Recently, ChatGPT, a representative LLM, has gained considerable attention due to its powerful emergent abilities. Some researchers suggest that LLMs could potentially replace structured knowledge bases like knowledge graphs (KGs) and function as parameterized knowledge bases. However, while LLMs are proficient at learning probabilistic language patterns based on large corpus and engaging in conversations with humans, they, like previous smaller pre-trained language models (PLMs), still have difficulty in recalling facts while generating knowledge-grounded contents. To overcome these limitations, researchers have proposed enhancing data-driven PLMs with knowledge-based KGs to incorporate explicit factual knowledge into PLMs, thus improving their performance to generate texts requiring factual knowledge and providing more informed responses to user queries. This paper reviews the studies on enhancing PLMs with KGs, detailing existing knowledge graph enhanced pre-trained language models (KGPLMs) as well as their applications. Inspired by existing studies on KGPLM, this paper proposes to enhance LLMs with KGs by developing knowledge graph-enhanced LLMs (KGLLMs). KGLLM provides a solution to enhance LLMs' factual reasoning ability, opening up new avenues for LLM research.
LLMs like ChatGPT are adept at generating human-like text but are limited in factual accuracy and knowledge-based content.
Knowledge graphs (KGs) are structured databases containing real-world facts and relationships, aiding in tasks that require factual correctness.
The paper proposes knowledge graph-enhanced pre-trained language models (KGPLMs) to integrate KGs into LLMs, improving factual reasoning and response accuracy.
Various approaches to integrating KGs into LLMs include before-training, during-training, and post-training enhancements.
KGPLMs could vastly benefit applications such as entity recognition, relation extraction, sentiment analysis, knowledge graph completion, question answering, and natural language generation.
In the world of artificial intelligence, language models have become increasingly sophisticated. LLMs such as ChatGPT have made strides in understanding and generating human-like text. These models have been trained on vast amounts of data, allowing them to produce content that is often coherent and contextually relevant. However, when it comes to factual accuracy and knowledge-based content generation, LLMs appear to have limitations. They generally do well with the information they were trained on but struggle to recall, apply, or update knowledge that wasn't covered in their training sets.
This is where knowledge graphs (KGs) come into play. KGs are structured databases of real-world facts and relationships and offer an explicit representation of knowledge. They store and present information in a way that is not only accessible but can be easily updated and maintained. As a result, KGs have inherent advantages for tasks requiring factual correctness and up-to-date information.
Enhancing language models with knowledge from KGs — termed here as knowledge graph-enhanced pre-trained language models (KGPLMs) — has been proposed to overcome the limitations of LLMs in factual content generation. The integration of KGs into LLMs is expected to improve the models’ abilities to reason with facts, thus producing more informed and accurate responses.
KGPLMs integrate knowledge into language models in various ways.
Applications of KGPLMs are vast and range across several tasks:
While LLMs have shown potential as knowledge bases, their ability to serve as reliable sources of information is questioned. They often default to probabilistic language patterns acquired during training rather than solid factual recall. By incorporating KGs, however, we can begin to build more robust, factual, and reliable language models, leading to what are being referred to as knowledge graph-enhanced LLMs (KGLLMs).
Developing effective KGLLMs requires careful consideration, such as selecting valuable knowledge to incorporate and methods to avoid the loss of previously learned information. It is also critical to explore ways to improve model interpretability and assess their performance on domain-specific tasks.
As LLMs evolve, it becomes increasingly clear that their full potential can only be unlocked through the integration of structured knowledge from KGs. The marriage of LLMs and KGs does not suggest the obsolescence of either; rather, it heralds a new phase of symbiotic development. KGs provide the foundation of fact-based reasoning that LLMs need to become truly intelligent systems, while LLMs offer flexible, self-updating mechanisms that keep the knowledge within KGs fluid and accessible. Together, they will redefine the possibilities for artificial intelligence in processing and generating human language.
T.-Y. Chang, Y. Liu, K. Gopalakrishnan, B. Hedayatnia, P. Zhou, and D. Hakkani-Tur, “Incorporating commonsense knowledge graph in pretrained models for social commonsense tasks,” in Proceedings of Deep Learning Inside Out (DeeLIO): The First Workshop on Knowledge Extraction and Integration for Deep Learning Architectures. Online: Association for Computational Linguistics, Nov. 2020, pp. 74–79. [Online]. Available: https://aclanthology.org/2020.deelio-1.9