How do you implement natural language generation for automated report writing?

Natural language generation (NLG) transforms structured data into readable text. Architecture: content planning (determine what information to include based on significance and audience), sentence planning (organize information into coherent sentence structures), and surface realization (generate actual text with proper grammar and style). Template-based NLG: predefined templates with slots for variable data ("Revenue increased by {percentage}% to {amount} in {period}"). Advantages: controllable, consistent, no hallucination risk. Limitations: rigid, requires manual template creation. ML-based NLG: use LLMs for flexible, natural-sounding text generation from structured data. Prompt engineering: provide data in structured format (JSON, tables) with clear instructions for report format, tone, and focus areas. Fine-tune on domain-specific report examples for consistent style. Hybrid approach: use templates for critical sections (financial figures, compliance statements) and LLM generation for analytical commentary and narrative sections. Implementation: define report structure (sections, order, required elements), implement data pipelines that aggregate relevant metrics, generate each section using appropriate method (template or LLM), and assemble into final report with formatting. Quality assurance: implement fact-checking pipeline that verifies all numbers in generated text match source data, check for logical consistency (e.g., percentages add to 100, comparisons are directionally correct), and implement human review workflows for high-stakes reports. Applications: financial reporting, sports summaries, weather forecasts, business intelligence dashboards, medical report generation, and e-commerce product descriptions. Personalization: adjust language complexity for audience (executive summary vs technical report), highlight metrics relevant to the reader's role, and adapt length and detail level.