Using InfraNodus, you can generate knowledge graphs and ontologies for any topic, text, external files, or knowledge bases. This is particularly useful for research, as it helps uncover crucial elements and relationships within any context. However, knowledge graphs can also enhance AI- and LLM-related tasks by augmenting datasets and improving RAG functionality. For instance, you can generate a knowledge graph from a set of PDF files and then use this structural information to refine AI prompts—reducing hallucinations and enforcing strict rules based on the relationships identified in your documents. The knowledge graph of ontologies visualized as a network can also be used to evaluate how optimal the structure is.
For instance, in the example below, it has a diverse topical structure but is too focused on a central concept, which is an indication that peripheral clusters could be developed further.
In this article, we will demonstrate our approach for generating a knowledge graph (and a simple ontology) from any text or extracting it from existing data. We will also show how the InfraNodus graph can provide a quick overview of the main clusters in this ontology, help identify structural gaps (topical clusters that are not well-connected) and the most influential nodes (which may also be points of vulnerability), and reveal underlying ideas. All this can be very useful to have an objective (and not AI-hallucinated) overview of a knowledge base, as it doesn't use any AI in the process. The insights obtained from the graph can be used to optimize the knowledge base structure by adding new content or improving semantic links within the data set. The resulting knowledge graphs can be leveraged to generate context-aware AI prompts and inspire new ideas based on the structure you’re analyzing.
For example, here's a graph generated on acupuncture using this approach below. The bigger nodes are the more influential nodes in the network (e.g. we can quickly see that qi stands out, indicating an important concept). We can also see how the ideas are organized into distinct clusters: "healing techniques", "energy pathways", etc. — giving us a high-level idea of the main topics present inside.
The statements contain detailed information about the relations and can be seen when the nodes of the graph are selected. Note the simplified syntax we use that is human-readable and is compatible with PKM tools such as Obsidian and Roam Research:
As there is an AI module built into InfraNodus, you can query this knowledge graph directly to generate more ideas. For example, you can focus on a structural gap to develop it further. With a standard AI chat interface, such degree of precision would not be possible, as you'd have to rely on an AI model to make that selection:
A Simple Knowledge Graph Syntax Format
Most of the texts on knowledge graphs use complex syntax and software for building the knowledge graphs. Typical formats include RDF, OWL, JSON, Neo4J-backed Cypher database language, and others. The main idea behind all of them is to describe the nodes (main elements) and the relations. But these are not easily understandable by humans and make everything more complex. Moreover, there is no evidence that AI models understand these formats better. That's why in the case of InfraNodus we implemented a very simple and familiar syntax that is already widely used in other applications, such as Obsidian and Roam Research.
The syntax is generated using the following logic:
1. The entities (nodes) are highlighted using [[wiki links]]
2. The relation is described in plain language and are tagged using [squarebrackets]
3. The relations can represent types, axioms, rules, classes, etc.
As a result, you can take a plain statements, such as
Qi is balanced by therapies such as acupuncture, Tai Chi, and Herbal Medicine.
and convert it to:
[[Qi]] is balanced by therapies such as [[Acupuncture]], [[Tai Chi]], and [[Herbal Medicine]] [balancedBy]
This syntax can be understood by humans, it can be used in any popular PKM, and it can also be visualized as a graph.
You can generate the ontology for any topic or your own text with on your own or using our Custom GPT: https://chatgpt.com/g/g-679925713a4c8191a0a8fbb1f97ff027-ontologies-generator (in beta, so we welcome your feedback).
Ontology Graph Visualization Logic
Once you have the ontology generated in this format, you can visualize it as a graph in InfraNodus. The entities are represented as nodes and co-occurrences are the relations. Based on this representation, you can identify the most influential nodes (which can also be the most vulnerable points of failure), topical clusters (groups of entities that belong together), structural gaps (clusters that could be better connected), and estimate how optimally connected the structure is (bias | focused | diverse | dispersed).
The logic for graph visualization in InfraNodus is the following:
1. Every [[wiki link]] is a node (or an entity) — e.g., [[Qi]] and [[Acupuncture]]
2. When [[wiki links]] appear in the same statement (paragraph), they will be connected (using 4-gram scanning as described in our paper, but it's possible to control that by limiting every paragraph to only two entities). E.g., [[Qi]] is balanced by [[Acupuncture]]
3. The statement acts as a description of the relation (which can be used by the underlying AI model), but it also is condensed into a standard relationship type specified in [squarebrackets] — e.g., "balanced by therapies" becomes [balancedBy]
The graph can then be filtered by relationship types, so it is possible to only see the "balancing" relations or "instance of" kind of relations, which is very useful for analysis.
To generate a graph, go to https://infranodus.com/apps and choose "Add a new text" and then copy-paste the generated ontology.
Comments
0 comments
Please sign in to leave a comment.