Version 5.1 - October 2024
Natural Language Generation (NLG) is a field of artificial intelligence that focuses on the development of systems that produce text for different applications, for example the textual description of massive datasets or the automation of routine text creation.
The web is constantly growing and its content, getting progressively more dynamic, is well-suited to automation by a realizer. However existing realizers are not designed with the web in mind and their integration in a web environment requires much knowledge, complicating their use.
jsRealB is a text realizer designed specifically for the web, easy to learn and to use. This realizer allows its user to build a variety of French and English expressions and sentences, to add HTML tags to them and to easily integrate them into web pages.
jsRealB can also be used in Javascript application by means of a node.js
module available also as npm
package. It also accepts an input specification in JSON.
The documentation can be accessed here. You can switch between English and French in the upper right corner of the page. The specification of the JSON input format is described here.
The companion project pyrealb implements in Python a text realizer using the same notation for syntactic elements as jsRealB
.
jsRealB
can be used out of the box (the GitHub in fact!) in a web page by using jsRealB.js
in the dist
directory.
Caution
node.js
is necessary for the Javascript application examples.- The current build process relies on the availability of webpack.
-
README.md
: Description of the organization of the source files and describes the main methods.
-
data
: lexicographic information that is bundled with thedist/jsRealB.js
lexicon-en.json
: a comprehensive English lexicon (33926 entries) in JSON formatlexicon-fr.json
: a comprehensive French lexicon (52512 entries) in JSON formatLICENSE.txt
: Creative Commons licenserule-en.js
: English conjugation and declension tablesrule-fr.js
: French conjugation and declension tables
-
[
demos
] : see next section -
dist
: pre-built JavaScript files ready for production use, they already include the English and French lexicons and the English and French rule tables-
jsRealB.js
: packages all .js files of thebuild
directory as a module and exports only main functions and constants- For use in a web page :
<script src="/path/to/dist/jsRealB.js"></script>
- For use as a node.js module :
import jsRealB from "/path/to/dist/jsRealB.js"
- For use in a web page :
-
jsRealB-filter.mjs
: example of use of the node.js module to create a Unix filter forjsRealB
-
jsRealB-server.mjs
: example of use of the node.js module to start a web server that realizes sentences -
testServer.py
: Python script using thejsRealB
server -
package.json
: necessary for publishing thejsrealb
npm package. -
README.md
: short presentation and example of use of the npm package displayed athttps://www.npmjs.com/package/jsrealb
Information for the maintainer: When a new version is to be put on
npm
, in principle, it should be enough to issue the two following commands from within thedist
directory (after a npm login):npm version {major|minor|patch}
ideally the resulting version number should the same asjsRealB_version
injsRealB.js
npm publish
Because of the.npmignore
hidden file in this directory, onlyjsRealB.js
is published.
-
-
documentation
: in both English and French. The examples are generated on the fly by embeddingjsRealB
in the page. Consult the documentationjsRealB-jsonInput.hmtl
use of the JSON format for jsRealB:jsRealBfromPython.html
: documentation for creating the JSON input format in PythonlexiconFormat.html
: format for the entries in the lexiconuser.html
: HTML of the core of the page (div[id]
correspond to variables inuser-infos.js
)style.css
: style sheetuser-infos.js
: definitions of variables containing the examplesuser.js
: JavaScript helper script.
-
Examples
: Examples of integration of jsRealB into web pages or node.js applications. See index.html for use cases. -
IDE
: An Integrated Development Environment built upon theNode.js
read-eval-print loop that includesjsRealB
to easily get the realization of an expression, to consult the lexicon, the conjugation and declination tables. It is also possible to get a lemmatization: i.e. thejsRealB
expression corresponding to a form. See theREADME.html
file to see how to use it. The use of the Evaluation demo is probably more convenient for developing with a web brovser. -
node-modules
: used for transpiling with webpack -
src
: sources to create the JavaScript library; more details in the document on the architecture of the systemjsdoc
: documentation directory of the source files ofjsRealB.js
. Consult the documentation.
Build this directory by runningjsdoc -d jsdoc *.js
in thesrc
directory. For the moment, ignore warning about unable to parse .../Lexicon.js Unfortunately, the jsdoc does not recognize the dynamic classes used for multiple inheritance, so all language specific classes are described as variables.Constituent.js
: Constituent is the top class for methods shared between Phrases and TerminalsConstituent-en.js
,Constituent-fr.js
: language specificConstituent
classes.Dependent.js
: subclass of Constituent for creating complex phrases using the dependency notationDependent-en.js
,Dependent-fr.js
: language specificDependent
classes.JSON-tools.js
: functions for dealing with the JSON input formatjsRealB.js
: main module that gathers all exported symbols from other classes and exports them in a single list. It also defines other utility functions and constantsLexicon.js
: English and French lexicons with their associated functionsLICENSE.txt
: Apache 2.0 license for the source codeNonTerminal-en.js
,NonTerminal-fr.js
:Language specific classes for functions and constants that are shared betweenDependent.js
andPhrase.js
Number.js
: utility functions for number formattingPhrase.js
: subclass of Constituent for creating complex phrases using the constituent notationPhrase-en.js
,Phrase-fr.js
: language specificPhrase
classes.Terminal.js
: subclass of Constituent for creating a single unit (most often a single word)Terminal-en.js
,Terminal-fr.js
: language specificTerminal
classes.
-
Tests
: unit tests (using QUnit) of jsRealB in both French and English.testAll.html
: load this file in a browser to run all tests.
