KG 통한 생성형 AI 리스크 완화

A Knowledge Graph (KG) is a means to connect and represent knowledge in a domain of interest using a graph structure. It is typically built on top of existing data sources of an organization to link data together at web-scale, combining both structured and/or unstructured information. As opposed to the more commonly used relational data models, a graph model is built as a collection of concepts or entities and the relationships between them.

Knowledge graphs form the foundation of modern data analytics by connecting complex data and metadata in a meaningful way.

• Highly Connected: Knowledge Graphs allow optimization on storing and processing a highly connected data.
• Diverse: Knowledge Graphs allow harmonization and integration of data from diverse data sources.
• Dynamic: Knowledge Graphs provide a flexible and agile model to accommodate the history and evolution of data.
• Contextual: Knowledge Graphs support semantics and meaning of data to be embedded within the data itself.
• Visual: Knowledge Graphs provide an intuitive visualization that allows us to extract insights from data.

Data lineage, root-cause analysis, 360 view of data, compliance & data quality analysis, tracking changes overtime, impact analysis, data cleansing & migration, recommendation (link prediction), metadata management, semantics analysis & reasoning, entity resolution, semantic search and query answering are few use cases of knowledge graphs within companies.

Knowledge graphs enable drawing new conclusions and discovering implicit knowledge based on the explicit facts and relationships encoded within the graph. This is typically achieved by designing a so called Ontology which establishes a shared vocabulary and a set of rules that dictate how data can be interrelated within the knowledge graph.
This capability for inference and reasoning makes knowledge graphs especially powerful tools for tasks such as building recommendation systems and complex data analysis, where understanding context and relationships is crucial.

Knowledge graphs can enhance the reliability, validity, and transparency of generative AI in a business context, helping to mitigate many of its technical and regulatory risks to maximize its benefits.

• Data Validation
• Contextual Understanding
• Controlled/Guided Generation
• Bias Mitigation
• Transparency and Explainability
• Language Independent (Symbolic)

Risks of Generative AI

1. Hallucination
2. 1. Controlled/Guided Generation: Knowledge graphs can guide the generative process by providing a source of reliable and structured information. This can help control the output of the generative AI, reducing the risk of generating misleading or harmful content that could damage the business's reputation or customer relationships.
   2. Data Validation: Knowledge graphs can serve as a reference point to validate the output of generative AI models. If the generated content contradicts the information in the knowledge graph, it can be flagged for review, ensuring business decisions are based on accurate and consistent information.
2. Lack of Provenance and Predictability
3. 1. Transparency and Explainability: Knowledge graphs can make the workings of generative AI models more transparent and explainable by tracking the origin, lineage, and transformation of data within their structured framework. Each entity, relationship, and piece of information is tagged with metadata that captures its source, history, and any alterations made over time. This traceability allows generative AI to verify the authenticity, reliability, and context of the data, ensuring transparency and trustworthiness in the information represented within the graph. This can be crucial for businesses in regulated industries or those that need to justify their decisions to stakeholders.
3. Lack of Context
4. 1. Contextual Understanding: Knowledge graphs provide a structured and meaningful representation of a knowledge domain, enabling generative AI models to better understand the context. This can reduce the risk of hallucination by enforcing only a deterministic and relevant set of outputs, hence, improving customer satisfaction and engagement.
   2. Bias Mitigation: Knowledge graphs can help identify and mitigate biases in the training data for generative AI models. By representing knowledge in a unstructured way, it is easier to spot patterns that may indicate bias, helping businesses to avoid unfair practices and maintain their reputation. To achieve this, the provided knowledge requires to be unbiased too.
4. Language Dependency
5. 1. Language Independent: Knowledge graphs, through their logic and symbolic representation of knowledge, transcend linguistic barriers. Rather than relying on specific languages or terminologies (as done by LLMs), they capture and represent knowledge based on universally understood symbols, relationships, and structures. This inherent design abstracts the essence of knowledge from any particular language, making it universally interpretable and ensuring that the core insights and relationships are accessible, regardless of linguistic context.

* For the "lack of timeliness", when a fact emerges, instead of retraining the entire model, the knowledge graph can be updated with this fresh information. This isn't exclusive to graphs and can be also addressed by technologies such as vector databases.

Scenario: Alice, an employee at a large corporation, is facing intermittent internet connectivity issues on her work laptop.

Navigating Innovation & Knowledge Management Leveraging Knowledge-enriched Generative AI

In the dynamic landscape of organizational strategy, striking a balance between exploiting existing knowledge and exploring new opportunities becomes paramount. The below matrix delineates various strategic postures based on the use, or non-use, of knowledge graphs (KGs) and generative AI (genAI) to navigate this balance. Traditional strategies, often void of technological leverage, might impede innovation and adaptive capacities. Employing KGs alone can ensure that existing knowledge is methodically exploited, fortifying data-driven decisions, albeit potentially stifling innovative leaps. Conversely, utilizing genAI singularly might spur innovative ideation but can drift into unstructured and potentially misaligned explorations. An integrated application of both KGs and genAI into knowledge-enriched AI strives to synthesize the structured exploitation of existing knowledge with the inventive exploration of new terrains, potentially unlocking new horizons while securing foundational knowledge. For practitioners, tactically blending the structured rigidity of KGs with the explorative fluidity of genAI could pave the way towards a robust strategic stance, capable of both safeguarding existing domain knowledge and pioneering into new strategic frontiers. This integrative approach, while seemingly optimal, demands thorough technological integration and proficient strategic alignment to truly harness the synergistic potentials of KGs and genAI in managing the exploration-exploitation tension effectively.

A Comprehensive Approach to Construct Knowledge-enriched Generative AI

Implementing a knowledge-enriched generative AI in an enterprise involves several key steps:

1. Identify Use Case
   Identify the specific business problem or opportunity that the implementation will address. This could be anything from automating content creation to predicting future trends.
2. Data Collection and Preparation
   Collect and prepare the data that the AI model will learn from. This can involve data curation for removing bias and ensuring that the data is suitable for training. This can also include data curation for removal of bias and specific attributes like gender and ethnicity as well as removal of company-sensitive data, as needed.
3. Build the Knowledge Graph
   Use the prepared data to build the knowledge graph. This involves defining the relationships in the graph and populating it with relevant information. This can be facilitated by generative AI to provide a draft of the graph that can be reviewed, modified, and enhanced by domain experts.
4. Train the Generative AI
   Use the training data to feed into the AI and train the AI on how to use the knowledge graph. This involves feeding the data into the AI and adjusting the AI's learning model to improve its performance.
5. Integrate the Knowledge Graph
   Integrate the knowledge graph with the generative AI. This involves creating an interface for generative AI systems that allows them to use the knowledge graph as a source of additional context for the AI or using it to validate and refine the AI's output.
6. Test and Validate
   Test the generative AI's performance against the knowledge graph to ensure that the AI's outputs align with the knowledge graph. This also involves adjusting the AI's model to improve its accuracy and performance.
7. Deploy and Monitor
   Deploy the AI in the intended business environment and monitor its performance over time. This could involve updating the knowledge graph to ensure accuracy, efficiency, and security as necessary.

The specific steps can vary depending on the specific use case, the type of generative AI being used, and the structure and content of the knowledge graph. It is also important to consider factors such as data privacy and security, AI ethics, and regulatory compliance throughout the implementation process.

Constructing a knowledge graph from the ground up is a time-intensive and effortful task. However, in an enterprise setting, generative AI can play a crucial role in streamlining and accelerating the development of these knowledge graphs.

• Populating a Knowledge Graph
  When it comes to populating a knowledge graph, generative AI can be used to extract instances of entities and relationships from text, given an ontology or schema for the graph.
• Data Enrichment for Graph Completion
  With regards to data enrichment for graph completion, generative AI can be used to fill gaps in a knowledge graph by generating new nodes and relationships based on the existing structure and content of the graph. This can help to enrich the knowledge graph and make it more comprehensive. In practice, this task could involve vetting of the added content, to ensure its accuracy.
• Predictive Analysis
  For predictive analysis, generative AI can be used to predict future states of the knowledge graph based on historical data. This can be particularly useful for strategic planning and decision-making in a business context.

