libpostal/README.md

# libpostal: international street address NLP

[![Build Status](https://travis-ci.org/openvenues/libpostal.svg?branch=master)](https://travis-ci.org/openvenues/libpostal)

libpostal is a fast statistical parser/normalizer for street addresses around the world.

:jp: :us: :gb: :ru: :fr: :kr: :it: :es: :cn: :de:

Addresses and the geographic coordinates they represent are essential for any
location-based application (map search, transportation, on-demand/delivery
services, check-ins, reviews). Yet even the simplest addresses are packed with
local conventions, abbreviations and context, making them difficult to
index/query effectively with traditional full-text search engines, which are
designed for document indexing. This library helps convert the free-form
addresses that humans use into clean normalized forms suitable for machine
comparison and full-text indexing.

While libpostal is not itself a full geocoder, it can be used as a preprocessing step to make any geocoding application smarter, simpler, and more consistent internationally.

The core library is written in pure C. Language bindings for [Python](https://github.com/openvenues/pypostal) and [NodeJS](https://github.com/openvenues/node-postal) are officially supported and it's easy to write bindings in other languages.

Examples of normalization
-------------------------

The expand_address API converts messy real-world addresses into normalized
equivalents suitable for search indexing, hashing, etc. Here's a code example
using the Python API for succinctness:

```python

from postal.expand import expand_address
expansions = expand_address('Quatre vignt douze Ave des Champs-Élysées')

assert '92 avenue des champs-elysees' in set(expansions)
```

libpostal contains an OSM-trained language classifier to detect which language(s) are used in a given
address so it can apply the appropriate normalizations. The only input needed is the raw address string. 
Here's a short list of some less straightforward normalizations in various languages.

| Input                               | Output (may be multiple in libpostal)   |
| ----------------------------------- |-----------------------------------------|
| One-hundred twenty E 96th St        | 120 east 96th street                    |
| C/ Ocho, P.I. 4                     | calle 8 polígono industrial 4           |
| V XX Settembre, 20                  | via 20 settembre 20                     |
| Quatre vignt douze R. de l'Église   | 92 rue de l' église                     |
| ул Каретный Ряд, д 4, строение 7    | улица каретныи ряд дом 4 строение 7     |
| ул Каретный Ряд, д 4, строение 7    | ulitsa karetnyy ryad dom 4 stroyeniye 7 |
| Marktstrasse 14                     | markt straße 14                         |

libpostal currently supports these types of normalizations in *60+ languages*,
and you can [add more](https://github.com/openvenues/libpostal/tree/master/resources/dictionaries) 
(without having to write any C).

Now, instead of trying to bake address-specific conventions into traditional
document search engines like Elasticsearch using giant synonyms files, scripting,
custom analyzers, tokenizers, and the like, geocoding can look like this:

1. Run the addresses in your database through libpostal's expand_address
2. Store the normalized string(s) in your favorite search engine, DB, 
   hashtable, etc.
3. Run your user queries or fresh imports through libpostal and search
   the existing database using those strings

In this way, libpostal can perform fuzzy address matching in constant time
relative to the size of the data set.

For further reading and some bizarre address edge-cases, see:
[Falsehoods Programmers Believe About Addresses](https://www.mjt.me.uk/posts/falsehoods-programmers-believe-about-addresses/).

Examples of parsing
-------------------

libpostal implements the first statistical address parser that works well internationally,
trained on ~50 million addresses in over 100 countries and as many
languages. We use OpenStreetMap (anything with an addr:* tag) and the OpenCage
address format templates at: https://github.com/OpenCageData/address-formatting
to construct the training data, supplementing with containing polygons and
perturbing the inputs in a number of ways to make the parser as robust as possible
to messy real-world input. Here's a code example, again using the Python API:

```python

from postal.parser import parse_address
parse_address('The Book Club 100-106 Leonard St Shoreditch London EC2A 4RH, United Kingdom')
```

