Answer: Use Amazon Comprehend toxicity detection.

## Detailed Explanation ### Question Analysis The question asks for a strategy to identify harmful language in social media comments using a machine learning model, with the specific constraint that the company will **not use labeled data to train the model**. This eliminates approaches that require custom model training with labeled datasets. ### Evaluation of Options **A: Use Amazon Rekognition moderation** - **Not suitable**: Amazon Rekognition is primarily designed for **image and video analysis**, including content moderation for visual media. It can detect inappropriate content in images/videos but is not optimized for **text-based language analysis**. Since the task involves analyzing text comments, Rekognition is not the appropriate service. **B: Use Amazon Comprehend toxicity detection** - **Optimal choice**: Amazon Comprehend is a **natural language processing (NLP) service** specifically designed to analyze text. It includes a **Toxicity Detection** feature that can identify harmful language (e.g., hate speech, harassment, insults) in text content. This service: - Requires **no labeled data** for training, as it uses pre-trained models. - Is specifically designed for **text analysis tasks** like sentiment analysis, entity recognition, and toxicity detection. - Can be directly applied to social media comments without any model training. - Aligns perfectly with the requirement to avoid using labeled training data. **C: Use Amazon SageMaker built-in algorithms to train the model** - **Not suitable**: While SageMaker provides built-in algorithms for various ML tasks, **training a model would require labeled data**. The question explicitly states the company will not use labeled data for training, making this approach contradictory to the requirements. Even with built-in algorithms, supervised learning for text classification necessitates labeled examples. **D: Use Amazon Polly to monitor comments** - **Not suitable**: Amazon Polly is a **text-to-speech service** that converts text into lifelike speech. It has no capability for language analysis, toxicity detection, or content moderation. This service is completely unrelated to the task of identifying harmful language in text. ### Key Considerations 1. **Service Specialization**: Different AWS AI services have distinct purposes. For text-based harmful language detection, Amazon Comprehend is specifically designed for this type of NLP task. 2. **No-Label Requirement**: The constraint eliminates approaches requiring supervised learning with labeled datasets. Amazon Comprehend's toxicity detection uses pre-trained models that don't require custom training data. 3. **Task Alignment**: The goal is to analyze **text comments** for harmful language, which falls squarely within Amazon Comprehend's text analysis capabilities. ### Conclusion Amazon Comprehend toxicity detection (Option B) is the only approach that: - Directly addresses text-based harmful language detection - Requires no labeled training data - Uses a purpose-built service for NLP tasks - Can be implemented immediately without model development All other options either address the wrong modality (images/video vs. text), require prohibited labeled data, or are completely unrelated to the task.