Hello, I have a big txt file of 2-person real conversation and I want to create a training data for the core (dialogue management) by transforming it into a stories format and feed it to rasa X ( I want to avoid using interactive learning since I have a big number of conversation scenarios and I can’t do it manually). I found something called “migration script” migrates any persistent tracker stores to a Rasa X compatible database. but I couldn’t find what the data format should I prepare as an input to that script have finally a Rasa X compatible database any help plzzzzzz
How to transform a real conversion (question & response) to stories format and fill in automatically in Rasa X
Hi @Hamza, the migration script is for migrating conversation trackers from actual conversations with an assistant; it won’t work with a text conversation like you have. Rasa requires stories to train a dialogue model; if you know the names of the intents and actions for each turn in the conversation, you could write a script to convert that to story format, but from what you’re describing it doesn’t sound like that’s the case? Do you have any stories, nlu data, or a domain yet?
Thank you @mloubser for your answer,
the migration script is for migrating conversation trackers from actual conversations with an assistant
Any assistant? like, google assistant? Salesforce bot?
- I’m actually using Rasa X to annotate the intents for only one person text conversation ( user questions) and I don’t have any idea how to annotate actions in Rasa X which is in my case (human response in my text conversations). Let’s imagine I create actions for each intent and I have finally intents and actions, how can I write a script to convert that to story format? since I need to make an intent classification in every turn in my real text conversation to identify the intent and after I can associate an action to it . this is the way I can create story format from a real text conversation. Do you think Rasa has this flexibility to allow me to create that script?
- Yeah we can create stories, nlu and domain for a small bot with few examples, but how can leverage our real txt conversation to make the bot better? ( I don’t want to use interactive learning, it’s very time expansive since we have a big data)
The migration script is for migrating conversations with a Rasa assistant; it’s not relevant to what you’re trying to do here.
By the time you have annotated all your conversations with intents & actions, you have essentially written your stories; you would not need to use interactive learning for that, but it would still require manual work.
Here’s what you could use your data for:
- NLU training data (sounds like you’re already doing this)
- In the case of simple questions/chit chat conversations, training a Response selector and using a retrieval action.
The short answer is, you can’t automatically turn text conversations into a Rasa assistant without some intermediate work. You can use the data you have to inform how you design your assistant, and you can use your data as training data for various components, but you cannot directly convert a set of text conversations into rasa stories.
This blog post: Integrate response retrieval models in assistants built with Rasa might be helpful, see the end section on progress towards end-to-end training (which it sounds like is what you’re asking for) for context.
Thank you so much @mloubser for your answers, that was very helpful