Developer builds entity resolver for investing knowledge graph to reduce bad data links

A software developer has detailed the fourth stage of building an investing knowledge graph, focusing on constructing an entity resolver that decides when two mentions refer to the same real-world entity. The system uses a scoring service that returns match probabilities for candidate entity pairs, applying a deliberately high confidence threshold before merging them into a single node. The developer explains that false merges — where two distinct entities are incorrectly combined — are treated as more harmful than false splits, since fabricated graph connections are harder to detect than missing ones. When a new article is processed, the pipeline handles mentions in one of three ways: instant lookup for known aliases, model-scored matching against existing registry entries, or creation of a new canonical entity. This conservative design prioritizes precision over recall, accepting that some real aliases will remain unlinked in order to avoid confidently wrong data in the graph.