cirrus

quick and dirty entity recognition

Cirrus attempts to serialize arbitrary natural text into a Go value using dictionary matching. At the moment it doesn't really do much, and shouldn't be used by anybody.

Cirrus tries to parse tokens into various values based on pretty simple heuristics like capitalization and the presence of a dollar sign. Some entities, like cardinality ("one", "many", "ten", etc) have to be matched against every token.

Roadmap

• grouping of sequential entities
  • e.g. sequential cardinals can be grouped, two dozen becomes Result{Value: 24}
• text classification
  • use wordnet?

Examples

• 2021-10-22 is a date
• $20 is a value in USD
• 20mph is a value in miles per hour
• Hong Kong is a city
• Charles Dickens is a person
• Microsoft or MSFT is a company
• Australia is a country

Numbers

1.2
10e-3
1.00001
20E10
15
2/3
one
two thousand three hundred seventy five

Dates

Christmas eve
10/22/15
10.22.15
2005
december 5 2005 // variations thereof

Times

12:10
ten minutes and five seconds

Phone Numbers

Addresses

Email Addresses

• probably best to use an established regexp for this

Web Addresses

• best to use an established regexp or std's url.Parse

Tables

| name | age |
| john | 2 |
| jane | 12 |

name,age
john,2
jane,12

name age
"john" 2
"jane" 12

name    age
john    2
jane    12