These example parse results are taken from the interactive address_parser program 
that builds with libpostal when you run make. Note that the parser doesn't care about commas
vs. no commas, casing, or different permutations of components (if the input is e.g. just
a city or just city/postcode).

```
> 781 Franklin Ave Crown Heights Brooklyn NYC NY 11216 USA

Result:

{
  "house_number": "781",
  "road": "franklin ave",
  "suburb": "crown heights",
  "city_district": "brooklyn",
  "city": "nyc",
  "state": "ny",
  "postcode": "11216",
  "country": "usa"
}

> The Book Club 100-106 Leonard St, Shoreditch, London, Greater London, England, EC2A 4RH, United Kingdom

Result:

{
  "house": "the book club",
  "house_number": "100-106",
  "road": "leonard st",
  "suburb": "shoreditch",
  "city": "london",
  "state_district": "greater london",
  "state": "england",
  "postcode": "ec2a 4rh",
  "country": "united kingdom"
}

> Museo del Prado C. de Ruiz de Alarcón, 23 28014 Madrid Madrid, Spain

Result:

{
  "house": "museo del prado",
  "road": "c. de ruiz de alarcón",
  "house_number": "23",
  "postcode": "28014",
  "state": "madrid",
  "city": "madrid",
  "country": "spain"
}

> Double Shot Tea & Coffee 15 Melle St. Braamfontein Johannesburg, 2000, South Africa

Result:

{
  "house": "double shot tea & coffee",
  "house_number": "15",
  "road": "melle st.",
  "suburb": "braamfontein",
  "city": "johannesburg",
  "postcode": "2000",
  "country": "south africa"
}

> Eschenbräu Bräurei Triftstraße 67, 13353 Berlin, Deutschland
        
Result:
        
{
  "house": "eschenbraeu braeurei",
  "road": "triftstrasse",
  "house_number": "67",
  "postcode": "13353",
  "city": "berlin",
  "country": "deutschland"
}

> Szimpla Kert Kazinczy utca 14 Budapest 1075, Magyarország

Result:

{
  "house": "szimpla kert",
  "road": "kazinczy utca",
  "house_number": "14",
  "city": "budapest",
  "postcode": "1075",
  "country": "magyarország"
}

> Государственный Эрмитаж Дворцовая наб., 34 191186, St. Petersburg, Russia

Result:

{
  "house": "государственный эрмитаж",
  "road": "дворцовая наб.",
  "house_number": "34",
  "postcode": "191186",
  "city": "st. petersburg",
  "country": "russia"
}
```

The parser achieves very high accuracy on held-out data, currently 98.9%
correct full parses (meaning a 1 in the numerator for getting *every* token
in the address correct).

Installation
------------

Before you install, make sure you have the following prerequisites:

**On Linux (Debian)**
```
sudo apt-get install libsnappy-dev autoconf automake libtool
```

**On Mac OSX**
```
sudo brew install snappy autoconf automake libtool
```

Then to install the C library:

```
git clone https://github.com/openvenues/libpostal
cd libpostal
./bootstrap.sh
./configure --datadir=[...some dir with a few GB of space...]
make
sudo make install

# On Linux it's probably a good idea to run
sudo ldconfig
```

Bindings
--------

Libpostal is designed to be used by higher-level languages. Here are the officially-supported language bindings:

- Python: [pypostal](https://github.com/openvenues/pypostal)
- NodeJS: [node-postal](https://github.com/openvenues/node-postal)

If you don't see your language of choice, or if you're writing a language binding, please let us know!

Command-line usage (expand)
---------------------------

After building libpostal:

```
cd src/

./libpostal "Quatre vignt douze Ave des Champs-Élysées"
```

If you have a text file or stream with one address per line, the command-line interface also accepts input from stdin:

```
cat some_file | ./libpostal --json
```

Command-line usage (parser)
---------------------------

After building libpostal:

```
cd src/

./address_parser
```

address_parser is an interactive shell. Just type addresses and libpostal will
parse them and print the result.

