Source-linked AI summary
SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval)
Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, Ritesh Kumar
TL;DR
OffensEval addresses the need to identify offensive language while distinguishing its type and target, using the purpose-built OLID dataset and a three-level task structure. Across a highly popular shared task, the reported results indicate that ensembles and state-of-the-art deep-learning models such as BERT were among the best-performing approaches.
Problem
Previous offensive-language research had not studied both the type and target of offensive language together, despite the need to automate labor-intensive and potentially harmful manual filtering.
Method
The paper organizes OffensEval around OLID’s hierarchical three-level annotations, using separate sub-tasks for offensive-language identification, offense-type categorization, and offense-target identification.
Results
The task attracted about 800 registered teams and 115 result submissions, with the best systems using ensembles and state-of-the-art deep-learning models such as BERT.
Takeaways & Limitations
OffensEval provides a shared benchmark and OLID resource for offensive-language identification across offense presence, type, and target.
Abstract
from arXiv · showhide
We present the results and the main findings of SemEval-2019 Task 6 on Identifying and Categorizing Offensive Language in Social Media (OffensEval). The task was based on a new dataset, the Offensive Language Identification Dataset (OLID), which contains over 14,000 English tweets. It featured three sub-tasks. In sub-task A, the goal was to discriminate between offensive and non-offensive posts. In sub-task B, the focus was on the type of offensive content in the post. Finally, in sub-task C, systems had to detect the target of the offensive posts. OffensEval attracted a large number of participants and it was one of the most popular tasks in SemEval-2019. In total, about 800 teams signed up to participate in the task, and 115 of them submitted results, which we present and analyze in this report.
1 Introduction
OffensEval addresses automated offensive-language detection by introducing OLID and a hierarchical annotation scheme that captures both offense type and target. It defines three sub-tasks covering offensive-language identification, offense categorization, and target identification, attracting 800 registered teams and 115 result submissions.
- Motivation: OLID uses a hierarchical three-level annotation schema that jointly represents the type and target of offensive language.The authors note that earlier work had not studied both dimensions together.
- Task design: OffensEval defines three sub-tasks corresponding to the annotation hierarchy.Sub-task A identifies offensive language, sub-task B categorizes offense types, and sub-task C identifies offense targets.
- Paper scope: The paper presents the task, dataset, participating systems, results, and concluding directions for future work.Its sections cover prior work, task description, OLID, system results, and conclusions.
- Participation: 800 teams signed up for OffensEval, while 115 submitted results.The paper analyzes the results of the teams that submitted systems.
2 Related Work
Prior work addressed several forms of abusive language, including aggression, bullying, hate speech, offensive language, and toxic comments. OffensEval builds on these efforts with OLID’s hierarchical model, which connects related phenomena while adding offense type and target distinctions.
- Existing tasks: Earlier research studied aggression, cyberbullying, hate speech, offensive language, and toxic comments using varied datasets and task definitions.Examples include Facebook posts, Twitter datasets, German tweets, and Wikipedia comments.
- Hate speech: Hate-speech identification was described as the most studied abusive-language detection task.The cited hate-speech dataset contained over 24,000 English tweets labeled non-offensive, hate speech, or profanity.
- Related shared tasks: GermEval included binary offensive-language identification and a second task separating profanity, insult, and abuse, but OffensEval uses a three-level hierarchy and different second-level labels.The passage identifies three differences, although the supplied text truncates the third.
- OLID’s contribution: OLID’s hierarchical annotation model captures commonalities across abuse categories, such as individual insults associated with cyberbullying and group insults associated with hate speech.The authors propose OLID as a resource for various offensive-language identification tasks.
3 Task Description and Evaluation
OffensEval uses OLID’s three annotation layers as three linked classification tasks, evaluated with macro-averaged F1 because class distributions are highly imbalanced. The task attracted nearly 800 teams, and participant materials included result tables and confusion matrices.
- Task framework: OLID supplies the training and testing material, and its three annotation layers define the three OffensEval sub-tasks.The dataset was built specifically for the task.
