Amazon Comprehend

Teams can train custom text classification models and custom entity recognition models by uploading labeled training data in CSV or augmented manifest format. Comprehend handles all ML pipeline steps automatically — feature engineering, training, hyperparameter tuning, and evaluation — producing precision, recall, and F1 metrics for each trained model. Custom classifiers support multi-class and multi-label modes, while custom entity recognizers can identify domain-specific entities not covered by the pre-trained models. Trained models can be deployed to real-time inference endpoints or used for asynchronous batch processing, and up to 5 training jobs per month are included in the free tier.

Identifies over 30 types of personally identifiable information — including names, addresses, Social Security numbers, credit card numbers, phone numbers, email addresses, dates of birth, bank account numbers, and driver's license numbers. Supports both detection mode (returns entity types with confidence scores and character offsets) and redaction mode (returns text with PII replaced by entity type labels or redaction markers). This enables GDPR, CCPA, and HIPAA compliance workflows without building custom regex patterns or integrating third-party data masking tools. PII detection is available via both synchronous and asynchronous APIs.

A specialized variant that extracts medical entities such as conditions, medications, dosages, procedures, test results, and anatomical terms from clinical text. Links extracted entities to standard medical ontologies including ICD-10-CM (diagnoses), RxNorm (medications), and SNOMED CT (medical concepts), enabling structured data extraction from unstructured clinical notes, discharge summaries, and pathology reports. The service is HIPAA-eligible when used under an AWS Business Associate Agreement, making it one of the few managed NLP services certified for processing protected health information in production healthcare environments.

Goes beyond document-level sentiment to identify sentiment expressed toward specific entities mentioned in the text. For example, in a product review mentioning both battery life and screen quality, targeted sentiment can separately classify the sentiment toward each attribute — positive for screen quality and negative for battery life — rather than returning a single mixed sentiment score for the entire review. This entity-level granularity is valuable for product feedback analysis, brand monitoring, and competitive intelligence where understanding sentiment toward specific features or aspects is more actionable than overall document sentiment.

Processes large collections of documents stored in Amazon S3 via asynchronous batch jobs, supporting up to 5 GB of input data per job with individual documents up to 100 KB. Results are written back to S3 in JSON format. Supports all standard NLP APIs in batch mode, enabling cost-effective processing of millions of documents without managing infrastructure. Batch jobs scale automatically and are ideal for periodic analysis of large document repositories, ETL pipelines, and data lake enrichment workflows where real-time latency is not required.