Tests
-----

libpostal uses [greatest](https://github.com/silentbicycle/greatest) for automated testing. To run the tests, use:

```
make check
```

Adding [test cases](https://github.com/openvenues/libpostal/tree/master/test) is easy, even if your C is rusty/non-existent, and we'd love contributions. We use mostly functional tests checking string input against string output.

libpostal also gets periodically battle-tested on tens of millions of addresses from OSM (clean) as well as anonymized queries from a production geocoder (not so clean). During this process we use valgrind to check for memory leaks and other errors.

Data files
----------

libpostal needs to download some data files from S3. The basic files are on-disk
representations of the data structures necessary to perform expansion. For address
parsing, since model training takes about a day, we publish the fully trained model 
to S3 and will update it automatically as new addresses get added to OSM. Same goes for
the language classifier model.

Data files are automatically downloaded when you run make. To check for and download
any new data files, run:

```
libpostal_data download all $YOUR_DATA_DIR/libpostal
```

And replace $YOUR_DATA_DIR with whatever you passed to configure during install.

Features
--------

- **Abbreviation expansion**: e.g. expanding "rd" => "road" but for almost any
language. libpostal supports > 50 languages and it's easy to add new languages
or expand the current dictionaries. Ideographic languages (not separated by
whitespace e.g. Chinese) are supported, as are Germanic languages where
thoroughfare types are concatenated onto the end of the string, and may
optionally be separated so Rosenstraße and Rosen Straße are equivalent.

- **International address parsing**: sequence model which parses
"123 Main Street New York New York" into {"house_number": 123, "road":
"Main Street", "city": "New York", "state": "New York"}. The parser works
for a wide variety of countries and languages, not just US/English. 
The model is trained on > 50M OSM addresses, using the
templates in the [OpenCage address formatting repo](https://github.com/OpenCageData/address-formatting) to construct formatted,
tagged traning examples for most countries around the world. Many types of [normalizations](https://github.com/openvenues/libpostal/blob/master/scripts/geodata/osm/osm_address_training_data.py)
are performed to make the training data resemble real messy geocoder input as closely as possible.

- **Language classification**: multinomial logistic regression
trained on all of OpenStreetMap ways, addr:* tags, toponyms and formatted
addresses. Labels are derived using point-in-polygon tests in Quattroshapes
and official/regional languages for countries and admin 1 boundaries
respectively. So, for example, Spanish is the default language in Spain but
in different regions e.g. Catalunya, Galicia, the Basque region, the respective 
regional languages are the default. Dictionary-based disambiguation is employed in
cases where the regional language is non-default e.g. Welsh, Breton, Occitan.
The dictionaries are also used to abbreviate canonical phrases like "Calle" => "C/"
(performed on both the language classifier and the address parser training sets)

- **Numeric expression parsing** ("twenty first" => 21st, 
"quatre-vignt-douze" => 92, again using data provided in CLDR), supports > 30
languages. Handles languages with concatenated expressions e.g.
milleottocento => 1800. Optionally normalizes Roman numerals regardless of the
language (IX => 9) which occur in the names of many monarchs, popes, etc.

- **Fast, accurate tokenization/lexing**: clocked at > 1M tokens / sec,
implements the TR-29 spec for UTF8 word segmentation, tokenizes East Asian
languages chracter by character instead of on whitespace.

- **UTF8 normalization**: optionally decompose UTF8 to NFD normalization form,
strips accent marks e.g. à => a and/or applies Latin-ASCII transliteration.

- **Transliteration**: e.g. улица => ulica or ulitsa. Uses all
[CLDR transforms](http://www.unicode.org/repos/cldr/trunk/common/transforms/), the exact same source data as used by [ICU](http://site.icu-project.org/),
though libpostal doesn't require pulling in all of ICU (might conflict 
with your system's version). Note: some languages, particularly Hebrew, Arabic
and Thai may not include vowels and thus will not often match a transliteration 
done by a human. It may be possible to implement statistical transliterators
for some of these languages.

