孙科 · Sun Ke

AI × Energy Systems

EN

Why Structured Semantics Is the Starting Point for Reliable AI Analysis

When people discuss automated AI analysis, they often begin by asking whether the model is intelligent enough, large enough, or capable of longer reasoning chains. In real urban and building applications, however, the workflow is often blocked by data before the model can meaningfully begin.

The reason is straightforward: the same object may have several names, the same indicator may use different boundaries, and related datasets may operate at different spatial or temporal scales. A model can generate syntactically correct code while selecting the wrong data or computing the wrong indicator, producing an answer that looks complete but cannot be audited.

Three barriers appear before reasoning begins

Several problems I have encountered look very different on the surface, yet share the same underlying difficulty.

In urban energy research, yearbooks, surveys, academic studies, and government datasets must be reconciled before energy-service demand can be compared across 63 cities. In building-management systems, raw point names such as AHU_SA_T or VAV-03-SP must be mapped to shared semantics before downstream tasks can understand what they represent. In government and enterprise research, GIS, CAD, survey, archive, population, and market data are scattered across systems, forcing analysts to spend substantial effort finding, renaming, aligning units, and reconciling scales.

The common feature is that the most time-consuming work is not computation itself; it is giving data a meaning that makes computation possible.

Working conclusion: shared semantics and computation rules are not post-processing steps after automated analysis. They are prerequisites for analysis that can begin and deserves to be trusted.

Structure directly changes downstream performance

Structured semantics does more than make data tidy. It constrains how a model selects tables, connects entities, generates queries, and chooses valid reasoning paths.

First, domain knowledge graphs provide entities, relationships, and constraints. The model no longer relies only on surface-level similarity between words; it can retrieve context along explicit domain relationships. In contract review or building-equipment analysis, an answer can point to specific objects and relations rather than vague linguistic associations.

Second, explicit schemas and valid join paths constrain queries. The hardest part of complex SQL is often not syntax, but deciding which tables may be joined, which keys are valid, and whether data at different scales can be combined. Making those relationships explicit establishes an executable computation path before the query is generated.

Third, graph structure limits the boundary of reasoning. When a model must reason over a given graph, only steps supported by permitted relationships are allowed. This does not necessarily make the model know more, but it reduces unsupported leaps, improves explainability, and makes the source of errors easier to trace.

Structured semantics therefore contributes three qualities more important than retrieval alone: accuracy, explainability, and restraint when evidence is insufficient.

Existing systems can use data, but may not know how to compute correctly

Urban data platforms, knowledge graphs, and multi-source databases already support many tasks. Data can be collected, organized, queried, and used for decision-making, community energy estimation, or urban relationship analysis. Yet one direct test is often missing: how should the data actually be computed?

An automated analysis system should answer at least four questions:

  1. Which dataset is appropriate for the task?
  2. Are the spatial, temporal, and object boundaries consistent?
  3. What computation rule applies, and where does it stop being valid?
  4. When information is missing or rules conflict, can the system stop rather than invent an answer?

The first three determine whether the system computes correctly. The fourth determines whether it is reliable. In practice, the greater risk is usually not that a model refuses to answer, but that it produces a fluent answer without the necessary evidence.

The question I want to test next

This leads to the research question that currently interests me most:

Can structured representations of data scope, scale, and computation rules reduce the misuse of data and indicators by LLMs, while helping them abstain when information is insufficient?

One possible system architecture would place an inspectable semantic-and-rule layer between the model and raw data. Data would first be mapped to shared semantics, after which the system would generate a computation plan. Each step would validate entities, scales, units, and joins, and the final result would retain both its data provenance and computation path. If necessary conditions were not met, the system would return the missing information instead of guessing.

The goal is not to make AI appear more intelligent. It is to give automated analysis the boundary awareness expected of an engineering system: knowing what it used, why it computed something in a particular way, and when it should not compute at all.


Literature cues from the first four presentation slides: Tsybina et al. (Nature Cities, 2025); Sun et al. (Energy, 2025); Balaji et al. (Applied Energy, 2018); Zheng et al. (Automation in Construction, 2025); Awotunde et al. (Advanced Engineering Informatics, 2026); Luo et al. (ICML 2025); Ning et al. (NeurIPS 2023); Daneshfar et al. (Advanced Engineering Informatics, 2022).

0%