Controlled vocabularies reduce spelling variation inside one dataset. Integration creates a harder problem: two organizations may divide the same domain differently. “South,” “Northeast,” or “machine learning” can have different boundaries even when the labels match.
A taxonomy is more than a list
A taxonomy contains concepts and hierarchical relationships such as is_a or part_of. Alignment relates concepts across two taxonomies.
Useful relation types include:
| Symbolic idea | Meaning |
|---|---|
| equals | Both concepts cover the same instances |
| includes | A is broader than B |
| is included in | A is narrower than B |
| overlaps | Some, but not all, instances are shared |
| disjoint | No instance can belong to both |
An alignment should state one of these meanings rather than storing a generic “match.”
In plain terms: two folders can have the same name while containing different documents; two different folder names can contain the same documents.
Labels generate candidates, not conclusions
Lexical similarity, synonyms, embeddings, and shared instances can propose candidate mappings. The final relation depends on scope and definition.
For example, one US regional taxonomy may have five regions and another four. Mapping by label alone can miss that one system’s South equals the union of another system’s Southeast and Southwest—or may still not match exactly because state membership differs.
Constraints reveal impossible alignments
Within a taxonomy, sibling concepts may be declared disjoint and children may cover their parent. Cross-taxonomy articulations combine with those constraints. A reasoner can then detect inconsistency or infer relations that were not entered directly.
If evidence is incomplete, several globally consistent alignments may remain. Treat them as possible worlds rather than forcing an arbitrary choice. Additional expert input can eliminate alternatives.
A practical alignment workflow
- Define the scope, version, and intended use of both taxonomies.
- Normalize labels but preserve original identifiers.
- Generate candidate articulations from labels, definitions, and shared instances.
- Ask an expert to choose semantic relations, including uncertainty.
- Run logical consistency checks.
- Inspect inferred relations and remaining possible worlds.
- Publish mappings with provenance, confidence, and version dates.
Mappings are versioned data products. When either taxonomy changes, revalidate the alignment rather than carrying old relations forward silently.
Where alignment differs from entity resolution
Entity resolution asks whether two records denote the same real-world entity. Taxonomy alignment asks how two concept sets relate. Boston in two city tables may be the same entity; Northeast in two regional systems may overlap without being equal.
Review card
- A taxonomy includes concepts plus hierarchical semantics.
- Cross-taxonomy relations should distinguish equality, inclusion, overlap, and disjointness.
- Label similarity proposes candidates but does not prove meaning.
- Logical constraints can infer relations and expose contradictions.
- Ambiguity may require several possible worlds.
- Version alignments and retain expert decisions as provenance.
Controlled vocabulary 能減少單一 dataset 內的 spelling variation；integration 更難，因為兩個組織可能用不同方式切分同一 domain。South、Northeast 或 machine learning 即使 label 相同，boundary 也可能不同。
Taxonomy 不只是一張清單
Taxonomy 包含 concept 與 is_a、part_of 等 hierarchical relationship；alignment 則連接兩套 taxonomy 的 concept。
常見 relation：
| 概念 | 意義 |
|---|---|
| equals | 兩者涵蓋相同 instance |
| includes | A 比 B 廣 |
| is included in | A 比 B 窄 |
| overlaps | 部分 instance 共享，但不完全相同 |
| disjoint | 不可能有共同 instance |
Alignment 應明確保存其中一種語意，而不是籠統的 match。
白話來說： 兩個 folder 可以同名卻裝不同文件；兩個不同名 folder 也可能裝相同文件。
Label 產生 candidate，不產生結論
Lexical similarity、synonym、embedding 與 shared instance 都能提出 candidate mapping；最後 relation 仍取決於 scope 與 definition。
例如一套美國區域 taxonomy 有五區，另一套只有四區。只按 label 對應，可能漏掉某方的 South 等於另一方 Southeast + Southwest 的 union；也可能因 state membership 不同而不是 exact match。
Constraint 找出不可能 alignment
Taxonomy 內可以宣告 sibling disjoint、children cover parent。Cross-taxonomy articulation 與這些 constraint 結合後，reasoner 能找 inconsistency，或推導未直接輸入的 relation。
證據不足時，可能還有多個 globally consistent alignment。應把它們當 possible worlds，不要任選一個冒充真相；再用 expert input 排除選項。
實務 alignment workflow
- 定義兩套 taxonomy 的 scope、version、intended use。
- Normalize label，但保存 original ID。
- 從 label、definition、shared instance 產生 candidate articulation。
- 由 expert 選 semantic relation，包含 uncertainty。
- 執行 logical consistency check。
- 檢查 inferred relation 與剩餘 possible world。
- 發布帶 provenance、confidence、version date 的 mapping。
Mapping 是 versioned data product。任一 taxonomy 更新時都要 revalidate，不能安靜沿用舊 relation。
Alignment 與 entity resolution 的差別
Entity resolution 問兩筆 record 是否代表同一 real-world entity；taxonomy alignment 問兩個 concept set 如何相關。兩張 city table 的 Boston 可能是同一 entity；兩個 regional system 的 Northeast 可能只 overlap，不相等。
複習卡
- Taxonomy 包含 concept 與 hierarchy semantics。
- Cross-taxonomy relation 要分 equality、inclusion、overlap、disjointness。
- Label similarity 只提 candidate，不證明 meaning。
- Logical constraint 能推 relation，也能找 contradiction。
- Ambiguity 可能需要多個 possible world。
- Alignment 要 version，expert decision 要留 provenance。