- **Script detection**: Detects which script a given string uses (can be
multiple e.g. a free-form Hong Kong or Macau address may use both Han and
Latin scripts in the same address). In transliteration we can use all
applicable transliterators for a given Unicode script (Greek can for instance
be transliterated with Greek-Latin, Greek-Latin-BGN and Greek-Latin-UNGEGN).

Roadmap
-------

- **Geographic name aliasing (coming soon)**: New York, NYC and Nueva York alias
to New York City. Uses the crowd-sourced GeoNames (geonames.org) database, so alternate
names added by contributors can automatically improve libpostal.

- **Geographic disambiguation (coming soon)**: There are several equally
likely Springfields in the US (formally known as The Simpsons problem), and
some context like a state is required to disambiguate. There are also > 1200
distinct San Franciscos in the world but the term "San Francisco" almost always
refers to the one in California. Williamsburg can refer to a neighborhood in
Brooklyn or a city in Virginia. Geo disambiguation is a subset of Word Sense
Disambiguation, and attempts to resolve place names in a string to GeoNames
entities. This can be useful for city-level geocoding suitable for polygon/area
lookup. By default, if there is no other context, as in the San Francisco case,
the most populous entity will be selected.

- **Ambiguous token classification (coming soon)**: e.g. "dr" => "doctor" or
"drive" for an English address depending on the context. Multiclass logistic
regression trained on OSM addresses, where abbreviations are discouraged,
giving us many examples of fully qualified addresses on which to train.

Non-goals
---------

- Verifying that a location is a valid address
- Street-level geocoding

Raison d'être
-------------

libpostal was originally created as part of the [OpenVenues](https://github.com/openvenues/openvenues) project to solve the problem of venue deduping. In OpenVenues, we have a data set of millions of
places derived from terabytes of web pages from the [Common Crawl](http://commoncrawl.org/).
The Common Crawl is published monthly, and so even merging the results of
two crawls produces significant duplicates.

Deduping is a relatively well-studied field, and for text documents 
like web pages, academic papers, etc. there exist pretty decent approximate
similarity methods such as [MinHash](https://en.wikipedia.org/wiki/MinHash).

However, for physical addresses, the frequent use of conventional abbreviations
such as Road == Rd, California == CA, or New York City == NYC complicates
matters a bit. Even using a technique like MinHash, which is well suited for
approximate matches and is equivalent to the Jaccard similarity of two sets, we
have to work with very short texts and it's often the case that two equivalent
addresses, one abbreviated and one fully specified, will not match very closely
in terms of n-gram set overlap. In non-Latin scripts, say a Russian address and
its transliterated equivalent, it's conceivable that two addresses referring to
the same place may not match even a single character.

As a motivating example, consider the following two equivalent ways to write a
particular Manhattan street address with varying conventions and degrees
of verbosity:

- 30 W 26th St Fl #7
- 30 West Twenty-sixth Street Floor Number 7

Obviously '30 W 26th St Fl #7 != '30 West Twenty-sixth Street Floor Number 7'
in a string comparison sense, but a human can grok that these two addresses
refer to the same physical location.

libpostal aims to create normalized geographic strings, parsed into components,
such that we can more effectively reason about how well two addresses
actually match and make automated server-side decisions about dupes.

So it's not a geocoder?
-----------------------

If the above sounds a lot like geocoding, that's because it is in a way,
only in the OpenVenues case, we have to geocode without a UI or a user 
to select the correct address in an autocomplete dropdown. Given a database 
of source addresses such as OpenAddresses or OpenStreetMap (or all of the above), 
libpostal can be used to implement things like address deduping and server-side
batch geocoding in settings like MapReduce or stream processing.

