resources for Multilingual Hate Speech Detection, while Hrushikesh Patil et al. built a large-scale Marathi hate speech corpus in L3Cube-MahaHate: A Tweet-based Marathi Hate Speech Detection Dataset and BERT Models.

Gathering this data efficiently requires creative heuristics to avoid keyword bias. For instance, in Arabic Offensive Language on Twitter: Analysis and Experiments, Ahmed Abdelali et al. leveraged the Arabic vocative particle "ya" (يا) as a structural signal to collect a massive, unbiased corpus of offensive Arabic tweets, demonstrating how language-specific syntactic features can optimize data curation.

Key Foundational Work

L-HSAB: A Levantine Twitter Dataset for Hate Speech and Abusive Language (2019) — contributed 2 concept elements to this direction.
Fermi at SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media using Sentence Embeddings (2019) — contributed 1 concept elements to this direction.
SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval) (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_offenseval_dataset_278f340d,
  title = {OffensEval dataset},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/offenseval-dataset/rdr_278f340d473a472691b5392c0dc492bf},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_offenseval_dataset_278f340d_v1,
  title = {OffensEval dataset},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/offenseval-dataset/rdr_278f340d473a472691b5392c0dc492bf/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake