How BMJ Clinical Intelligence can work with and complement large language models
Knowledge graphs may leverage, work with, or complement large language models both to extract features, define relationships, and make associations, as well as to potentially fuse knowledge graph approaches and large language models to improve prompt engineering. In large language models, getting the prompt or question right is all important. If you do not get the question right, you are never going to get the answer that you need. So knowledge graphs can help frame the question or prompt by providing continually updated domain specific context and information that can then be submitted to a large language model. The submitted question should be much more likely to get a context-specific, relevant and appropriate response that will answer the question directly and thus augment clinical reasoning appropriately. Knowledge graphs might also actually help to discover new insights that previously were not seen or known about – so discoveries are an exciting part of this. The result should ultimately be a more holistic understanding or complete picture of the patient state within a domain of knowledge. This in turn should make it easier to augment clinical reasoning appropriately and help healthcare professionals make better decisions. The advent of large language models operating at scale with trillions of nodes or parameters that can be used to infer what might be the next word in a chain of thought means we can really start to emulate clinical reasoning in many interesting ways. But the challenge is the notion of hallucinations or fabrications that may result when using a large language model. Sometimes it appears that the large language model just attempts to fill in a gap and so it makes stuff up. This is of course completely unacceptable in clinical decision-making – we would not want a large language model to guess, imagine, or fabricate a drug or test for a particular condition. New research is looking at how large language models and knowledge graphs can complement one another. And the research is showing that they can do so in a variety of interesting ways. For example, the learning from a large language model might be guided by structured knowledge coming from the knowledge graph to make the resulting large language model far more robust and allow access to newer or updated information held in the knowledge graph. The knowledge graph can also assess the output of large language models to validity check or to look for hallucinations or fabrications. Using a knowledge graph with a large language model also allows attribution of the references and source material. And the knowledge graphs and large language models can work together in a fused way – so-called fusion models where parts of the inference might be done by a large language model and parts done by a knowledge graph, with both working in synchrony.
11
Why the BMJ Knowledge Graph
Powered by FlippingBook