Duration 20:04
16+
Play
Video

End-to-end dialogue systems, or a feature which wasn’t meant to happen | Rasa Summit

Evgeniia Razumovskaia
Machine Learning Research Intern at Rasa
  • Video
  • Table of contents
  • Video
Rasa Summit 2021
February 12, 2021, Online, USA
Rasa Summit 2021
Request Q&A
Video
End-to-end dialogue systems, or a feature which wasn’t meant to happen | Rasa Summit
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Add to favorites
259
I like 0
I dislike 0
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
  • Description
  • Transcript
  • Discussion

About speaker

Evgeniia Razumovskaia
Machine Learning Research Intern at Rasa

About the talk

You know the feeling when you ask for something and you’re pretty sure “no” will be the answer, but you still do it, because why not try? Well… the story of end-to-end is exactly this! Before starting on it, we read several papers about the technology not being ready for end-to-end dialogues in production. So, when we started working on it as a research project, “negative results are also interesting results” was our mantra. Suddenly, the results started to look more and more promising. Then, we developed the end-to-end training further – so that one can combine the classic Rasa format with intents and actions with the new end-to-end and gradually get rid of intents they don’t need.

In short, I will tell you a story of how end-to-end grew from a little internship project into an experimental feature of Rasa (and spanned far beyond the internship).

Presented by Evgeniia Razumovskaia, PhD on Computation, Cognition and Language at University of Cambridge at the 2021 Rasa Summit https://rasa.com/summit/

- Learn more about Rasa: [https://rasa.com​](https://www.youtube.com/redirect?even...​)

- Rasa documentation: [http://rasa.com/docs​](https://www.youtube.com/redirect?even...​)

- Join the Rasa Community: [https://forum.rasa.com​](https://www.youtube.com/redirect?even...​)

- Twitter: [https://twitter.com/Rasa_HQ​](https://www.youtube.com/redirect?even...​)

- Facebook: [https://www.facebook.com/RasaHQ​](https://www.youtube.com/redirect?even...​)

- Linkedin: [https://www.linkedin.com/company/rasa​](https://www.youtube.com/redirect?even...​)

#conversationalai #aichatbot #nlp

Share

He said they'll be talking about Anton, which was my essentially main project during my 10 months is internship at tricep and my hope is its truth is presentation to show you which one one is interested in it as a new teacher arrested, but it was too. I think it's a good showcase example of how we treat imperative things. Go to her house is from a small project at the corner of a table to a full-on experimental feature. So what is, what is Anton, what we are actually started,

and why we store and then how did she did? And what happened after? And I have to say this, whenever I say we have this presentation, it means more roster team and I was and the price of registering for most of the journey, but no recipe irrespective of with me or without me on board. He's just finishing work, still. I think it's good to put it in the context of traditional conversation, which probably, everyone who is here has had some experience with and the tradition pipelines. Just like this, where are you working? For

is user text and then it goes first for your first Mojo, which is Nation language understand. But it's understand what was the user intense and what's your second Mojo, which comes next in the pipeline is done with manager, which predicts will the next action, the boss. And then your third Moto is response to leftovers decides. What is the surface form? Which is sometimes text sometimes images or like no button which this action to take, but when we talked me into it, what which ends do we mean is actually connecting the two ends of this pipeline going directly

from user text on the left. Watch response on the right. And when we look at it from the point of view was, it means that we are essentially getting rid of the three motors that are rented and just using one Modo to go directly strong, the user text to the response of the board. And I know it might sound very tempting that instead of training three, Mojo's you just use one. I don't think I'm in any position to say, just go through surgery. 14 years of research and practice and just change all of us. So I'll tell you what, I think, are advantages and

disadvantages of being both in a pipeline and using Unturned. one advantage which I've mentioned is that you have one Mojo instead of 3 which means that sometimes you might have less parameters this way and this This might have this might help sometime another benefit which I think is one of the most important is that you need much less manual labor. You don't need to annotate everything, go to use a reference, I don't need to find, where would the sports are? In which words, you don't need to come up with the extra labels and sometimes this manual.

Labor is a change of Prestige prestigious. So you might want to avoid it. And stirred mean benefit. It says your actual training is much more legible. something like this with me, always reminds do dialects in Peru in books or where, you know, what is character said and you know exactly which text they meant rather than choosing some of those more attracted representation such an intense or exiting But surely this, if I'm to do this, don't come without the other side.

Yes, on one hand, this case is much more legible but it was. So you have a good and turn motor which works. We need loads and loads of dialog and judges required to make sure that our motive is actually capable of generalizing to something. It hasn't seen that training, Jacob, inferring, all those combinations, grows exponentially. The more options you seen essentially. On the other hand because the murder was done to end and it isn't my thing directly. Use your advance to the next door to spoil. It means if we have much less control and we can't

really explain why the most fun thing over the other in different contexts. And then he was going to think about being a real-life applications where in some point it's the same phrase which is totally fine. In one place, will be insulting someone in another So when we started working on enter, we had all the disadvantages and advantages, it's fun. and there was gold boots, which said that maybe we are not yet ready for unsupervised learning and then to him, but at the same time, more

bit more than a year ago, I wrote a blog post about how rich was time to start getting rid of intense within Chadwick and I think it is important to say it, so I want to never forget to that. Bloke goes for all the reasons why sometimes instead of Zara Boston Neck. And why haven't, you know, that was hard to find. Classes would be would be negative and kind of constraining, but I forgot to tell you that. I think right now is the time because now it's a boy. It's I started to be used for in the world

