/nlut

NLU Transcoder — easier import/export from and to Wit.ia, Dialogflow, Alexa skills

Primary LanguagePHPMIT LicenseMIT

NLU Transcoder

For PHP7.1.

This little PHP CLI tool transcodes Alexa Skill / Wit.ai / DialogFlow backup files in another format for ease of switching to another platform if needed.

It does its best to recover the model data and to transcode it as accurately as possible, even though technically this is not possible since 1. each platform uses custom entities for common types (dates, numbers, ...) that are generally not exported in the archive files and 2. the inner workings of each platform makes it hard to match all expressions in the training model.

Use at your own risk. PRs welcome ! :)

Installation

No steps are necessary prior to using the tool.

Usage

./nlut.php --help

Get info on an archive file

./nlut.php --source MyWitProject.zip
----------------------------------------
This is a Wit.ai archive
App name     : MyWitProject
Entities     : 17
Intents      : 5
Expressions  : 3624
----------------------------------------

Transcode a DialogFlow archive into the Wit format

./nlut.php --source MyDFArchive.zip --export MyProject --format WIT
----------------------------------------
This is a DialogFlow archive
App name     : project-13d59
Entities     : 17
Intents      : 5
Expressions  : 3624
----------------------------------------
Exporting to the Wit format
File MyProject.zip written.

Do not put an extension on the exported file name, it will be put automatically

Tests

As of 23/07/2018 :

to Wit to DialogFlow to Alexa
from Wit ✅ ✅ ✅
from DialogFlow ✅ ✅ ✅
from Alexa ✅ ✅ ✅

Known limitations / bugs

  • Platform-specific entities (such as wit/datetime or DialogFlow's @sys.duration) are treated as user-defined entities (free-text for Wit, for instance). In the case of Alexa specific entities, such as AMAZON.DATE, placeholders are put in the expressions because the exported model does not contain utterances for these entities
  • Alexa's AMAZON.FallbackIntent, AMAZON.CancelIntent, AMAZON.HelpIntent and AMAZON.StopIntent are ignored when exporting
  • When exporting from an Alexa skill, expressions (or samples) are reconstructed from the first synonym of each slot, thus leading to sentences that can be meaningless in natural language. They are accurate training-wise, though
  • Not tested on high-complexity models

License

MIT