Add a new language
Adding a new language in RosaeNLG consists of 2 main tasks:
-
All (well, most of) language specific content must be in dedicated packages: Create Linguistic Resources. Typically you will define how to agree words, verbs, and adjectives. Create your linguistic resources, one by one. You can start with implementing these modules as they are small and require very few "RosaeNLG specific" knowledge.
-
Implement the high level language specific behaviors and bundle everything together: Implement the Language in RosaeNLG
Create Linguistic Resources
Overview
The target is to put 99% of the language specific content in specific packages, in order to:
-
favor reuse beyond RosaeNLG
-
keep language specific packages for the browser small enough
Language specific packages manage agreement of words, verbs and adjectives. For language with a gender they also manage gender of words. At last, you also have dedicated packages for determiners and ordinal numbers.
Example
In French:
-
Words / Adjectives / Verbs:
-
words gender are managed using lists, by
french-words
andfrench-words-gender-lefff
, while word agreement is managed using rules, bypluralize-fr
-
adjectives agreement are managed using rules, by
french-adjectives
andfrench-adjectives-wrapper
-
verbs agreement are managed using lists, by
french-verbs
andfrench-verbs-lefff
, and intransitive verbs are infrench-verbs-intransitive
-
-
Misc:
-
determiners with
french-determiners
-
contractions by
french-contractions
-
ordinal numbers by
french-ordinals
-
helper for simplified syntax (used by Simplified Syntax) is in
lefff-helper
-
In English, the a/an mechanism, specific to English, is managed by english-a-an
(wrapper) and english-a-an-list
(the real list).
Rules or Lists?
Agreements can be managed using rules (as we learned at school) or using explicit lists derived from linguistic resources. Rules can be tedious to implement (there are always plenty of exceptions), while lists are heavy, bring both errors and missing elements, and carry their own license.
Both are ok:
-
French is mostly implemented using lists, except for adjectives agreement
-
German uses only lists
-
Spanish uses only rules
You have to check what exists on npm and elsewhere, and check license compatibility with Apache 2.0.
Wrappers and Lists
Each time a package is important enough, you will have to create 2:
-
the real package that contains the rules or the just list (e.g.
french-verbs-lefff
); it can be as heavy as required -
a wrapper that uses the real package
The real package will only be used when compiling templates (you need the whole lists when compiling), while the wrapper will only be used when rendering compiled templates (you only need the elements which are in your templates, not the whole lists). This keeps rendering only browser packages small.
The wrapper will used the JSON resource file for testing purposes, but only call it in devDependencies
(do not to link them together).
If your module uses lists, you must declare it in fake_resources.js , or build times of other packages will take ages as tsc will try to parse them (and fail).
|
Implement the Language in RosaeNLG
As a general advice:
-
Make something very simple, without language specific behaviors, just to plug everything together and "make things work" (compile, simple tests etc.)
-
Build progressively all language specific behaviors
In rosaenlg-filter
package:
-
create a new class that extends
LanguageFilter
-
it will contain everything specific about filtering for this language: contractions, punctuations
-
create something very simple at the beginning; it can become much more complex (like in French or Italian) or remain very simple (like in German)
-
update helpers
index.ts
languageFilterHelper.ts
etc. -
add test cases in
filter.js
In rosaenlg-commons
package:
-
create a new class that extends
LanguageCommon
-
it must remain quite simple as a target; you can keep it almost empty at the beginning, but at some point update
validPropsWord
andvalidPropsAdj
-
update helpers helpers
helper.ts
index.ts
etc. -
add test cases
In rosaenlg-pug-code-gen
package:
-
create a new class that extends
LanguageCodeGen
-
its only goal is to fetch structured Verb, Word and Adjective data from dedicated modules; it should not contain any language related logic
-
update
package.json
accordingly -
update helper
helper.ts
-
add test cases
In rosaenlg
package:
-
create a new class that extends
LanguageImpl
-
indicate properties, and also override methods
-
this class can become complex, as it will hold all language specific complex behaviors; but at the beginning you can keep it simple
-
use
build:tsc
to build (do not make a complete build, with browser packages, each time) -
update
package.json
accordingly -
create test cases along your development in
test/test-rosaenlg/xx_XX
Browser packaging, also in rosaenlg
package:
-
declare all language specific packages properly in
rollup.config.js
: the linguistic resource must not be included in the standard package, but only in the package to compile templates (the big one) -
enrich
tiny.js
-
build browser packages: adapt
yarn run rollup_xx
inpackage.json
-
check manually the content of each bundled package
-
mocha on
tiny.js
should work