Natural language understanding and artificial intelligence for text analysis have witnessed remarkable advances thanks to large language models (LLMs). These models can generate fluent and coherent texts on various topics and domains, as well as perform a range of natural language processing (NLP) tasks. However, the current paradigm of generative AI for text is limited by the conversational interface, which often fails to capture the rich semantics and pragmatics of natural language. In this blog post, we argue that the future of generative AI for text lies in a hybrid approach that combines graph-based knowledge integration, such as knowledge graphs, and conversational AI.
At Lettria, we have developed a novel platform that combines knowledge graph technology with LLMs to create smarter AI models that generate insights from both knowledge and text. Our platform leverages the power of knowledge graphs to represent and store structured information from various sources, as well as the ability of LLMs to extract and infer new knowledge from unstructured texts. By connecting knowledge graphs with pre-trained language models and custom NLP models tailored to your needs, Lettria delivers a “text to knowledge graph” solution enabling:
- Knowledge graph enrichment: LLMs can discover new relationships, attributes, and connections in text, enriching knowledge graphs with highly accurate information. As models process more data, knowledge graphs become truly intelligent.
- Text structuration: LLMs excel at NLP tasks like disambiguation, co-reference resolution, and parsing - turning unstructured text into structured data and seamlessly integrating it with knowledge graphs.
- Question answering: LLMs, especially those fine-tuned on domain-specific data, provide a natural interface for asking complex questions and gaining knowledge graph-based answers with contextual understanding.
Rather than relying on pre-built conversational AI or using LLMs in isolation, connecting generative models, knowledge graphs, and custom NLP enables responsibly developing solutions that align with business needs. Our hybrid human-AI approach focuses resources on augmenting knowledge and structuring data for a transparent, collaborative system improving over time.
Graph-Based vs Conversational AI: Which Is Right For Your Business?
While graph-based and conversational AI both enable intelligent interfaces, the approaches have significant advantages and disadvantages depending on a company's needs. Understanding use cases where each excels helps determine the optimal solution for developing AI capabilities responsibly.
Graph-based AI connects knowledge as a network of semantic relationships, navigating complexity at scale. This provides:
- Explainability: Graph-based knowledge offers transparency into how the AI generates results or answers questions. This visibility into the "reasoning process" builds trust and enables troubleshooting.
- Scalability: Graphs can scale infinitely as new knowledge connects, suited for domains where information is growing or changing rapidly.
- Flexibility: Graph-based knowledge adapts as relationships or data change, ideal for dynamic use cases. The AI system evolves to new contexts.
- Complex implementation: Developing knowledge graphs requires technical expertise and time. The approach may be over-engineered for simple use cases.
Conversational AI provides an intuitive user interface for querying information or requesting actions. This is ideal when:
- Simplicity: Conversational AI typically has a fast, easy implementation, well-suited for customer service or basic questions at scale.
- Accessibility: The conversational interface is highly accessible for many users across platforms and devices.
- Limited context: Conversational AI struggles with complex questions or discussions requiring an understanding of relationships, context or history. Responses are often generic without personalization.
- Rigidity: Conversational interfaces depend on pre-defined behavioral paths and responses. They lack flexibility for new contexts, limiting value in dynamic use cases.
- Black box: Conversational AI systems can be considered as "black boxes" because their inner workings and decision-making processes are often opaque and difficult to understand. This lack of transparency makes it challenging to troubleshoot issues or align the AI system to new objectives.
Clearly, both paradigms have their advantages and disadvantages, depending on the use case and the domain. Here are some examples of use cases where conversational AI or graph-based AI are more suitable:
- Customer service: Conversational AI excels at handling basic, common questions for many users with a simple, accessible interface. However, it may struggle with complex, uncommon, or ambiguous questions that require deeper understanding or reasoning. In these cases, graph-based AI can provide more context, personalization, and explainability, leveraging knowledge graphs to find relevant information from various sources.
- Virtual assistant: Conversational AI can handle generic queries or requests and trigger simple actions like setting reminders, playing music, or ordering food. However, it may not be suitable for sophisticated tasks that require integration of multiple information sources, such as planning a trip or finding the best deals. Graph-based AI can provide more intelligence and flexibility, connecting information from different domains and services.
- Diagnosis and recommendations: In domains like healthcare and finance, which require understanding complex user histories, relationships between symptoms or risk factors, and data-supported recommendations, graph-based AI is ideal. It offers accuracy, reliability, and explainability and can scale as new conditions, treatments, or insights emerge. Conversational AI may not provide the same level of accuracy and reliability, especially when dealing with sensitive or critical information.
As we can see, conversational AI and graph-based AI have different strengths and weaknesses when it comes to natural language understanding and generation. However, they are not mutually exclusive. In fact, they can be integrated to create hybrid solutions that combine the best of both worlds. For example, Neo4j shows how knowledge graphs can power conversational AI by providing contextual awareness, personalization, natural language understanding, and multimodal interactions. Similarly, Deloitte predicts that the future of conversational AI will involve more integration of knowledge graphs and other sources of structured data to enable more intelligent and human-like interactions.