In the following section, we introduce a variety of architectural designs that merge knowledge graphs with LLMs. While these serve as a source of inspiration, tailoring them to your organization's unique needs and specifications is essential.

Option 1 is based on Retrieval-Augmented Generation (RAG) to combine the strengths of both knowledge graph-based retrieval and generation-based methods in LLM. Instead of solely generating responses from scratch, RAG first retrieves relevant information from the graph and then uses this retrieved information as a context to guide the generation of detailed and accurate responses.
To convert user's natural language queries into graph queries, prompt engineering must be conducted using a set of representative sample queries.
By merging the precision of knowledge graphs with the fluency of generative models, RAG offers a more informed and contextually-aware output, also resolving the lack of timeliness of LLM.
While this method eliminates the need for fine-tuning the LLM, the enriched outputs are limited to the specific domain addressed by the knowledge graph.

Option 2 is based on Supervised Fine-tuning of LLM with additional structured knowledge fed from a knowledge graph. This requires verbalizing the knowledge graph which means converting the structured information within the graph into some natural language descriptions or narratives, digestible by LLM.

For example, the following graph with four entities and four relationships can be verbalized into:

“Bank X, which operates in the Netherlands, provides mortgage to Company Y. Company Y is located in Amsterdam, the capital city of the Netherlands.”

This approach allows to encompass a wider range of topics, however, it comes with the costly fine-tuning of LLM when information is updated.

Option 3 is an extended version of Option 1 where an additional vector database is added to cover for more topics that cannot get directly retrieved from the knowledge graph by running graph queries. While the knowledge graph explicitly depicts relationships and semantic context between entities, the added vector database provides faster similarity search capabilities, allowing for more nuanced comparisons between data points and providing more implicit and relevant results.

Another point of extension here is the capability of semantic reasoning to amplify the depth and contextual accuracy of the generated outputs. This requires designing an ontology that represents knowledge in a specific domain, outlining entities, their properties, and the relationships between them which allows the drawing of meaningful conclusions beyond explicit data in the knowledge graph.

This approach provides enhanced coverage, deeper semantic reasoning, and access to continually updated comprehensive knowledge, all while eliminating the need for LLM fine-tuning. However, it introduces the challenges of maintaining and synchronizing an additional database as data evolve. Additionally, there are considerations of added processing time, the initial development and upkeep of the ontology, and the associated expenses of implementing a reasoning engine.

As we navigate the dynamic field of artificial intelligence, the integration of knowledge graphs with generative AI stands out as a transformative approach to enable trustworthy and responsible enterprise decisions. Knowledge graphs are characterized by their structured representation of knowledge, where entities (often denoting concepts or objects) and the relationships between them are explicitly defined.
They provide a reliable and actionable map of knowledge to not only capture explicit facts but also enable semantic understanding, allowing for the deduction of implicit knowledge.
On the other hand, generative AI is characterized by its ability to produce new content, be it text, images, music, or other forms of data, that mirrors or emulates the patterns seen in its training data. Merging these two powerful realms into a knowledge-enriched AI ecosystem facilitates the production of outputs that are not only precise but imbued with contextual richness.

Knowledge graphs can consistently and accurately influence and guide the behavior of generative AI models such as LLMs, without the need for dedicated training. As LLMs continue to evolve daily, knowledge graphs offer a means to ensure consistent quality across top-tier models. Consequently, investing in the development of high-quality enterprise knowledge graphs is beneficial, irrespective of the particular LLM in use.

This Whitepaper is just the start of our journey into knowledge-enriched AI. While we have touched upon its importance and have come up with some ideas for its realization, there is much more to unfold in this domain in the upcoming years.
As technology advances, the blend of AI with deep knowledge will offer more innovations and breakthroughs. We are at the beginning of an exciting journey, and there is a lot to look forward to...