Detecting Derogatory Compounds - An Unsupervised Approach

Michael Wiegand, Maximilian Wolf, Josef RuppenhoferNAACL 2019 · DOI

Introduction

Abusive language detection systems frequently rely on blacklists of derogatory unigrams (e.g., "idiot"). However, a substantial portion of online abuse—approximately 39% in the German GermEval corpus—utilizes derogatory compounds. These multi-word insults, such as curry muncher or booze hound, present significant challenges for automated detection due to their sparsity and compositional nature. Many are "creative" one-off constructions not found in standard dictionaries, and their individual constituents are often harmless in isolation.

Methodology

This research proposes an unsupervised pipeline that leverages the distributional properties of words within a massive 30-billion-token German web corpus (COW16). The approach consists of two primary phases: high-precision linguistic diagnostics and graph-based refinement.

Linguistic Diagnostics

Four diagnostics were developed to rank compounds based on linguistic intuition:

  1. Negative Polarity (NEG): Semantic proximity to known negative sentiment terms.
  2. Compound vs. Constituent Occurrence (COMCON): Identifying compounds that appear together significantly more often than their parts would by chance.
  3. Personhood (PERSON): Similarity to embeddings of human-centric nouns (e.g., professions).
  4. Outlier Detection (OUT): Identifying semantic outliers within groups sharing the same "head" noun.

Graph-Based Refinement

The top-ranked compounds from these diagnostics serve as "seeds" for graph-based algorithms. Personalized PageRank (PRANK) is used to find semantically similar nodes in a similarity graph, while Label Propagation (LP) spreads derogatory labels across the unlabeled dataset.

Results and Evaluation

The effectiveness of the unsupervised approach was evaluated against a newly created gold standard of 3,500 German noun-noun compounds. A key finding is that derogatory compounds are a distinct linguistic phenomenon; a classifier trained solely on unigrams performs poorly on compounds (F1 63.3).

As shown in Figure 4, the unsupervised LP method achieved an F1 score of 73.5, nearly matching the performance of the best supervised SVM model (F1 74.7). This indicates that the semantic "insult" signal is strongly encoded in distributional contexts.

Comparison of supervised and unsupervised classification performance across different training sets and features.

The precision of these methods at the top of the ranked list is particularly impressive. As detailed in Figure 6, both PRANK and LP achieve 100% precision for the top 25 and top 50 results. This high precision makes these methods ideal for the automated expansion of insult lexicons without manual labeling.

Precision at top ranks (P@n) comparing various linguistic diagnostics and graph-based refinement methods.

Conclusion and Limitations

The study demonstrates that unsupervised methods can effectively detect derogatory compounds by exploiting their distributional properties. However, limitations remain, including the current focus on noun-noun compounds and the reliance on massive corpora for high-quality embeddings. Future work may extend these techniques to other compound types and languages.

Summary updated

BibTeX
@inproceedings{wiegand2019detecting,
  author = {Michael Wiegand and Maximilian Wolf and Josef Ruppenhofer},
  title = {{Detecting Derogatory Compounds - An Unsupervised Approach}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
  doi = {10.18653/v1/n19-1211},
}
Figure