Exploring Large Language Models for Knowledge Graph Completion

In the realm of artificial intelligence, knowledge graphs have emerged as a crucial component, though they often grapple with the issue of incompleteness. A recent study by Liang Yao, Jiazhen Peng, Chengsheng Mao, and Yuan Luo attempts to address this problem by leveraging Large Language Models (LLM) for knowledge graph completion.

The researchers have developed a unique framework known as Knowledge Graph LLM (KG-LLM), which treats triples in knowledge graphs as text sequences. This innovative approach utilizes entity and relation descriptions of a triple as prompts, and the response from these prompts is then used for predictions.

The researchers have found that Large Language Models can address the issue of incompleteness by modeling knowledge graph completion as a sequence-to-sequence problem. The entity, relation, and triples are treated as textual sequences, and LLMs can then perform instruction tuning on these sequences to predict the plausibility of a triple or a candidate entity/relation, enhancing the accuracy of knowledge graph completion tasks.

The KG-LLM framework models triples as text sequences by transforming entity and relation descriptions into simple questions for a large language model to complete. For example, given the triple <Steve Jobs, founded, Apple Inc.>, the prompt formation would be "Is this true: Steve Jobs founded Apple Inc.?". The ideal output of LLM would then be "Yes, this is true."

When tested on various benchmark knowledge graphs such as YAGO3-10, FB15K-237, and WN18RR, the KG-LLM framework demonstrated superior performance in tasks like triple classification and relation prediction. Interestingly, the researchers found that fine-tuning relatively smaller models such as LLaMA-7B and ChatGLM-6B yielded better results than recent models like ChatGPT and GPT-4. This is because these smaller models can utilize instruction tuning to bridge the gap between the pre-trained weights in LLMs and KG triple descriptions, enabling them to extract knowledge stored in model parameters more efficiently.

This study highlights the potential of Large Language Models in addressing the incompleteness issue in knowledge graphs, providing a promising avenue for future research and developments in artificial intelligence.

Read the whole article here: http://arxiv.org/abs/2308.13916v2