The “Brain” of AI‑Native Apps: Turning Human Data Translation into Self‑Translating Data

Why Human‑Centred Data Translation Fails

In a 2013 logistics data‑warehouse project, an ETL engineer spent weeks manually mapping the field “运费结算方式” because the source system stored free‑text values like “到付”, “货到付款”, “收货方付”, and “老规矩”. He encoded 2,000 rules in his mind; when the system was upgraded, the knowledge vanished with his retirement.

The Invisible Semantic Translators

Every company has “invisible people” who know why a particular ERP configuration exists and can translate business jargon across systems. Their expertise is not captured in code, and they eventually leave, creating a knowledge gap.

Semantic Hub: Let Data Translate Itself

The proposed semantic hub aims to replace these human translators with a system that can translate data autonomously.

Three‑Layer Architecture

First layer – Entity Resolution. Similar to Master Data Management, this layer dynamically identifies that different IDs in CRM, ERP, or finance refer to the same real‑world entity. Unlike static golden‑record syncing, it evaluates new sources on‑the‑fly using attribute similarity, relational context, and time windows, guided by an ontology‑based inference engine.

Second layer – Semantic Mapping. The hub maps heterogeneous strings (e.g., “货到付款”, “到付”, “收货方付”) to a unified business concept such as “到付结算”. Instead of a hard‑coded mapping table, it learns equivalence relations (synonyms, equivalents) within the constraints of an ontology that acts as a dictionary, while a large language model provides the understanding needed to infer meanings from context.

Third layer – Rule Inference. After the data’s meaning is resolved, explicit business rules (e.g., “信用额度 = 基础额度 + 临时额度 – 已占用额度，且临时额度需审批人当日确认”) are applied. The large model performs the reasoning, and the ontology validates the results, ensuring the inference is both smart and correct.

From Manual ETL to Native Capability

Traditionally, projects build a translation dictionary for each system, write ETL scripts for conversion, and add numerous cleaning rules. The semantic hub consolidates these three steps—reference, conversion, cleaning—into a native capability that the system can execute automatically.

Where a human‑crafted dictionary might contain a few hundred entries, a machine‑oriented semantic model must make all implicit rules explicit, which is the challenging and valuable part of the solution.

Conclusion

As one SAP veteran put it, “For the past twenty years we have enabled non‑technical users to operate systems; in the AI era we must enable systems to understand humans.” The semantic hub serves as the foundational “brain” for AI‑native applications—without it, even the most powerful models, compute, or UI are just a bare machine lacking an operating system.