Skip to content

scoring.md

Latest commit

 

History

History
67 lines (50 loc) · 5.47 KB

scoring.md

File metadata and controls

67 lines (50 loc) · 5.47 KB

Description of Evaluation Measure

This project focuses on leveraging state-of-the-art (SOTA) NLP techniques for hate speech detection using the MetaHate dataset, a unified benchmark for this task. Specifically, given a social media post, the goal is to classify it as hateful or non-hateful. Since this is a binary classification problem, the F1 score will be used as the primary evaluation metric, as it balances Precision and Recall—two critical aspects in hate speech detection. Recall captures the proportion of actual hate speech posts correctly identified by the model. Optimizing for recall ensures harmful content is not overlooked. Precision measures the proportion of posts labeled as hate speech that are truly hateful. Optimizing for precision reduces the risk of mislabeling benign content and unnecessarily restricting users’ expression. For social media platforms represented in the MetaHate dataset, both metrics are essential: missing hate speech allows harmful content to spread, while false positives can unfairly penalize users. Striking the right balance ensures the platform remains safe and inclusive without over-censoring legitimate expression, maintaining trust and engagement among users. In this work, the F1 score is computed specifically for the hate speech class (labeled as 1 in our dataset), as correctly identifying hateful content is our primary concern.

Mathematically, Precision, Recall, and F1 score are defined as follows:

  • Precision = TP / (TP + FP)
  • Recall = TP / (TP + FN)
  • F1 Score: (2 * Precision * Recall) / (Precision + Recall)

where TP, FP, and FN denote the number of true positives, false positives, and false negatives, respectively. The F1 score provides a single metric that balances Precision and Recall, making it particularly suitable for tasks like hate speech detection where both false negatives and false positives carry significant consequences.

We want to note that accuracy alone is insufficient in this scenario because the dataset is imbalanced, with only about 20% of posts labeled as hate speech. A naive classifier that always predicts “non-hateful” could achieve 80% accuracy without detecting any actual hate speech. By using the F1 score as the primary metric, we ensure that the model’s performance reflects its ability to correctly identify hateful content while minimizing false positives, providing a more meaningful evaluation for this imbalanced dataset.

To conclude, we note that our strong baseline draws direct inspiration from the paper MetaHate: A Dataset for Unifying Efforts on Hate Speech Detection, which fine-tunes a BERT model on the MetaHate dataset. In their experiments, they also use F1 score as an evaluation metric, emphasizing its importance for balancing Precision and Recall in hate speech detection tasks.

Hate Speech Evaluation Metrics: Literature Review

In this section, we review prior works that utilize the F1 score as a primary evaluation metric for hate speech detection:

Both studies leverage the MetaHate dataset and report the F1 score to assess the effectiveness of their proposed approaches in identifying hate speech.

Running score.py

Pass NumPy files (.npy) containing the ground truth and predicted labels. Each file contains a NumPy array of integers 0 or 1, where:

  • 0 represents a non-hateful post
  • 1 represents a hateful post

Example Command

python score.py \
    --true_labels_file ground-truth-test.npy  \
    --pred_labels_file simple-baseline-test-preds.npy

Note to Grader: ground-truth-test.npy has been submitted along with the other deliverables of this assignment!

Example Output

Accuracy: 0.7500
Pos Precision: 0.7500
Pos Recall: 0.7500
Neg Precision: 0.7500
Neg Recall: 0.7500
Pos F1: 0.7500
Neg F1: 0.7500
F1 Macro: 0.7500
F1 Micro: 0.7500
F1 Weighted: 0.7500

Output Interpretation

  • Accuracy: Overall proportion of correctly classified posts (both hateful and non-hateful).
  • Pos Precision: Of the posts predicted as hateful (1), the fraction that are actually hateful. Formula: TP / (TP + FP).
  • Pos Recall: Of all truly hateful posts, the fraction correctly identified as hateful. Formula: TP / (TP + FN).
  • Neg Precision: Of the posts predicted as non-hateful (0), the fraction that are actually non-hateful. Formula: TN / (TN + FN).
  • Neg Recall: Of all truly non-hateful posts, the fraction correctly identified as non-hateful. Formula: TN / (TN + FP).
  • Pos F1: F1 score for the hateful class, harmonic mean of Pos Precision and Pos Recall. Formula: 2 * (Precision * Recall) / (Precision + Recall).
  • Neg F1: F1 score for the non-hateful class, harmonic mean of Neg Precision and Neg Recall. Formula: 2 * (Precision * Recall) / (Precision + Recall)
  • F1 Macro: Average of Pos F1 and Neg F1, treating both classes equally regardless of class frequency.
  • F1 Micro: Global F1 considering total true positives, false positives, and false negatives across all classes.
  • F1 Weighted: Average of Pos F1 and Neg F1 weighted by the number of instances in each class.

Note to Grader: While the script returns multiple evaluation metrics, our primary metric for assessing model performance is the F1 score(i.e. Pos F1). The other metrics are provided solely for additional analysis.