Ultimately we need enough local context to narrow it down to one match.

Overall we use basic tokenization to create an array of Pronouns (including all-caps ones because human beings are special) to break down the input text. Then we compare those pronouns against a US City dictionary to perform the named entity extraction. Once a city has been identified we refine the potential candidates down to a single match based on other information near to the found entity, like the State and Zip Code.

The goal of this project is high precision identification vs. loose identification (a.k.a Micro Understanding). So you won't get matches for Cities that don't have any refining context surrounding them.