LLM-Based Normative Recommendation Systems: Making AI Recommendations More Attuned to Ethics and Values

This study explores the use of large language models (LLMs) to extract and formalize stakeholder norms, aiming to build more transparent and value-aligned recommendation systems with a focus on the food recommendation scenario of the DJ4ME project. The research addresses the problem that traditional recommendation systems ignore deep normative factors such as user values and ethical constraints, using LLM capabilities to bridge the gap between natural language norms and machine-executable rules.

Recommendation systems have permeated all areas of life, but traditional algorithms are mostly based on behavioral data and item features, often ignoring normative factors such as user values, ethical constraints (e.g., vegetarianism), and health goals. Normative recommendation systems attempt to address this dilemma. As a case study for food recommendation, the DJ4ME scenario has particularly critical normative factors: including health constraints (medical conditions, nutritional goals), ethical preferences (vegetarianism, halal), cultural factors (eating habits), and personal goals (weight loss, etc.), which are difficult to capture with traditional methods.

The project adopts a systematic methodology: 1. Literature review: covering normative multi-agent systems (NorMAS) theory, LLM value alignment applications, progress in explainable recommendation systems, and ethical AI fairness research; 2. Norm mining pipeline: data preprocessing → prompt engineering → norm identification → conflict detection → priority ranking; 3. LLM implementation: leveraging context understanding, reasoning capabilities, multilingual support, and few-shot learning; 4. Empirical evaluation: validation using the DJ4ME dataset, including accuracy (compared to manual annotations), coverage, and user study feedback.

The technical path includes: 1. Norm representation language: exploring formal methods such as logical rules (first-order/modal logic), weight systems, conditional norms, and priority hierarchies; 2. Integration with the Jiminy framework: integrating formalized norms into the Autonomous Jiminy (autonomous agent norm reasoning engine) and Dialogue Jiminy (dialogue-based norm negotiation) frameworks to demonstrate practical application value.

Expected contributions include: 1. Methodology: validated LLM-driven norm extraction methods, norm formalization guidelines, and an evaluation framework for normative recommendation systems; 2. System: improved transparency and interpretability, better value alignment capabilities, and support for complex ethical scenarios; 3. Practice: providing ethical guarantees for sensitive areas such as food recommendation, helping developers build responsible AI, and offering users personalized recommendations aligned with their values.

Current challenges: norm conflict resolution, dynamic norm adaptation, cultural sensitivity, and scalability. Future directions: introducing reinforcement learning to adjust norm priorities, exploring norm extraction from multimodal inputs (images/voice), building cross-domain norm knowledge bases, and researching norm-guided explanation generation techniques.

This study represents an important direction for recommendation systems to shift from technical optimization to value-sensitive design. In today's world where AI increasingly influences decisions, attention to ethics and values is a necessity. Normative recommendation systems provide a framework for thinking about how technology can serve human values, rather than just optimizing metrics.