Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

README #48

Merged
merged 1 commit into from
Jun 11, 2019
Merged

README #48

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
259 changes: 178 additions & 81 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,111 +1,208 @@
| WARNING: This repo is a work in progress! |
| --- |
# A natural language classification engine for geocoding

An address parsing engine designed for geocoding.
This library contains primitive 'building blocks' which can be composed together to produce a powerful and flexible natural language parser.

Strategic goals:
- Seperate unit, housenumber, road from 'everything else'
- Does not require a corpus of 'real places' to operate
- Do not attempt to classify administrative areas
- Provide multiple solutions in the case of ambiguous parses
- Basic typo correction
- Honour delimiters
- Extensible to handle queries such as 'pizza near new york'
- Record offsets to the original token positions in the input text
- Pluggable classifiers
- Support for partially complete 'autocomplete' tokens
The project was designed and built to work with the [Pelias geocoder](https://github.com/pelias/pelias), so it comes bundled with a parser called `AddressParser` which can be included in other npm project independent of Pelias.

#### CLI
It is also possible to modify the configuration of `AddressParser`, the dictionaries or the semantics. You can also easily create a completely new parser to suit your own domain.

[![NPM](https://nodei.co/npm/pelias-parser.png?downloads=true&stars=true)](https://nodei.co/npm/pelias-parser)

[![Gitter](https://badges.gitter.im/Join%20Chat.svg)](https://gitter.im/pelias/pelias)

## AddressParser Example

```
30 w 26 st nyc 10010

(0.95) ➜ [
{ housenumber: '30' },
{ street: 'w 26 st' },
{ locality: 'nyc' },
{ postcode: '10010' }
]
```

## Application Interfaces

You can access the library via three different interfaces:
- all parts of the codebase are available in `javascript` via `npm`
- on the `command line` via the `node bin/cli.js` script
- through a `web service` via the `node server/http.js` script

> the web service provides an interactive demo at the URL `/parser/parse`

## Quick Start

A quick and easy way to get started with the library is to use the command-line interface:

```
node bin/cli.js West 26th Street, New York, NYC, 10010
```

![cli](./docs/cli.png)

#### Server
---

```bash
$ PORT=6100 npm run server;
```
# Architecture Description

![demo](./docs/demo.png)
![api](./docs/api.png)
Please refer to the CLI screenshot above for a visual reference.

### open browser
## Tokenization

the server should now be running and you should be able to access the http API:
Tokenization is the process of splitting text into individual words.

```bash
http://localhost:6100/
```
The spliting process used by the engine maintains token positions, so it's able to 'remember' where each character was in the original input text.

> Tokenization is coloured `blue` on the command-line.

### Span

The most primitive element is called a `span`, this is essentially just a single string of text with some metadata attached.

try the following paths:
The terms `word`, `phrase` and `section` (explained below) are all just ways of using a `span`.

### Section Boundaries

Some parsers like [libpostal](https://github.com/openvenues/libpostal) ignore characters such as `comma`, `tab`, `newline` and `quote`.

While it's unrealistic to expect commas always being present, it's very useful to record their positions when they are.

These boundary positions help to avoid parsing errors for queries such as `Main St, East Village` being parsed as `Main St East` in `Village`.

Once sections are established there is no 'bleeding' of information between sections, avoiding the issue above.

### Word Splitting

Each section is then split in to individual `words`, by default this simply considers whitespace as a word boundary.

As per the `section`, the original token positions are maintained.

### Phrase Generation

May terms such as 'New York City' span multiple words, these multi-word tokens are called `phrases`.

In order to be able to classify `phrase` terms, permutations of adjacent words are generated.

Phrase generation is performed per-section, so it will not generate a `phrase` which contains words from more than one `section`.

Phrase generation is controlled by a configuration which specifies things like the minimum & maximum amount of words allowed in a `phrase`.

### Token Graph

A graph is used to associate `word`, `phrase` and `section` elements to each other.

The graph is free-form, so it's easy to add a new relationship between terms on the future, as required.

