0

What is "lost" or what kind of measurable impact has it, when restoring a previously heavily trained Watson Conversation Workspace from its JSON dump-file?

As it seems to me on a small example workspace, the bot is running again. Most probably not as good as before.

Considering a much larger workspace in the future: Is there a way

  1. to quantify and/or measure such lost quality?
  2. to "retrain" the bot (restore the original bot quality after restoring a trained workspace from its dump)? And when yes: what would the best solution be to do that?
René Baron
  • 115
  • 1
  • 5

2 Answers2

1

When you restore or retain a larger Watson workspace ( if it contains a huge number of intents or entities ) it will create as it is, unless your account plan or credentials is expired

Manjunath PV
  • 326
  • 1
  • 2
  • 13
1

The workspace JSON file contains the entire definition of the dialog. If you trained the dialog using the improve feature, then that training is also part of the exported / imported workspace. The structure contains elements for examples and counterexamples which are captured during the training and improve process.

Bottom line: When you export a trained workspace, importing the workspace nothing is lost.

data_henrik
  • 16,724
  • 2
  • 28
  • 49
  • Hi Henrik, Is the following conclusion correct? "Watson Conversation training/improving is nothing else than adding further unlabelled user utterance either to an intent's examples or to counterexamples. So WCS training is just bloating the workspace but does not train the Watson Conversation Service System, right? So the next question would be then: Why is the WCD System considered to be a KI system, when it is not able to learn and finally depends on simple, no-brains pattern matching? – René Baron Feb 13 '18 at 10:00
  • 1
    I am not a developer of WCS, but based on the input data (workspace definition) a learning model is built. See the note here on the models and that they change over time: https://console.bluemix.net/docs/services/conversation/release-notes.html#release-notes IMHO the more the system is used, the better it is to read beween the lines. That is the "fluffy" part built on top of the JSON data. – data_henrik Feb 13 '18 at 11:12
  • Reading the details on Watson Conversation's version notes, currently seems to me, to be the only official piece of information, that might give some minor details on how Watson Conversation might work internally. I really do not understand why IBM is not more specific on this. – René Baron Feb 13 '18 at 12:08
  • I mean, WCS is great when being used out of the box for small mockups and demo project. For larger, enterprise level projects, the question about how intelligent WCS is and will be in the future is of essence for deciding whether to invest into WCS or not. So I do not understand why IBM does not provide more information on this, to make sure, all potential customers can trust and predict the WCS-value (or do the "little" WSC part on their own or with a competitor). – René Baron Feb 13 '18 at 12:08
  • 1
    I recommend getting into the conversation channel on the public IBM Watson Slack: http://wdc-slack-inviter.mybluemix.net/ There are some experts that could get deeper into it. Also note that there is the Conversation service on IBM Cloud, then a Virtual Agent offering and AFAIK some industry-specific offerings. – data_henrik Feb 13 '18 at 13:20