Why C?
------

libpostal is written in C for three reasons (in order of importance):

1. **Portability/ubiquity**: libpostal targets higher-level languages that
people actually use day-to-day: Python, Go, Ruby, NodeJS, etc. The beauty of C
is that just about any programming language can bind to it and C compilers are
everywhere, so pick your favorite, write a binding, and you can use libpostal
directly in your application without having to stand up a separate server. We
support Mac/Linux (Windows is not a priority but happy to accept patches), have
a standard autotools build and an endianness-agnostic file format for the data
files. The Python bindings, are maintained as part of this repo since they're
needed to construct the training data.

2. **Memory-efficiency**: libpostal is designed to run in a MapReduce setting
where we may be limited to < 1GB of RAM per process depending on the machine
configuration. As much as possible libpostal uses contiguous arrays, tries
(built on contiguous arrays), bloom filters and compressed sparse matrices to
keep memory usage low. It's conceivable that libpostal could even be used on
a mobile device, although that's not an explicit goal of the project.

3. **Performance**: this is last on the list for a reason. Most of the
optimizations in libpostal are for memory usage rather than performance.
libpostal is quite fast given the amount of work it does. It can process
10-30k addresses / second in a single thread/process on the platforms we've
tested (that means processing every address in OSM planet in a little over
an hour). Check out the simple benchmark program to test on your environment
and various types of input. In the MapReduce setting, per-core performance
isn't as important because everything's being done in parallel, but there are
some streaming ingestion applications at Mapzen where this needs to
run in-process.

C codebase
----------

libpostal is written in modern, legible, C99 and uses the following conventions:

- Roughly object-oriented, as much as allowed by C
- Almost no pointer-based data structures, arrays all the way down
- Uses dynamic character arrays (inspired by [sds](https://github.com/antirez/sds)) for safer string handling
- Confines almost all mallocs to *name*_new and all frees to *name*_destroy
- Efficient existing implementations for simple things like hashtables
- Generic containers (via [klib](https://github.com/attractivechaos/klib)) whenever possible
- Data structrues take advantage of sparsity as much as possible
- Efficient double-array trie implementation for most string dictionaries
- Tries to stay cross-platform as much as possible, particularly for *nix

Python codebase
---------------

The [geodata](https://github.com/openvenues/libpostal/tree/master/scripts/geodata) package in the libpostal repo is a confederation of scripts for preprocessing the various geo
data sets and building input files for the C lib to use during model training.
Said scripts shouldn't be needed  for most users unless you're rebuilding data
files for the C lib.

Language dictionaries
---------------------

It's easy to add new languages/synonyms to libpostal by modifying a few text
files. The format of each dictionary file roughly resembles a
Lucene/Elasticsearch synonyms file:

```
drive|dr
street|st|str
road|rd
```

The leftmost string is treated as the canonical/normalized version. Synonyms
if any, are appended to the right, delimited by the pipe character.

The supported languages can be found in the [resources/dictionaries](https://github.com/openvenues/libpostal/tree/master/resources/dictionaries).

Each language can define one or more dictionaries (sometimes called "gazetteers" in NLP) to help with address parsing, and normalizing abbreviations. The dictionary types are:

