Predicative integration: the bridge between the two approaches

What is predicative integration?

Learning in AI systems can benefit from integrating Linked Data technologies — associated with knowledge graphs — with approaches that use machine learning and deep learning techniques, among which generative GPT (Generative Pre-trained Transformer) technologies stand out for their capabilities.

The former can aggregate facts to entities recognised as equivalent, while the latter can help integrate new facts identified through statistical patterns. We have termed this operation predicative integration, and it forms the basis for extending knowledge graphs and ontologically consolidating and interpreting inductive learning in such a way that it becomes possible to build operations containing logic on top of it.

This predicative integration is bidirectional and benefits both knowledge graphs and LLMs. On one hand, it makes it possible to use LLMs to enrich knowledge graphs with new facts and relationships discovered through pattern analysis in large volumes of text. On the other hand, knowledge graphs provide the context and semantic structure needed to improve the performance of LLMs and prevent hallucinations.

This operation allows the results of subsymbolic (probabilistic) AI to be integrated into a knowledge graph through ontological interpretation, making it possible to consolidate the results of subsymbolic AI as axioms in inference processes. Conversely, it allows the benefits of symbolic AI to be integrated into subsymbolic AI, providing LLMs with context to improve their coherence and accuracy.

LLMs for knowledge graphs: enhancing graphs with generative AI

Large language models are gaining increasing popularity for their ability to automate and optimise processes across a wide range of domains. In the context of knowledge graphs, LLMs offer significant assistance by improving the extraction of information and knowledge from large volumes of data.

First, they can automatically identify entities in unstructured texts through entity recognition, making it possible to create new nodes in the knowledge graph. Once these entities have been identified, LLMs can be used to detect and extract the relationships between them as mentioned in the text, enriching the relational structure of the graph. In addition to identifying entities and relationships, they can extract attributes associated with entities, completing their representation within the graph.

LLMs also facilitate natural language query processing, enabling relevant information to be retrieved from the knowledge graph. This capability transforms questions expressed in natural language into structured queries that extract precise answers from the graph. As a result, users can explore and navigate the graph more intuitively and efficiently, thanks to the conversational interface that LLMs make possible.

Finally, they are valuable tools for integrating and disambiguating data and entities from multiple sources within the knowledge graph. They can identify entities that, although referred to by different names, refer to the same entity in semantic terms — helping to resolve problems of ambiguity or data duplication through entity disambiguation.

Knowledge graphs for LLMs: providing context and meaning

It is important, however, to acknowledge the limitations of generative AI on its own. LLMs, impressive as they are in their ability to generate language, lack memory and genuine symbolic capability and are therefore unable to evaluate problems in their full context or produce responses based on logical chains of reasoning.

Moving towards a more robust and meaningful AI requires overcoming these limitations by incorporating meaning, logical reasoning and the ability to learn in context. Only in this way is it possible to develop Auditable AI that is traceable and reproducible by third parties, as called for by the National AI Strategy and the EU AI Act.

Knowledge graphs help optimise the performance of LLMs at three critical points in the process.

1. Before processing prompts

Knowledge graphs provide a semantic framework that enables LLMs to understand the context of text more accurately through the standardisation and linking of entities. They can also complement prompts by adding supplementary information, such as the relevant relationships between identified entities.

2. During prompt processing

In scenarios where prompts come from domains in which LLMs show limitations, knowledge graphs can provide additional data sources for fine-tuning the model.

3. After processing prompts

Once the LLM has generated its response, the knowledge graph can carry out a knowledge correction, stabilising and enriching the output by combining entities, relationships and attributes represented in the graph. Issues such as hallucinations or incorrect links can also be mitigated by cross-referencing the generated information against the explicit relationships in the graph. Finally, the knowledge graph also plays a key role in source retrieval: if a source appears linked to attributes or entities in the graph, it can be integrated directly into the response.

One of the main challenges of this approach is enabling systems to recognise unknown entities and integrate them coherently into their knowledge graph. In doing so, this approach transforms scattered information into structured, accessible and contextualised knowledge. It is precisely this capability that justifies describing it as Neurosymbolic AI, as it has the tools needed to generate contexts and link information autonomously and automatically.