Amazon Comprehend
Uses Natural Language Processing (NLP) to understand documents, extract information and organize them, this is a completely managed and serverless service provided by Amazon Web Services , we also have custom classification defined by us, to classify emails as complaints or anything else, has a realtime analysis and async analysis, it can also do Named Entity Recognition (NER) to identify people, companies, dates and what not, also has support for Custom Entity Recognition (CER)
AI Practitioner focus
- Choose Comprehend for managed NLP analysis of existing text: sentiment, targeted sentiment, entities, key phrases, language, syntax, classification, and PII detection.
- It analyzes text rather than extracting text from scans (Amazon Textract), translating it (Amazon Translate), or searching a corpus (Amazon Kendra).
Key points
- Extracts insights about the content of documents, recognizing key phrases, languages, and sentiment.
- Useful for customer-service engagement, surfacing sentiment and topics from support interactions at scale.
- Provides both real-time (synchronous) analysis and asynchronous batch analysis.
- Real-time entity/key-phrase/language detection is capped at 100 KB per document; real-time sentiment, targeted sentiment, and syntax detection are capped at 5 KB per document. Asynchronous jobs accept much larger inputs (e.g. up to 1 MB per document for entities and key phrases).
- Supports custom classification, letting users define their own categories (e.g. classifying emails as complaints), including a semi-structured mode that classifies PDFs, Word documents, and images directly (built on Amazon Textract extraction under the hood).
- Topic modeling, event detection, and prompt safety classification are no longer available to new customers (accounts without prior use in the last 12 months); AWS recommends using Amazon Bedrock LLMs for topic/event extraction and Bedrock Guardrails for prompt safety classification instead. All other Comprehend features are unaffected.
- Distinct from Amazon Textract, which extracts text, handwriting, and form/table data from scanned documents (OCR), rather than analyzing already-digitized text for sentiment and entities.
- Distinct from Amazon Kendra, which powers intelligent document/keyword search rather than surfacing sentiment, entities, or key phrases.