[Paper Review] Constructing interval variables via faceted Rasch measurement and multitask deep learning: a hate speech application

Chris J. Kennedy, Geoff Bacon+2 more|arXiv (Cornell University)|Sep 22, 2020

Hate Speech and Cyberbullying Detection55 references35 citations

TL;DR

The paper presents a method to construct continuous interval measurements of hate speech by combining faceted Rasch item response theory with multitask deep learning, using ordinal survey items and debiased predictions from text data.

ABSTRACT

We propose a general method for measuring complex variables on a continuous, interval spectrum by combining supervised deep learning with the Constructing Measures approach to faceted Rasch item response theory (IRT). We decompose the target construct, hate speech in our case, into multiple constituent components that are labeled as ordinal survey items. Those survey responses are transformed via IRT into a debiased, continuous outcome measure. Our method estimates the survey interpretation bias of the human labelers and eliminates that influence on the generated continuous measure. We further estimate the response quality of each labeler using faceted IRT, allowing responses from low-quality labelers to be removed. Our faceted Rasch scaling procedure integrates naturally with a multitask deep learning architecture for automated prediction on new data. The ratings on the theorized components of the target outcome are used as supervised, ordinal variables for the neural networks' internal concept learning. We test the use of an activation function (ordinal softmax) and loss function (ordinal cross-entropy) designed to exploit the structure of ordinal outcome variables. Our multitask architecture leads to a new form of model interpretation because each continuous prediction can be directly explained by the constituent components in the penultimate layer. We demonstrate this new method on a dataset of 50,000 social media comments sourced from YouTube, Twitter, and Reddit and labeled by 11,000 U.S.-based Amazon Mechanical Turk workers to measure a continuous spectrum from hate speech to counterspeech. We evaluate Universal Sentence Encoders, BERT, and RoBERTa as language representation models for the comment text, and compare our predictive accuracy to Google Jigsaw's Perspective API models, showing significant improvement over this standard benchmark.

Motivation & Objective

• The motivation to measure complex social constructs as continuous interval variables rather than binary labels.
• The goal to debias human labeling and estimate labeler quality within a scalable predictive framework.
• The aim to integrate Rasch-based measurement with deep learning for debiased, interpretable predictions.

Proposed method

• Decompose hate speech into eight theorized components and label them with ordinal survey items.
• Use faceted Rasch measurement theory to transform multi-item ordinal labels into a continuous interval scale.
• Train a multitask deep learning model with shared weights to predict the latent components from text.
• Employ ordinal softmax activation and ordinal cross-entropy loss to exploit the ordinal structure of targets.
• Apply a partial credit IRT transformation to predictions to obtain plausible value scores.
• Enable debiasing by estimating reviewer labeler bias and filter low-quality responses.

Experimental results

Research questions

• RQ1 Can hate speech be modeled as a continuous spectrum using a faceted Rasch framework combined with supervised deep learning?
• RQ2 Does debiasing labeler bias improve the accuracy and reliability of the continuous hate speech score?
• RQ3 Can a multitask architecture provide interpretable continuous predictions aligned with constituent components?
• RQ4 Do ordinal activation and loss functions improve prediction for ordinal targets compared to standard approaches?
• RQ5 How does this method perform relative to existing benchmarks such as Google Jigsaw’s Perspective API?

Key findings

• The method yields a continuous hate speech scale derived from eight theorized levels and 32–48 labeling items.
• A 50,000-comment dataset labeled by 10,000 crowdworkers across YouTube, Twitter, and Reddit was constructed.
• The approach shows significant improvement over Perspective API in predictive accuracy on the task.
• Faceted Rasch scaling provides invariant measurement with a debiasing mechanism for labelers and comments.
• The multitask model offers interpretable predictions by linking each continuous score to its constituent components in the penultimate layer.