The Linguistic Ideologies of Deep Abusive Language Classification

Michael CastelleEMNLP 2018

TL;DR

This paper critically examines how modern NLP models for abusive language detection implicitly adopt specific "language ideologies" that prioritize surface-level patterns over deeper linguistic and social context. By analyzing common datasets and neural network approaches, the authors argue for incorporating pragmatic and metapragmatic theories to build more robust and nuanced classification systems.

What Problem Does This Paper Address?

The paper tackles the challenge of accurately classifying abusive language, particularly hate speech, using automated methods. While the field of Natural Language Processing (NLP) has made strides, the authors argue that current approaches often rely on a simplistic view of language as mere pattern recognition, akin to computer vision. This overlooks the crucial role of social context, pragmatics (how meaning is conveyed through use), and metapragmatics (reflexivity about language use) in understanding linguistic abuse. The paper highlights a tension between the NLP research culture of achieving "state-of-the-art" scores on decontextualized datasets and the reality that abusive language is deeply embedded in social interactions and power dynamics.

Methods

The authors employ standard deep learning architectures for text classification: Convolutional Neural Networks (CNNs) and Bidirectional Gated Recurrent Units (RNN Bi-GRUs). These models are trained on word embeddings derived from Wikipedia data. The core of their methodological contribution lies not in novel algorithms, but in the analysis of how these standard methods perform on different datasets and the interpretation of their results through the lens of linguistic theory. They conduct an error analysis, particularly on the StackOverflow dataset, to understand why certain classifications fail.

Datasets

The paper uses three publicly available datasets for abusive language classification:

  • Kaggle Toxic Comment Classification Challenge: Decontextualized Wikipedia talk page comments labeled for toxicity (toxic, severe toxic, obscene, threat, insult, identity hate).
  • TRAC1 (Trolling, Aggression and Cyberbullying): Stand-alone Facebook posts and comments in English and Hindi, labeled as "Overtly Aggressive," "Covertly Aggressive," or "Not Aggressive."
  • StackOverflow dataset: User-flagged comments from StackOverflow, specifically those flagged as "Rude or offensive." This dataset is unique as flags are provided by users who have seen the interaction context, making it a form of metapragmatic data.

The authors note that for comparative analysis, labels across datasets were sometimes grouped (e.g., "aggressive" categories combined).

Key Insights and Ideas

  • "Language Ideologies" in NLP: The paper argues that NLP, particularly in abusive language detection, operates under specific "language ideologies" – sets of beliefs about language. These often lean towards a Chomskyan view of an innate language faculty and a behaviorist stimulus-response model, prioritizing formal structure over social context.
  • Decontextualization is a Problem: Datasets are often "decontextualized," meaning utterances are stripped of their surrounding conversation. This aligns with a view of language as static text, ignoring how meaning and offensiveness are dynamically constructed in interaction.
  • Pragmatics and Metapragmatics Matter: The authors emphasize that understanding abusive language requires theories of pragmatics (how context shapes meaning) and metapragmatics (how speakers reflect on and manage language use). These are crucial for interpreting indirect or implicit abuse.
  • Model Performance Varies by Context: Standard neural network models perform well when offensiveness is signaled directly within an utterance (like explicit slurs). However, they struggle significantly when offensiveness relies on context or metapragmatic cues, as seen in the StackOverflow dataset.
  • User Flags vs. Annotator Labels: The StackOverflow dataset, with user-generated flags based on interaction context, reveals limitations in models that rely solely on decontextualized labels. Comments flagged as offensive by users might not be so when viewed in isolation, and vice-versa.

Key Results

  • Similar Performance for CNN and Bi-GRU: The two neural network architectures (CNN and Bi-GRU) performed similarly across the datasets, suggesting that the choice of architecture might be less critical than the nature of the data and the underlying linguistic assumptions.
  • Kaggle Dataset Yields Best Results: The Kaggle dataset, despite class imbalance, showed the best overall performance (micro-averaged F1 score), with high precision and recall for the "normal" class and decent performance for the "offensive" class.
  • StackOverflow Shows Poor Recall: The StackOverflow dataset exhibited the lowest micro-averaged F1 score, particularly for the "offensive" class, with very low recall (0.2). This indicates the models frequently missed offensive comments in this dataset.
  • Error Analysis Highlights Context Dependency: Qualitative analysis of StackOverflow errors revealed that many false negatives (comments flagged as offensive by users but missed by models) were context-dependent. Conversely, many comments predicted as offensive by models were not flagged by users, suggesting the models were picking up on subtle cues that users, with context, did not find offensive.

Limitations and Future Work

Limitations: * Dataset Limitations: The paper highlights that current datasets often lack sufficient contextual information, hindering the development of models that can handle nuanced abusive language. * Model Limitations: While neural networks can model context, their effectiveness is limited by the available data. The current models struggle with context-dependent offensiveness. * Class Imbalance: While addressed to some extent, class imbalance remains a challenge in abusive language datasets.

Future Work: * Incorporate Pragmatic and Metapragmatic Theories: Future NLP research should actively integrate theories of pragmatics and metapragmatics into model design and evaluation. * Model Interaction Dynamics: For platforms like StackOverflow, future work should explore modeling the affective and interactional dynamics of communication, potentially using embeddings of interaction networks. * Consider Communities of Practice: A deeper understanding of abusive language requires attending to "communities of practice," where norms of behavior and language use are embodied and context-specific. * Beyond Stand-Alone Utterances: Future research should move beyond classifying isolated utterances and consider the broader communicative context, including user interactions and platform-specific norms.

Summary updated

BibTeX
@inproceedings{castelle2018linguistic,
  author = {Michael Castelle},
  title = {{The Linguistic Ideologies of Deep Abusive Language Classification}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2018},
}
Figure