固有表現の認識

It finds mentions under a chosen schema. Linking a mention to a particular real-world record is a separate entity-linking task.

主なポイント

• Define types and span boundaries.
• Preserve offsets into the original text.
• Keep recognition separate from identity linking.

Define the entity types and span rules before training or evaluation. Should a company suffix be included? Is a product an organization, a separate type, or outside the schema? Inconsistent annotation rules can make a dataset internally contradictory. NER systems may assign token-level labels and combine adjacent tokens into spans. Subword tokenization requires care when aligning labels with the original text. Preserve character offsets so applications can show exactly which passage produced an extracted value. Evaluate both boundaries and types. Identifying only “Northstar” when the annotated organization is “Northstar Research Labs” may count as a span error even if the general category is correct. Report the matching convention with precision and recall so scores can be interpreted. Context can change the label. “Jordan” might identify a person, country, or organization in different passages. A recognized name is not verified identity information. When using extraction for redaction, search, or record matching, test the downstream outcome and handle ambiguous or missed mentions explicitly.

1. Use the invented sentence “Jordan joined Northstar Research Labs in June.”
2. Mark Jordan as a person mention and Northstar Research Labs as an organization mention under a documented schema.
3. Do not attach a particular biography or company registration unless a separate linking step has evidence for that match.

The constructed example separates locating a name from resolving who or what it identifies.