CERA
About
CERA (Context-Engineered Reviews Architecture) is a training-free framework for generating realistic, controllable synthetic review datasets for Aspect-Based Sentiment Analysis (ABSA). Developed as part of an MSc thesis at the University of Windsor, it addresses data scarcity, class imbalance, and domain sparsity in ABSA research using only prompt engineering and multi-agent verification — no model fine-tuning required. Features a full-stack web GUI (React + Vite + Convex) and a Python CLI/API backend with Docker support and optional GPU acceleration. Preliminary results show synthetic data achieves up to 93.2% of real human-annotated dataset performance.
Features
Subject Intelligence Layer (SIL)
Agentic web search for factual grounding that reduces hallucination in generated reviews.
Multi-Agent Verification (MAV)
Cross-model consensus using 2/3 majority voting across multiple LLMs for fact verification.
Authenticity Modeling Layer (AML)
Configurable polarity control with noise injection to address the 'polite phenomenon' in LLM-generated text.
Multi-Dimensional Quality Assessment (MDQA)
Comprehensive evaluation suite covering lexical quality (BLEU, ROUGE-L), semantic similarity (BERTScore, MoverScore), corpus diversity, and downstream task performance.