🟦 Chain of Knowledge (CoK)

Chain of Knowledge (CoK) addresses limitations in existing prompting techniques, particularly the hallucination issues common in Chain-of-Thought (CoT) prompting. CoK implements a structured approach to knowledge representation and verification through two main components: evidence triples and explanation hints.

Core Components

Evidence Triples (CoK-ET)

Evidence triples are structured knowledge representations in the format (subject, relation, object). These triples provide verifiable atomic facts that serve as the foundation for reasoning.

Example triples:

• (water, boilingPoint, 100°C)
• (mammals, class, vertebrates)
• (photosynthesis, requires, sunlight)

Explanation Hints (CoK-EH)

Explanation hints provide logical connections between evidence triples to construct valid reasoning chains. They explicitly state how the presented evidence supports the final conclusion.

The CoK Prompting Process

Construct In-Context Exemplars

Choose one or several labeled examples (exemplars) that include:

• The original query.
• Annotated evidence triples (either manually created or retrieved from a knowledge base).
• Explanation hints that tie the triples to the answer.

Here's an example you can use in one-shot CoK prompting:

Generating a Response

The final prompt is constructed by concatenating the in-context exemplars with the test query. This "chain-of-knowledge prompt" guides the model to produce:

• A set of evidence triples.
• Corresponding explanation hints.
• The final answer.

Limitations of CoK

1. Requires accurate knowledge base

2. Additional processing for triple generation

3. Implementation complexity like structured format maintenance

Applications

CoK is particularly effective in:

• Technical documentation
• Scientific reasoning
• Fact-based analysis
• Educational content verification

Conclusion

Chain of Knowledge provides a structured methodology for enhancing language model reasoning through explicit knowledge representation and verification. While it requires more computational resources than simpler prompting methods, it offers improved reliability and verifiability in knowledge-intensive applications.

Footnotes

1. Wang, J., Sun, Q., Li, X., & Gao, M. (2024). Boosting Language Models Reasoning with Chain-of-Knowledge Prompting. https://arxiv.org/abs/2306.06427 ↩