Inducing a Lexicon of Abusive Words - a Feature-Based Approach

Michael Wiegand, Josef Ruppenhofer, Anna Schmidt, Clayton GreenbergNAACL 2018 · DOI

Abstract

Detecting abusive language is a persistent challenge due to the evolving nature of slurs and the tendency of machine learning models to overfit to domain-specific topics. This paper presents a feature-based approach to automatically generate a large, high-quality lexicon of abusive words. By training a classifier on a small set of manually annotated negative terms and leveraging linguistic features such as polar intensity and sentiment perspective, we expanded a base lexicon of 551 words to nearly 3,000. Our results demonstrate that this lexicon-based approach significantly improves cross-domain robustness compared to standard supervised models.

Introduction: The Problem of Domain Mismatch

The detection of hate speech and cyberbullying is often hampered by domain mismatch. State-of-the-art classifiers, such as RNNs or FastText, frequently learn to associate specific topics (e.g., "sports") with abuse if those topics are prevalent in their training data. Consequently, a model trained on sports-related tweets often fails when applied to Wikipedia comments. To address this, we propose that a robust, domain-independent lexicon of abusive words is essential for bridging the gap between disparate social media environments.

Methodology

We frame the task of lexicon induction as a binary classification problem: identifying which "negative polar expressions" are specifically "abusive."

  1. Base Lexicon & Calibration: We began with a base lexicon of 1,650 negative words, manually labeled via crowdsourcing.
  2. Feature Engineering: We utilized high-level linguistic features rather than raw text:
  3. Polar Intensity: Distinguishing "strong" negative words from "weak" ones.
  4. Sentiment Views: Separating "speaker views" (e.g., liar) from "actor views" (e.g., disappointed).
  5. Surface Patterns: Mining Twitter for syntactic patterns like "called {me|him|her} a <noun>" to identify candidate nouns.
  6. Lexical Resources: Incorporating WordNet and Wiktionary tags (e.g., "pejorative") and dictionary definitions.
  7. Lexicon Expansion: An SVM classifier was trained on these features to expand the list of abusive words from 551 to 2,989.
Figure 1: Lexicon induction process using seed words identified from surface patterns and negative polar expressions.

Datasets

Our research utilized diverse resources for training and evaluation:

  • Lexical Resources: Subjectivity Lexicon, WordNet, and Wiktionary.
  • Unlabeled Corpora: Amazon Review Corpus (1.2B tokens), Web As Corpus (2.3B tokens), and RateItAll (4.7M tokens).
  • Evaluation Datasets: Four labeled micropost datasets covering diverse sources, Twitter (sexism/racism), UseNet, and Wikipedia Talk pages.

Key Insights

A central finding is that abuse is a specific subset of negativity. Not all negative words are abusive; the classifier's primary role is to filter this subset.

A critical observation in our research was the importance of the "Speaker View." Abusive terms are almost exclusively expressions of the speaker's opinion. If a word describes the internal state of the subject (actor-view), it is rarely abusive. Furthermore, abusive words tend to have higher polar intensity. As shown in the data, words categorized under the "speaker" view or "strong" intensity are significantly more likely to be classified as abusive.

Figure 2: Distribution of abusive vs. non-abusive words across polar intensity and sentiment views.

Key Results

  • Word-Level Accuracy: Our feature-based SVM achieved an F1-score of 77.5%, outperforming generic word embeddings (75.7%) and baselines trained on labeled microposts (65.8%).
  • Cross-Domain Robustness: In cross-domain evaluations, the lexicon-based approach consistently outperformed state-of-the-art models. For instance, on the Razavi dataset, our model achieved an F1 of 71.70, whereas a strong in-domain SVM only reached 60.05.
  • Explicit vs. Implicit Abuse: The lexicon is highly effective at identifying explicit abuse (slurs and insults), with F1-scores reaching 80.7%. However, it remains less effective against implicit abuse, such as sarcasm or metaphor.

Limitations and Future Work

While the lexicon-based approach provides a stable foundation across domains, it is inherently limited by its inability to detect abuse conveyed through implicit means. Additionally, while adjectives help identify some obfuscated nouns, intentional misspellings remain a challenge. Future work will explore adapting these features to other languages and combining the lexicon with more sophisticated models to better address implicit abuse.

Summary updated

BibTeX
@inproceedings{wiegand2018inducing,
  author = {Michael Wiegand and Josef Ruppenhofer and Anna Schmidt and Clayton Greenberg},
  title = {{Inducing a Lexicon of Abusive Words - a Feature-Based Approach}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
  doi = {10.18653/v1/n18-1095},
}
Figure