Training a goal oriented bot based on the gobot_extended_tutorial

axelbender · November 3, 2020, 10:14am

Hi, new to the forum and using DeepPavlov in general. I followed the tutorial on Github: https://github.com/deepmipt/DeepPavlov/blob/master/examples/gobot_extended_tutorial.ipynb

I’ve created my own data, structured in the same way as the dstc2 dataset, with the exception of the api calls to a db.sqlite database. I’ve removed the db entry from the gobot_config and created a slotfill that works for the dataset I’ve created.

My problem is that when I train this modified bot it fails to navigate a conversation similar to the select restaurant scenario, every input results in the bot outputting the first line of the training data (equivalent to: “Hello, welcome to the Cambridge restaurant system. You can ask for restaurants by area, price range or food type. How may I help you?”) Are there any obvious errors that I’ve made or steps I haven’t considered, when modifying the tutorial to work with new data?

danielkornev · November 5, 2020, 5:58pm

Hi!

Thanks for trying out our Go-Bot. In the last few months we’ve started making changes in our Go-Bot that are focused on supporting RASA DSLs (currently v1, soon v2): stories.md, nlu.md, and domain.yml to make configuration of key components of a goal-oriented bot simple and easy to use.

Take a look at this tutorial:

github.com

deepmipt/DeepPavlov/blob/fixes/gobot-md-yaml-config-danielko-1/examples/gobot_md_yaml_configs_tutorial.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "gobot_md_yaml_configs_tutorial.ipynb",
      "provenance": [],
      "toc_visible": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "pycharm": {
      "stem_cell": {
        "cell_type": "raw",
        "source": [],
        "metadata": {
          "collapsed": false
        }

This file has been truncated. show original

In addition to that we’ve started transitioning Go-Bot into a framework for building goal-oriented Skills for our DeepPavlov Dream which enables you building Multiskill AI Assistants.

We’ve just showcased an example of such goal-oriented skill built using Go-Bot based on RASA DSLs at ODSC West 2020 and at ML Conf EU. In our demo we’ve got our harvesters maintenance skill from Deepy 3000 Demo (see at https://github.com/deepmipt/dp-dream-demos, check out branch “moonbase_ai_demo”) and re-implemented it with the Go-Bot built one.

We’ll release a public version of the said demo in the coming weeks, and we ask you to stay tuned and subscribe to our blog to learn how to build your goal-oriented bot with our DeepPavlov Dream and Go-Bot.

Thanks,
Daniel Kornev,
CPO @ DeepPavlov.ai