Skip to main content

Identify locations and tag them with UN-LOCODEs and ISO-3166-2 subdivisions.

Project description

berlin-rs

A Python/Rust microservice to identify locations and tag them with UN-LOCODEs and ISO-3166-2 subdivisions.

Getting started

To test the Rust API locally:

  make run

This will make an API available on port 3001. It serves simple requests of the form:

curl 'http://localhost:3001/berlin/search?q=house+prices+in+londo&state=gb' | jq

replacing localhost with the local endpoint (jq used for formatting).

This will return results of the form:

{
  "time": "32.46ms",
  "query": {
    "raw": "house prices in londo",
    "normalized": "house prices in londo",
    "stop_words": [
      "in"
    ],
    "codes": [],
    "exact_matches": [
      "house"
    ],
    "not_exact_matches": [
      "house prices",
      "prices in",
      "prices",
      "in londo",
      "londo"
    ],
    "state_filter": "gb",
    "limit": 1,
    "levenshtein_distance": 2
  },
  "results": [
    {
      "loc": {
        "encoding": "UN-LOCODE",
        "id": "gb:lon",
        "key": "UN-LOCODE-gb:lon",
        "names": [
          "london"
        ],
        "codes": [
          "lon"
        ],
        "state": [
          "gb",
          "united kingdom of great britain and northern ireland"
        ],
        "subdiv": [
          "lnd",
          "london, city of"
        ]
      },
      "score": 1346,
      "offset": {
        "start": 16,
        "end": 21
      }
    }
  ]
}

A Python wheel can also be built, using

  make wheels
  pip install build/wheels/berlin-0.1.0-xyz.whl

where xyz is your architecture.

Afterwards berlin should be functional inside a python shell/script. Example:

import berlin

db = berlin.load('../data')
loc = db.query('manchester population', 'gb', 1)[0];
print("location:", loc.words)

Description

Berlin is a location search engine which works on an in-memory collection of all UN Locodes, subdivisions and states (countries). Here are the main architectural highlights: On startup Berlin does a basic linguistic analysis of the locations: split names into words, remove diacritics, transliterate non-ASCII symbols to ASCII. For example, this allows us to find “Las Vegas” when searching for “vegas”. It employs string interning in order to both optimise memory usage and allow direct lookups for exact matches. If we can resolve (parts of) the search term to an existing interned string, it means that we have a location with this name in the database.

When the user submits the search term, Berlin first does a preliminary analysis of the search term: 1) split into words and pairs of words 2) try to identify the former as existing locations (can be resolved to existing interned strings) and tag them as “exact matches”. This creates many search terms from the original phrase. Pre-filtering step. Here we do three things 1) resolve exact matches by direct lookup in the names and codes tables 2) do a prefix search via a finite-state transducer 3) do a fuzzy search via a Levenshtein distance enabled finite-state transducer. The pre-filtered results are passed through a string-similarity evaluation algorithm and sorted by score. The results below a threshold are truncated. A graph is built from the locations found during the previous step in order to link them together hierarchically if possible. This further boosts some locations. For example, if the user searches for “new york UK” it will boost the location in Lincolnshire and it will show up higher than New York city in the USA. It is also possible to request search only in a specific country (which is enabled by default for the UK)

Berlin is able to find locations with a high degree of semantic accuracy. Speed is roughly equal to 10-15 ms per every non-matching word (or typo) + 1 ms for every exact match. A complex query of 8 words usually takes less than 100 ms and all of the realistic queries in our test suite take less than 50 ms, while the median is under 30 ms. Short queries containing an exact match (case insensitive) are faster than 10 ms.

The architecture would allow to easily implement as-you-type search suggestions in under 10 milliseconds if deemed desirable.

License

Prepared by Flax & Teal Limited for ONS Alpha project. Copyright © 2022, Office for National Statistics (https://www.ons.gov.uk)

Released under MIT license, see LICENSE for details.

License

Prepared by Flax & Teal Limited for ONS Alpha project. Copyright © 2022, Office for National Statistics (https://www.ons.gov.uk)

Released under MIT license, see LICENSE for details.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

berlin-0.3.0.tar.gz (15.6 kB view hashes)

Uploaded Source

Built Distributions

berlin-0.3.0-pp37-pypy37_pp73-manylinux_2_5_x86_64.manylinux1_x86_64.whl (7.4 MB view hashes)

Uploaded PyPy manylinux: glibc 2.5+ x86-64

berlin-0.3.0-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.whl (7.4 MB view hashes)

Uploaded CPython 3.10 manylinux: glibc 2.5+ x86-64

berlin-0.3.0-cp39-cp39-manylinux_2_5_x86_64.manylinux1_x86_64.whl (7.4 MB view hashes)

Uploaded CPython 3.9 manylinux: glibc 2.5+ x86-64

berlin-0.3.0-cp38-cp38-manylinux_2_5_x86_64.manylinux1_x86_64.whl (7.4 MB view hashes)

Uploaded CPython 3.8 manylinux: glibc 2.5+ x86-64

berlin-0.3.0-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (8.0 MB view hashes)

Uploaded CPython 3.7m manylinux: glibc 2.17+ x86-64

berlin-0.3.0-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.whl (7.4 MB view hashes)

Uploaded CPython 3.7m manylinux: glibc 2.5+ x86-64

berlin-0.3.0-cp36-cp36m-manylinux_2_5_x86_64.manylinux1_x86_64.whl (7.4 MB view hashes)

Uploaded CPython 3.6m manylinux: glibc 2.5+ x86-64

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page