Graph Example:

```javascript
/demo
/parser/parse?text=12 main st
// find the next word in this section
word.findOne('next')

// find all words in this phrase
phrase.findAll('child')
```

### Architecture overview
## Classification

#### 1. start with an input:
```
30 West 26th Street, New York, NYC, 10010
```
Classification is the process of establishing that a `word` or `phrase` represents a 'concept' (such as a street name).

#### 2. split tokens in to logical groups:
```
[
"30 West 26th Street",
"New York",
"NYC",
"10010"
]
```
Classification can be based on:
- Dictionary matching (usually with normalization applied)
- Pattern matching (such as regular expressions)
- Composite matching (such as relative positioning)
- External API calls (such as calling other services)
- Other semantic matching techniques

#### 3. tokenize groups:
```
[
[ "30", "west", "26th", "street" ],
[ "new", "york" ],
[ "nyc" ],
[ "10010" ]
]
```
> Classification is coloured `green` and `red` on the command-line.

#### 4. generate phrase permutations:
```
[
[
"30 west 26th street",
"30 west 26th",
"30 west",
"30",
"west 26th street",
"west 26th",
"west",
"26th street",
"26th"
],
[
"new york",
"new",
"york"
],
[ "nyc" ],
[ "10010" ]
]
### Classifier Types

The library comes with three generic classifiers which can be extended in order to create a new `classifier`:

- WordClassifier
- PhraseClassifier
- SectionClassifier

### Classifiers

The library comes bundled with a range of classifiers out-of-the box.

You can find them in the `/classifier` directory, dictionary-based classifiers usually store their data in the `/resources` directory.

Example of some of the included classifiers:

```javascript
// word classifiers
HouseNumberClassifier
PostcodeClassifier
StreetPrefixClassifier
StreetSuffixClassifier
CompoundStreetClassifier
DirectionalClassifier
OrdinalClassifier
StopWordClassifier

// phrase classifiers
IntersectionClassifier
PersonClassifier
GivenNameClassifier
SurnameClassifier
PersonalSuffixClassifier
PersonalTitleClassifier
ChainClassifier
PlaceClassifier
WhosOnFirstClassifier
```

#### 5. run classifiers against all phrases and record potential classes per phrase
## Solvers

Solving is the final process, where `solutions` are generated based on all the classifications that have been made.

Each parse can contain multiple `solutions`, each is provided with a `confidence` score and is displayed sorted from highest scoring solution to lowest scoring.

The core of this process is the `ExclusiveCartesianSolver` module.

This `solver` generates all the possible permutations of the different classifications while taking care to:
- ensure the same `span` position is not used more than once
- ensure that the same `classification` is not used more than once.

After the `ExclusiveCartesianSolver` has run there are additional solvers which can:
- filter the `solutions` to remove inconsistencies
- add new `solutions` to provide additional functionality (such as intersections)

### Solution Masks

It is possible to produce a simple `mask` for any generated solution, this is useful for comparing the `solution` to the original text:

```javascript
VVV VVVV NN SSSSSSS AAAAAA PPPPP
Foo Cafe 10 Main St London 10010 Earth
```
'10010' -> postcode
'west 26th street' -> street
'26th street' -> street
'street' -> street_postfix

# Contributing

Please fork and pull request against upstream master on a feature branch. Pretty please; provide unit tests.

## Unit tests

You can run the unit test suite using the command:

```bash
$ npm test
```

#### 6. generate solutions

Given the classifications for each phrase, compute an array of potential parses for the input, a confidence score can also be provided.
### Continuous Integration

Travis tests every release against all supported Node.js versions.

[![Build Status](https://travis-ci.org/pelias/parser.png?branch=master)](https://travis-ci.org/pelias/parser)


### Versioning

We rely on semantic-release and Greenkeeper to maintain our module and dependency versions.

[![Greenkeeper badge](https://badges.greenkeeper.io/pelias/parser.svg)](https://greenkeeper.io/)