Pick out named entities

Status

What other tools call it

Meridian's classification; an independent blind rating agreed on 95.7% of 324 rows (variant, version 3).

Background

When to use it

Alteryx's Named Entity Recognition tool writes the entities it finds in each text as a JSON list, giving each entity's label, character index and length.^[2] RapidMiner's Token Classification operator stores its result in a new column as JSON values describing all identified tokens.^[3] Expect a text to yield several entities or none, and unnest the list into one row per entity before counting or joining on it. (Meridian's view)

In NER, the same name can refer to different entities of the same type: "JFK" can be the former president of the United States or his son.^[1] A name can also refer to entities of completely different types, as "JFK" might be the airport in New York.^[1] Alteryx's Named Entity Recognition tool cannot, by default, recognise that Michael Jordan and Air Jordan are the same person.^[2] Treat each extracted name as a string still to be resolved, and match it to a list of known entities before counting by entity. (Meridian's view)

In 2001, research indicated that NER systems developed for one domain did not typically perform well on other domains.^[1] spaCy's models are statistical and strongly depend on the examples they were trained on, so its entity recognition does not always work perfectly and might need tuning.^[4] Alteryx's Named Entity Recognition tool looks only for text in the language the user specifies, and French text in the same data can be run through a second NER process.^[2] Check the extraction on a sample of your own texts before trusting its counts. (Meridian's view)

The method

Full named-entity recognition is often broken down into two problems: detecting names, and classifying them by the type of entity they refer to.^[1] In NER, detection is typically simplified to a segmentation problem in which names are contiguous spans of tokens with no nesting, so "Bank of America" is a single name.^[1] The usual measures of an NER system's output are precision, recall and the F1 score.^[1]

NER systems have been built with linguistic grammar-based techniques and with statistical models such as machine learning, and state-of-the-art systems may combine approaches.^[1] Hand-crafted grammar-based NER systems typically obtain better precision, at the cost of lower recall and months of work by experienced computational linguists.^[1] Statistical NER systems typically require a large amount of manually annotated training data.^[1] Conditional random fields have been a typical choice of classifier for machine-learned NER, and transformers offer token classification using deep learning models.^[1] RapidMiner's Token Classification operator applies models trained on predicting the type of all input tokens, which can be used for named entity recognition.^[3]

Early work on NER systems in the 1990s was aimed primarily at extraction from journalistic articles.^[1]

Compared with the alternatives

A gazetteer is a list of names and their types, such as "General Electric", and can be used to augment any NER system.^[1] Where every name you care about is already on a list, matching the text against that list may be simpler to check than a model. (Meridian's view) In spaCy, entity linking grounds named entities in the real world by resolving a textual entity to a unique identifier from a knowledge base.^[4] Where the question is which company a text mentions, rather than which words name a company, follow NER with entity linking or a lookup against a register. (Meridian's view)

Sources

1. [1]Named-entity recognition - Wikipediasecondary source
2. [2]Named Entity Recognitiondocumentation
3. [3]Token Classification - Rapidminer Documentationdocumentation
4. [4]Linguistic Features · spaCy Usage Documentationdocumentation