How is a Knowledge Graph with Millions of Nodes Built?

Why Enterprises Need Knowledge Graphs

In enterprise operations, brand tone, product selling points, customer insights, and business scenarios—each type of information itself constitutes a vast knowledge system, and there are complex associations between these systems. When AI intervenes in content production and business decision-making, its first challenge is not generation capability, but knowledge structure: does it truly understand the relationships between concepts like 'brand,' 'scenario,' and 'audience,' or does it merely stitch words together?

The knowledge graph addresses this issue. It organizes concepts, entities, and relationships into a graph structure, allowing machines to reason along relationship paths rather than just performing word matching. Tezign's enterprise knowledge graph has over a million nodes, covering multiple dimensions such as brands, products, scenarios, people, and industry terminology, constructed using unsupervised Open Information Extraction (OIE) technology.

Why Traditional Approaches Don't Work

Before knowledge graphs, enterprises typically managed knowledge using manual dictionaries or structured databases. The former is labor-intensive, slow to update, and incomplete; the latter relies on pre-designed table structures, requiring schema changes when encountering new concepts.

The more fundamental issue is that both methods record 'facts' but do not record 'relationships.' Knowing 'a certain brand is a technology company' and knowing 'the technology direction of that brand,' 'the core scenarios corresponding to the technology direction,' and 'the capabilities that the scenario depends on' are entirely different levels of cognitive understanding. The latter is the knowledge structure that can truly drive AI reasoning.

There is also a scale issue. Enterprise information updates rapidly, with new brands, new products, and new demands continuously emerging. Achieving a scale of a million nodes through manual annotation is impossible—the upper limit of manual annotation is determined by annotation costs, while the growth rate of enterprise data far exceeds this limit.

What is Unsupervised OIE

Open Information Extraction (OIE) automatically extracts structured triples from natural language text, in the form of 'subject-relation-object.' For example, from 'a certain brand launched a new series in 2024,' it extracts (certain brand, launched, new series).

The meaning of 'open' is that it does not predefine relationship types. Traditional relation extraction requires a set of labels (such as 'belongs to' or 'competitor is') to be determined in advance, and then trains a model to recognize predefined relationships. OIE does not impose this constraint; it extracts all visible semantic relationships from the text as much as possible, regardless of whether these relationships appear in a predefined list.

'Unsupervised' means that no manually annotated training samples are needed. This is key to achieving a million nodes—OIE directly infers triples from the language structure of the original text, without relying on labeled data, with the scale limit determined by the corpus size rather than annotation costs.

From Multimodal Assets to Graph Structure

Tezign's knowledge graph processes not only text but also multimodal assets such as images and videos. The complete extraction path consists of three stages.

First stage, multimodal content understanding. Recognize content in images and videos—identifying brand elements, product categories, scenario types, and character features, transforming unstructured visual content into processable text descriptions.

Second stage, triple extraction. Run the OIE model on the text to identify entities and relationships, outputting a large number of raw triples. The quality of this step varies, containing both genuine semantic relationships and noise generated by linguistic phenomena (metaphors, ellipses, references).

Third stage, graph integration and cleaning. Merge the raw triples into a unified graph structure. This process requires handling entity alignment—'Apple,' 'Apple Inc.,' and '苹果公司' may refer to the same entity in different documents, requiring identification and merging. At the same time, filter out relationships with low confidence, retaining high-quality connections.

How the Graph Functions in AI Products

The completed knowledge graph is not a static archive but the infrastructure for AI reasoning, functioning on three levels.

First, enhanced retrieval. When users present content needs, the system not only searches for assets that match keywords but also recalls semantically related content along the graph's relationship paths—what brands, product categories, and visual styles have historically been strongly related to this scenario, with retrieval quality far exceeding pure text search.

Second, generation constraints. When generating content for a certain brand, the model retrieves the brand's associated attributes (price range, target audience, brand tone) from the graph, ensuring that the generated content remains consistent across these dimensions, avoiding tonal shifts.

Third, continuous updates. New assets are continuously ingested, and the OIE model continuously extracts new triples from them, integrating them into the existing graph. Nodes and relationships grow as business data increases, thickening the knowledge foundation over time.

Notable Boundaries

The extraction accuracy of unsupervised OIE is not as high as that of supervised methods; the scale advantage comes at the cost of needing to ensure the reliability of each triple through post-processing. Confidence filtering and manual sampling verification are indispensable steps in the engineering implementation.

Another limitation comes from domain specificity. General OIE models typically perform below expectations on specialized enterprise corpora because industry language is filled with jargon, brand neologisms, and context-dependent implicit relationships. Adapting to vertical domain corpora is a necessary step; directly applying general models will yield suboptimal results.

The value of the knowledge graph also highly depends on the design of downstream tasks. The graph itself is merely a structure; whether it can function in AI products depends on whether the retrieval, orchestration, and generation modules know how to utilize it. Understanding the graph as an external memory that AI reasoning can call upon, rather than as the 'brain' of AI, will prevent unrealistic expectations of it.