In Visual Studio Code, the launch configuration takes for granted that a local web server has been launched in thejsRealB
directory (e.g. withhttp-server -c-1
)
-
jsRealB
is also available an annpm
package:use-npm.js
is a simple example of its use (after it is installed on the system)
-
Files in the current directory:
README.md
: this filepackage.json
: file with parameters for buildingjsRealB
usingnpm
using
npm run build-dev
ornpm run build-prod
test-demos.sh
: launch all web demos in Safari and the jsRealB server with the Weather Python demotest-node.js
: import the jsRealB package installed withnpm
and realize a simple English sentencetests-dev.js
:node.js
application that loadsjsRealB.js
from thedist
directory. It also has functions with many examples that were useful during the development.web-dev.html
: load the currentdist/jsRealB.js
webpack module in a web page, thus allowing interactive testing.webpack.config.cjs
: configuration file for building thejsRealB.js
package in thedist
directory.vscode
: hidden directory containing configuration for Visual Studio Code
-
Evaluate a
jsRealB
expression and display its realization in a web page in either English or French. -
Show the use of loops in Javascript to create repetitive texts
- English: 99 bottles of beer. Execute
- French: 1 km à pied. Execute
-
Tests of specific features
-
Sentences modified with time, number and conjugation: Date generation Execute
-
Sentence with sentence modifiers Sentence variants Execute
-
French or English conjugation and declension of a word Conjugation and declension Execute
-
Pronouns: Generate a table (both in English and French) showing the different forms of pronouns
- using the original specification
- using the tonic and clitic options
This table is now part of the documentation
-
-
User interface to create a sentence with options. The system shows the
jsRealB
expression and its realization. It is also possible to ask for a random sentence using words of the lexicon.
-
Generate spelling and grammar exercises from a simple sentence structure in both English and French.
-
Translation game from English to French and French to English. Simple sentences are randomly generated in the source language and generated in the target using the same options of
jsRealB
. The used must build the target sentence by selecting words, the system checks if the translation is correct, if not is displays the differences with the expected sentence.
-
Exercise in Style which creates the structure of the original story of Raymond Queneau in both French and English. Using menus, some elements of the text can be modified and the modifications are highlighted in the web page. Exercises in style Execute
-
L'augmentation : Generate a text in French for asking a pay raise following a flowchart as originally described by George Perec. Using menus, some elements of the text can be modified. The path in the flowchart is displayed in the web page and it is possible to highlight a step in the flowchart with the corresponding text. L'Augmentation Execute
-
Eliza : Use jsRealB to program a version of the classical Eliza doctor script in French. Mainly used to show how to generate questions.
-
Universal Dependencies structure used for generating the original sentence from its annotation:
- in English : Execute
- in French : Execute
- Paper describing the approach SyntaxFest-2021 paper
-
Classical fairy tale reproduction in which hovering over a sentence, shows the underlying
jsRealB
expression
-
jsRealB
for the E2E Challenge : browser for the datasets (training, development and test) used in the End to End Generation Challenge (2017-2018). The page also shows the English and French output produced by a "rule-based" generator usingjsRealB
for a selection of feature values. There is also a short description of the implementation of the realizer. Execute -
Personalized descriptions of restaurants : how jsRealB can be used for varying the linguistic style of the generated text according to a user profile defined as one of the Big Five model of personality. Execute
-
Examples suggested by RosaeNLG :
jsRealB
version of an example used in the RosaeNLG tutorials in English and French. RosaeNLG-demos Run with node.js Execute -
Description (in French) of a list of events and associated informations given as a json file Événements Execute
-
Description of list of steps for the building of a house, given information about tasks, the duration and the precedence relations between them.
The system first computes the critical path to find the start and end times of each task. It then creates a graphic for displaying the PERT diagram and an accompanying text to explain the steps to follow. It is possible to interactively change the start date and to explore the graphic with the mouse which also uses jsRealB to generate the text of the tooltips.
-
Itinerary description in an optimistic Montréal Métro network. The system shows an interactive map of the Montréal Métro station with a new line. When a user clicks two stations, the systems realizes a text describing the itinerary to go from the first station to the second.
The language of the web page and of the realization can be changed interactively by clicking in the top right of the page. Metro Execute
-
Weather bulletin generation in English and French. An example of use of the Python API for jsRealB. Taking weather information in JSON, it generates bulletin in both English and French. This tutorial describes the organization of the system which shows how jsRealB can be used in a real-life situation in conjunction with Python for data manipulation.
- Two Observable notebooks are available for trying
jsRealB
expressions and seeing their realizations.
pyrealb
source code is licensed under Apache-2.0 and the linguistic resources in the ./data
directory is
licensed under CC-BY-SA-4.0
Version 3.0 was a redesign and reimplementation of the previous version while keeping intact the external interface, i.e. same name of functions for building constituents, for option names and for global functions. This means that applications using only the external interface of jsRealB
can be run unchanged. Version 4.0 added the dependency notation. Version 5.0 reorganized the internal class structure to separate common processing from the language specific aspects for English and French.
More info:
- This document describes the transformation steps within the realizer using a few examples. It also gives an overview of the implementation explaining the role of the main classes and methods.
- https://arxiv.org/abs/2311.14808 illustrate how to use pyrealb for bilingual data-to-text applications.
jsRealB was updated, developed and brought to its current version by Guy Lapalme building on the work of:
- Francis Gauthier as part of his summer internship at RALI in 2016;
- Paul Molins as part of an internship from INSA Lyon spent at RALI, University of Montreal in 2015;
- Nicolas Daoust developed the original concept in the JSreal realizer for French only in 2013.
For more information, contact Guy Lapalme.
Thanks to Fabrizio Gotti, François Lareau and Ludan Stoeckle for interesting suggestions.