Source-linked AI summary
SOLID: A Large-Scale Semi-Supervised Dataset for Offensive Language Identification
Sara Rosenthal, Pepa Atanasova, Georgi Karadzhov, Marcos Zampieri, Preslav Nakov
TL;DR
Offensive-language identification lacks large datasets, while OLID’s lower taxonomy levels are small and its keyword-based collection may introduce bias. The paper creates SOLID through semi-supervised labeling and evaluates it with OLID. SOLID improves performance on lower taxonomy levels, but the authors acknowledge persistent dataset bias and potential misuse.
Problem
OLID’s lower taxonomy levels are small and imbalanced, and keyword-based collection may bias offensive-language data.
Method
The authors create SOLID by using democratic co-training with OLID as a seed dataset and an ensemble of diverse models to label tweets semi-supervised.
Results
Training on SOLID and testing on OLID yields sizeable improvements at the middle and lower taxonomy levels, including 12.1 versus 4.2 macro-F1 points for FastText and BERT at Level B.
Takeaways & Limitations
SOLID provides a nine-million-tweet resource and shows that adding semi-supervised data can improve offensive-language identification where gold training data is scarce.
Takeaways & Limitations
The dataset may retain biases from its supervised source models, offensive-to-non-offensive ratio, and collection method, and may be misused for unfair moderation.
Abstract
from arXiv · showhide
The widespread use of offensive content in social media has led to an abundance of research in detecting language such as hate speech, cyberbullying, and cyber-aggression. Recent work presented the OLID dataset, which follows a taxonomy for offensive language identification that provides meaningful information for understanding the type and the target of offensive messages. However, it is limited in size and it might be biased towards offensive language as it was collected using keywords. In this work, we present SOLID, an expanded dataset, where the tweets were collected in a more principled manner. SOLID contains over nine million English tweets labeled in a semi-supervised fashion. We demonstrate that using SOLID along with OLID yields sizable performance gains on the OLID test set for two different models, especially for the lower levels of the taxonomy.
1 Introduction
Offensive language creates moderation challenges at social-media scale, while existing datasets are limited by scarcity and collection bias. The paper introduces SOLID, a larger semi-supervised resource intended to improve offensive-language identification, especially for harder taxonomy levels.
- Motivation: Automatic moderation systems are needed because human moderators cannot handle the volume of offensive social-media content.Moderation failures can affect platforms, governments, and individuals.
- Prior dataset: OLID uses a hierarchical taxonomy that represents both the type and target of offensive content.The taxonomy distinguishes, for example, group-targeted hate speech from individual-targeted cyberbullying.
- Dataset gap: Lower taxonomy levels contain fewer instances, while randomly sampled tweets contain less than 3% offensive content, making robust training and data collection difficult.These constraints make offensive examples costly and time-consuming to obtain.
- Contributions: SOLID applies a semi-supervised collection method using OLID as a seed dataset, avoiding time-consuming annotation.The method is presented as the first application of semi-supervised collection for new offensive data in this work.
- Contributions: SOLID contains 9 million English tweets and is released as the largest dataset for offensive language identification.It also served as the official dataset for the SemEval-2020 OffensEval shared task.
- Results: Training on SOLID and testing on OLID produces sizeable improvements at the middle and lower taxonomy levels, where gold training data is scarce.The paper also analyzes EASY explicit and HARD implicit offensive tweets.
2 Related Work
Prior work spans offensive language, hate speech, cyberbullying, aggression, and toxic-language detection across datasets, models, and shared tasks. This paper extends OLID’s hierarchical approach to offensive-language identification.
- Research landscape: Research has addressed offensive language alongside hate speech, cyberbullying, aggression, and toxic-comment detection.These tasks often overlap in the characteristics they detect.
- Datasets and tasks: Existing work uses diverse resources, including Twitter, Facebook, Wikipedia, and multilingual shared-task datasets.Examples include HatEval, TRAC, GermEval, HASOC, and the Kaggle Toxic Comment Classification Challenge.
- OLID extension: OLID differs from prior approaches by modeling both the target and type of offensive content within a hierarchical annotation schema.The schema supports more specific analysis than a single offensive/non-offensive label.
3 The OLID Dataset
OLID organizes offensive-language identification into a three-level hierarchy covering offensiveness, targeting, and target identity. Its lower levels are substantially smaller and imbalanced, motivating a larger dataset.