where they stayed to use it. And then it means that the conversation AI developers face more and more of the situations where would the user said doesn't quite fit in any of the year in terms. That you essentially, the two of the signs are that you've mentioned or Assigning Ron Hughes Road turns, 1 and 2 and would ruin. Another one to give me an example of this is to do exactly the same other user is searching for a restaurant and would suggest start quizzing. And then the user says on the left, the other says I have this yesterday

and book you and I understand that is like human nature Lounge. Language understanding out. Here comes the understanding that I have this yesterday, probably means, no. And then probably the board should suggest a different But example of the right, the user says I had that yesterday, but I'm always up for time which means that the boys can proceed and can offer a place to go to. However, if we'd imagined, if we had to assign the sentences too hard to find until we

realized that if we sign, I had the chance today to deny ancient than most probably when the Moodle encounters I had the chance to do but I'm always up for time is the classified as 8910 over which isn't And the other way around. So, just heart and there are a lot of disadvantages. Well, the challenge of real and the challenge was accepted. I think you're it's important to give you an overview of what, our Research Services included. Still did a bit of research

in a very rule fashion and then if the research would be thinking about, integrating it into the whole recipe framework, and then, once that is done, we would happily welcome to use their in a sentence as an experimental teacher. And that's the way Vermont. As I said, I was the first step was researching a 10:10 and who idea? And we started with this very kind of in a way paper or an academy Academia like where Jesus said cosmology was widely used in dialogue systems community. And would you rather

And their results record looks surprisingly good. I would say, I'm not giving you the number here, because if I gave you the number and then there were two, Explain, but if you're interested, I'm very happy to explain all of the actual experiments. Didn't we started to think about as well? If we have that, then we have to actually consider what could be the way of using Edge within us and we started going through the steps through which our users would go through. So, we took at 4,

which I think you might have encountered, if you've looked at Wrath of victimization and started trying to get rid of intense and essentially retraining it and check in with its as a regular to come for discrimination work and what was that, it performed North 334th and then some case is actually better then Pregnant with strange in whatever fashion. That was pretty inspiring and I'm not going to lie. I think about if I was happier, if this stage went a bit for it because then we wouldn't need to think about how do we actually bring it into reality, better than we expected?

What does it mean to be integrated into the framework? And as I said, I don't think that it's already the time to just throw out everything we had before and do this radical change or let's just all go and fans. And I think it was so we started thinking about how can we integrate what we had before but with the intended and what what is it that we actually went to achieve but having it I'll explain what you're going to do cheap with. With three examples of three

training stories that we have. One is a girl from left to right. It was a left. One is your usual vices story. If you've ever really wanted always goes from engine to action and a typo with you, too old for my butt. Is it is just at normal regular story. To Luke by Cohen. And then the one in the middle is a sandwich. Like A mix of having some texts, but also summoned, Jensen actions. And as you see here, most of it is exactly like the story. I'm glad. But there is just one, he's afraid insurgent. And then the last one use the

movie and times for much, where everything is just like the Just like you know what every person said. So we imagine that's enough. Perfect future world to have some way of having all of those included in Casa. And I'll show you how we actually integrated on an example of a mixed salad. Security me where previously used their invoice, which is each other, and there wasn't an treats and intense search restaurants recognized by the end of the year and our dog with bullets. He predicted the action after being and after suggesting what do we know

from seeing this sorted out already? We know that we have had at least some texturing and this in turn means that there was an interview motive which was strange at the same time. We know that we have been 10 stories so the Darla quality would be able to take text entire column during production. How does the prediction proceed in this case? So first we are in a few Mortals will predicting intent for the user message. Last day, for this case is predict the intent search restaurant.

Are adults supposed to take into account but unlike previously it's not only expert restaurant but also at the text I'm always up for sushi. And at this point. What was he is empowered to make a decision of that they can't and doesn't have any useful information and it can predict the next action directly from the text. I'm always up for sushi. What do we what do we get with having this Moto or this person? Break things that we catch is that we make intense optional, you don't have to sign every single user, a trance that your moto encounters 270 of predefined. And at the same

time we have a very powerful dad with policy which can be trained on intern actual name, text appraisers and enemies of those. I I do it. I don't think it will be an overstatement to say that I have not seen a policy like that before. And I haven't seen an algorithm like this before or after I have to say that before between coming up with that and it being ready, there were a couple of loose ends of the Spiral of. This is a beautiful idea but it's technically feasible and there is technically for but doesn't work the way

we live. But at the end of the Spiral where we are now, I think everyone interested free to use, happy to welcome to users. To use an end-to-end. An experimental feature within Vasa, starting from rest at some point too. And I hope that you give us I'm a bit confused. When insane. I think everyone is 4 + 10 And that's the end of my dog. Thank you very much for your attention, and I am very happy to get any questions, but if there will be anything, I want to answer, you can

find me anywhere.

Cackle comments for the website

Buy this talk

Access to the talk “End-to-end dialogue systems, or a feature which wasn’t meant to happen | Rasa Summit”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Ticket

Get access to all videos “Rasa Summit 2021”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Ticket

Interested in topic “Artificial Intelligence and Machine Learning”?

You might be interested in videos from this event

October 7 - 20, 2020
Online, Mountain View
19
5.41 K
google, googledevs, it, machinelearning, mlsummit, network, platform, tensorflow, tfx

Similar talks

Tom Bocklisch
Director of Engineering at Rasa
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Nikhil Mane
Conversational AI Engineer at Autodesk
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free
Hans Van Dam
Co-founder and CEO at Conversation Design Institute
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Buy this video

Video

Access to the talk “End-to-end dialogue systems, or a feature which wasn’t meant to happen | Rasa Summit”
Available
In cart
Free
Free
Free
Free
Free
Free
Free
Free

Conference Cast

With ConferenceCast.tv, you get access to our library of the world's best conference talks.

Conference Cast
735 conferences
30224 speakers
11293 hours of content