Co-reference resolution

Here's a demonstration of the concept of co-reference resolution, or finding multiple references to the same entity in unstructured text which appear in different variations.

The Allen Institute for AI has an online search engine which will find references to entities in inputted text. See here.

As we can see, co-reference resolution finds references to people, locations, and organizations, which appear in complete, incomplete, and pronoun form, and groups those forms together. The Allen Institute uses its end-to-end neural coreference resolution model which has achieved an F1 score as high as 78.87% on some data sets. As discussed in the Tip of the Night for June 11, 2016, a F1 score measures the weighted average of precision and recall. The Allen Institute F1 score reflects a precision score of around 80% (how many hits in the results are true hits) and a recall of about 73% (how many of the total true hits in the source data show up in the results). See the research paper posted here. The Allen Institute was founded by Microsoft head Paul Allen and has created an open source NLP library.