- Level A: Offensive Language Detection: Level A determines whether a text is offensive or non-offensive.OFF includes inappropriate language, insults, or threats; NOT is neither offensive nor profane.
- Level B: Categorization of Offensive Language: Level B determines whether offensive text is targeted or untargeted.TIN denotes a targeted insult or threat, while UNT denotes untargeted profanity or swearing.
- Level C: Offensive Language Target Identification: Level C identifies whether the target is an individual, group, or other entity.Group targets include characteristics such as ethnicity, gender, sexual orientation, or religion.
- Dataset scope: The taxonomy has been adopted for multiple languages and shared tasks, and OLID contains 13,241 training and 860 test tweets.Its structure provides a basis for multilingual learning and analysis.
- Dataset limitations: Level C is 30% smaller than Level A, with substantial class imbalance and especially small datasets at Levels B and C.These distributional drawbacks indicate the need for a larger dataset.
4 Models
The paper evaluates heterogeneous text models and combines them through democratic co-training to generate semi-supervised labels. The models range from frequency-based and subword methods to recurrent and Transformer architectures.
- Model suite: The semi-supervised setup uses PMI, FastText, LSTM, and BERT models with diverse inductive biases.The diversity is treated as essential for the ensemble used in democratic co-training.
- PMI: The PMI model classifies tweets by averaging n-gram PMI and PMI-SO scores for each class and selecting the highest-scoring class.It falls back to default labels when an instance contains no scored words.
- PMI: PMI measures n-gram association with a class using frequencies from class-specific and overall training data.PMI-SO additionally contrasts the target class with all other classes.
- FastText: FastText uses subword representations to handle noisy tweets and supplies a modeling representation distinct from PMI, LSTM, and BERT.The experiments use bigrams for Levels A and B and trigrams for Level C.
- Neural models: LSTM models long-distance word relations with bidirectional recurrence and attention, while BERT is fine-tuned through transfer learning for all three taxonomy levels.BERT is used for classification in a low-resource setting.
- Democratic co-training: Democratic co-training trains diverse supervised models on OLID, predicts confidence scores for unlabeled tweets, and uses the ensemble to label SOLID.The approach is intended to reduce noise associated with individual model biases.
5 The SOLID Dataset
SOLID is constructed from over 12 million collected tweets, with nine million labeled semi-supervised using democratic co-training. The section also describes model aggregation, confidence-based labeling, and manually annotated test data spanning different difficulty levels.
- Dataset collection: Over 12 million tweets were collected, and nine million were labeled using democratic co-training for SOLID.The resulting dataset supports offensive-language identification across the taxonomy’s levels.
- Dataset collection: Tweets were collected through Twitter’s streaming API using common English stopwords to obtain diverse random English tweets.The collection process also filtered for English, minimum length, and self-contained text without URLs.
- Semi-supervised labeling: The ensemble combined BERT, LSTM, PMI, and FastText, with BERT strongest on Levels A and C and PMI nearly matching LSTM on Level A.Level B was more challenging for frequency- and n-gram-based models, while BERT and LSTM outperformed FastText and PMI at Level C.
- Semi-supervised labeling: Average and standard-deviation confidence scores were aggregated across models to reduce model-specific bias and filter potentially noisy annotations.The procedure used confidence thresholds to assign labels progressively from Level A to Levels B and C.
- Test dataset: A held-out set of three million tweets was used to create a new test set because the OLID test set was especially small at Levels B and C.The manually annotated test data included 5,993 tweets and was divided into confidence-based Easy and Hard subsets.
- Test dataset: Hard offensive tweets were more difficult to label accurately: 25% were judged NOT, compared with 8% of Easy offensive tweets.The lower agreement at Level C also reflects its three-way annotation scheme.
6 Experiments and Evaluation
The experiments compare OLID-only training with OLID+SOLID training using BERT and FastText on OLID and SOLID test sets. SOLID generally helps most where training data is scarce or models are weaker, while gains vary by taxonomy level, model, and tweet difficulty.
- Experimental setup: Training on OLID+SOLID is evaluated against OLID-only training on the OLID test set for BERT and FastText.The experiments also evaluate SOLID test performance and Easy versus Hard subsets.
- OLID results: For Level A, adding SOLID improves FastText but changes BERT little and can slightly degrade its performance.The authors relate this pattern to co-training helping more when the initial classifier is weaker.
