Contents 4
Sufficient consensus
Named entity recognition identifies people, organizations, places and products in text and links them to knowledge base records. It supports disambiguation, structured data and knowledge panels.
Consistent naming across sources is what makes it work.
If the systems had learned what a name looks like, new names would score close to known ones. They score 22 to 35 points lower.
Markup can state the type of a thing on the page before any model has to infer it. See Entity Recognition.
| System | F1 on names seen in training | F1 on unseen names |
|---|---|---|
| Stanford NER | 83.79 | 48.48 |
| SENNA | 79.85 | 58.08 |
| CRFSuite | 75.02 | 45.61 |
Augenstein, I., Derczynski, L. and Bontcheva, K., Generalisation in named entity recognition: A quantitative analysis, Computer Speech and Language, 2017. Macro average over 19 English test corpora.
What the result pages leave out
These belong to the subject. People ask about them. They are missing from the agreed coverage.
- Confusion with a larger namesake. A shared name means the bigger entity absorbs the references, and correcting it takes years.
- Entities that do not exist yet. A new method, product or concept has no record, so every mention is either dropped or misattached.
- The cost of renaming. Rebranding resets the accumulated recognition, and the old name keeps attracting the references.
- Inconsistency across your own material. Three spellings of your organization on three platforms produce three weak entities instead of one.
- Recognition without reputation. Being identified reliably says nothing about being trusted, and the two are regularly confused.
What people actually want to know
- Does any system reliably know who we are?
- Which other organization are we being confused with?
- Do we spell our own name the same way everywhere?
- What is the name for the thing we invented?
Consistent naming is maintenance work with slow returns. Skipping it makes every other effort harder to attribute.