- Sub-task A: Sub-task A distinguishes offensive from non-offensive posts, including insults, threats, and untargeted profanity.The labels are Not Offensive (NOT) and Offensive (OFF).
- Sub-task B: Sub-task B predicts offense type for posts labeled Offensive in sub-task A, distinguishing Targeted Insult (TIN) from Untargeted (UNT).UNT covers non-targeted profanity and swearing.
- Sub-task C: Sub-task C identifies the target of TIN posts as Individual (IND), Group (GRP), or Other (OTH).Groups share characteristics such as ethnicity, gender, political affiliation, or religious belief; Other includes organizations, situations, events, and issues.
- Evaluation: Macro-averaged F1 is the official metric for all three sub-tasks because class counts are strongly imbalanced.This choice applies across the three classification layers.
- Participation and reporting: Nearly 800 teams were attracted to the task, and 115 submitted results; participants received tables, confusion matrices, and team-rank listings.The results package included materials for each submission.
4 Data
OLID is a purpose-built collection of English tweets annotated with a hierarchical three-layer model. It contains 14,100 annotated tweets, with quality control based on experienced annotators, test questions, multiple annotations, and majority voting.
- Annotation structure: OLID applies a hierarchical three-layer annotation model, with each layer used for one OffensEval sub-task.The dataset was built specifically for OffensEval.
- Dataset composition: OLID contains 14,100 annotated English tweets split into 13,240 training tweets and 860 testing tweets.A 320-tweet trial dataset was also released before the competition.
- Annotation quality: All tweets were annotated by two people, with a third annotation requested when they disagreed and majority voting used for the final label.Annotation was conducted through the Figure Eight crowdsourcing platform.
5 Results
OffensEval submissions covered traditional, deep-learning, and ensemble approaches across three subtasks, with deep learning especially popular. Performance patterns varied by subtask: BERT led subtask A, while ensembles and task-specific approaches were prominent in subtasks B and C.
- Overall results: Deep learning and ensemble models were the most popular approaches across the submissions, alongside traditional machine-learning systems.Submissions ranged from SVM and logistic regression to CNN, RNN, BiLSTM, ELMo, and BERT; similar trends appeared in subtasks B and C.
- Subtask A: 82.9% F1 was achieved by NULI in subtask A, 1.4 points above the next system.Seven of the top-10 teams used BERT, while the top non-BERT system ranked sixth with an ensemble of CNN and BLSTM+BGRU.
- Subtask B: 75.5% F1 was achieved by jhan014 in subtask B using a rule-based keyword filter, while five top-10 teams used ensembles.The subtask attracted 76 teams, and the best subtask-A team ranked fourth with 71.6%.
- Subtask C: 66 teams participated in subtask C, where ensembles were successful and the best team used BERT with preprocessing and pre-trained GloVe embeddings.Five of the top-10 teams used ensembles; the second-best system combined deep-learning and non-neural models.
- Top teams: NLPR@SRPOL, NULI, and vradivchev anikolov had the three best average ranks across all three subtasks.Their systems combined pretrained or neural models with ensembles, external or custom data, preprocessing, and class-imbalance techniques.
6 Conclusion
OffensEval used OLID to evaluate offensive-language identification across three hierarchical dimensions and attracted substantial participation. The dataset was released publicly, has been adopted in university teaching, and remains bounded by imbalance and limited test data, especially for subtasks B and C.
- OLID annotated English tweets through three hierarchical levels: offensiveness, offensive-message type, and target.
- 115 of approximately 800 registered teams participated, with leading systems using ensembles and deep-learning models such as BERT.
- OLID was made publicly available to support research beyond the SemEval competition.
- The dataset and competition setup were used in university curricula in the UK and USA.Examples included student competitions at Imperial College London and the University of Leeds, plus coursework at the University of Arizona.
- Future expansion targets dataset size, class imbalance, and the small test partition, particularly for subtasks B and C.The authors also plan to extend the dataset and task to other languages.