- **academic_degrees.txt**: for post-nominal strings like "M.D.", "Ph.D.", etc.
- **ambiguous_expansions.txt**: e.g. "E" could be expanded to "East" or could
be "E Street", so if the string it encountered, it can either be left alone or expanded
- **building_types.txt**: strings indicating a building/house
- **company_types.txt**: company suffixes like "Inc" or "GmbH"
- **concatenated_prefixes_separable.txt**: things like "Hinter..." which can
be written either concatenated or as separate tokens
- **concatenated_suffixes_inseparable.txt**: Things like "...bg." => "...burg"
where the suffix cannot be separated from the main token, but either has an
abbreviated equivalent or simply can help identify the token in parsing as,
say, part of a street name
- **concatenated_suffixes_separable.txt**: Things like "...straße" where the
suffix can be either concatenated to the main token or separated
- **directionals.txt**: strings indicating directions (cardinal and
lower/central/upper, etc.)
- **level_types.txt**: strings indicating a particular floor
- **no_number.txt**: strings like "no fixed address"
- **nulls.txt**: strings meaning "not applicable"
- **personal_suffixes.txt**: post-nominal suffixes, usually generational
like Jr/Sr
- **personal_titles.txt**: civilian, royal and military titles
- **place_names.txt**: strings found in names of places e.g. "theatre",
"aquarium", "restaurant". See [Nominatim Special Phrases](http://wiki.openstreetmap.org/wiki/Nominatim/Special_Phrases)
- **post_office.txt**: strings like "p.o. box"
- **qualifiers.txt**: strings like "township"
- **stopwords.txt**: prepositions and articles mostly, very common words
which may be ignored in some contexts
- **street_types.txt**: words like "street", "road", "drive" which indicate
a thoroughfare and their respective abbreviations.
- **synonyms.txt**: any miscellaneous synonyms/abbreviations e.g. "bros"
expands to "brothers", etc. These have no special meaning and will essentially
just be treated as string replacement.
- **toponyms.txt**: abbreviations for certain abbreviations relating to
toponyms like regions, places, etc. Note: GeoNames covers most of these.
In most cases better to leave these alone
- **unit_types.txt**: strings indicating an apartment or unit number

Most of the dictionaries have been derived with the following process:

1. Tokenize every street name in OSM for language x
2. Count the most common N tokens
3. Optionally use frequent item set techniques to extract phrases
4. Run the most frequent words/phrases through Google Translate
5. Add the ones that mean "street" to dictionaries
6. Augment by researching addresses in countries speaking language x

In the future it might be beneficial to move the dictionaries to a wiki
so they can be crowdsourced by native speakers regardless of whether or not
they use git.

Address parser accuracy
-----------------------

On held-out test data (meaning labeled parses that the model has _not_ seen
before), the address parser achieves 98.9% full parse accuracy.

For some tasks like named entity recognition it's preferable to use something
like an F1 score or variants, mostly because there's a class bias problem (most
tokens are non-entities, and a system that simply predicted non-entity for
every token would actually do fairly well in terms of accuracy). That is not
the case for address parsing. Every token has a label and there are millions
of examples of each class in the training data, so accuracy is preferable as it's
a clean, simple and intuitive measure of performance.

Here we use full parse accuracy, meaning we only give the parser a "point" in
the numerator if it gets every single token in the address correct. That should
be a better measure than simply looking at whether each token was correct.

Improving the address parser
----------------------------

Though the current parser is quite good for most standard addresses, there
is still room for improvement, particularly in making sure the training data
we use is as close as possible to addresses in the wild. There are four primary
ways the address parser can be improved even further (in order of difficulty):

1. Contribute addresses to OSM. Anything with an addr:housenumber tag will be
   incorporated automatically into the parser next time it's trained.
2. If the address parser isn't working well for a particular country, language
   or style of address, chances are that some name variations or places being
   missed/mislabeled during training data creation. Sometimes the fix is to
   add more countries at: https://github.com/OpenCageData/address-formatting,
   and in many other cases there are relatively simple tweaks we can make
   when creating the training data that will ensure the model is trained to
   handle your use case without you having to do any manual data entry.
   If you see a pattern of obviously bad address parses, the best thing to
   do is post an issue to Github.
3. We currently don't have training data for things like apartment/flat numbers.
   The tags are fairly uncommon in OSM and the address-formatting templates
   don't use floor, level, apartment/flat number, etc. This would be a slightly
   more involved effort, but would be worth starting a discussion.
4. We use a greedy averaged perceptron for the parser model primarily for its
   speed and relatively good performance compared to slower, fancier models.
   Viterbi inference using a linear-chain CRF may improve parser performance
   on certain classes of input since the score is the argmax over the entire
   label sequence not just the token. This may slow down training significantly
   although runtime performance would be relatively unaffected.

Todos
-----

- [ ] Hosted documentation