- OLID results: 12.1 vs. 4.2 macro-F1 points absolute are the Level B gains for FastText and BERT, respectively, after adding SOLID.Level B has less OLID training data and is more complex; FastText performs at the majority-class baseline with OLID only.
- OLID results: 5.4 macro-F1 points absolute is BERT’s Level C gain from adding SOLID, while FastText does not improve.Level C has fewer manually annotated OLID examples and three classes rather than two.
- SOLID results: 0.923 and 0.860 macro-F1 are the BERT and FastText Level A scores on the SOLID test, with only FastText improving slightly after OLID augmentation.Hard tweets are substantially more difficult, with BERT and FastText scoring 0.570 and 0.536, respectively; SOLID improves both models on Hard tweets.
- Easy versus Hard tweets: A 0.936 F1 curse-word baseline on Easy tweets contrasts with 0.580 macro-F1 on Hard tweets, indicating that models rely partly on curse words.Hard tweets often express offensiveness through negative biases rather than curse words, and remain an open classification challenge.
- SOLID results: Combining SOLID and OLID yields sizable Level B and C improvements, while variation across models highlights instability in the small OLID test set.The larger SOLID test set is intended to provide more stable evaluation, especially for Levels B and C.
7 Conclusion and Future Work
SOLID is a large-scale semi-supervised dataset created with an ensemble of four models, containing nine million English tweets. Using SOLID improves performance on Levels B and C of OLID, while distinguishing Easy from Hard offensive tweets and identifying Hard detection as an open challenge.
- SOLID contains nine million English tweets and was created using an ensemble of four different models.
- Using SOLID yields noticeable performance improvements for Levels B and C of the OLID annotation schema on the OLID test set.
- The approach distinguishes Hard offensive tweets from Easy tweets, unlike keyword-based collection.
- Detecting Hard offensive tweets remains an open challenge, motivating future work on categorizing Hard tweets.
Ethics Statement
The dataset was collected under Twitter’s terms and incorporates procedures intended to reduce bias, but the authors acknowledge that semi-supervised labeling and demographic variation leave important risks. They also caution that SOLID may be misused for unfair moderation and that human supervision remains necessary.
- SOLID tweets were collected through the Twitter API using frequent stop words, while only public tweets and user IDs were provided.
- Offensive-language judgments are subjective, making biases in the gold-labeled data inevitable despite a defined annotation schema and high inter-annotator agreement.
- Semi-supervised labeling can expand biases from the gold data, so the authors used random word sampling, diverse models, and manual labeling of offensive test tweets to reduce them.
- The authors acknowledge that current semi-supervised techniques do not eliminate biases inherited from supervised source models, and collection-related biases remain hard to avoid.
- Demographic variation, including the gender of a target, can affect how offensive language is expressed and remains a direction for future work.
- SOLID could be misused to unfairly moderate non-offensive text, so human moderators are required to prevent this harm.
A Appendix
The appendix provides additional details about data collection, analysis, and implementation.
- The appendix supplies further details on data collection, analysis, and implementation.
A.1 Data Collection and Analysis
SOLID tweet collection uses frequent English words sampled from the Project Gutenberg corpus rather than offensive keywords. The appendix also documents the curse-word baseline and its 22 offensive terms.
- Data collection: SOLID collection queries the Twitter API with frequent English words from the Project Gutenberg corpus.
- Data collection: The collection method randomly selects a word by sampling a number between 0 and 1 and choosing the corresponding frequency-ranked word.
- Data collection: Table 9 lists the top-20 frequent English words and their normalized frequencies used for collection.
- Data analysis: The Easy OFF/NOT analysis uses a simple baseline based on 22 common curse words.
A.2 Implementation Details
The implementation uses models fine-tuned on OLID and evaluates them with macro-F1. Their agreement varies across taxonomy levels, with greatest disagreement at Level C, motivating ensemble combination.
- Models were fine-tuned on 10% of the OLID dataset.
- Macro-F1 score, as implemented in scikit-learn, was used for all experiments.
- Table 11 reports validation-set macro-F1 scores for the ensemble models across Levels A, B, and C.
- Agreement was more common among all four models for Levels A and B than for Level C.
- Level C showed the greatest model disagreement and the least frequent unanimous agreement, consistent with lower model performance.
- The observed prediction variance, especially at lower taxonomy levels, supports resolving individual model differences through ensemble combination.