Core Concepts
A novel fine-tuning framework, Logic-Aware Curriculum Tuning (LACT), is proposed to enhance the complex logical reasoning ability of large language models over knowledge graphs by incorporating logical context and leveraging curriculum learning.
Abstract
The paper presents a novel framework called Logic-Aware Curriculum Tuning (LACT) to improve the complex logical reasoning ability of large language models (LLMs) over knowledge graphs (KGs).
The key highlights are:
LACT incorporates logical context into the fine-tuning corpus by leveraging binary tree decomposition to transform complex first-order logic (FOL) queries into a sequence of simpler subqueries. This stimulates the LLM's ability to decompose and reason over complex logical structures.
LACT employs a curriculum learning strategy to smooth the difficulty gap among different types of complex queries, gradually exposing the LLM to more challenging reasoning tasks.
Extensive experiments demonstrate that LACT significantly outperforms previous embedding-based and PLM-based methods, achieving new state-of-the-art performance on widely used datasets like FB15K, FB15K-237 and NELL995.
Further analyses reveal that LACT's performance is correlated with the completeness of relevant information extracted from the KG and the complexity of the logical queries. LACT shows stronger improvements on more challenging reasoning tasks.
Case studies illustrate how LACT can leverage the provided KG context to perform step-by-step logical reasoning, in contrast to the limitations of pure prompt-based approaches.
Overall, the LACT framework effectively enhances the complex logical reasoning capabilities of LLMs by incorporating logical structure and leveraging curriculum learning, demonstrating the benefits of combining knowledge graphs and large language models.
Stats
The set of entities E is connected to entity Africa by relation adjoins.
The set of entities F is connected to entity Beckham by relation owner.
The set of entities G is the intersection of sets E and F.