Transformer-based models have pushed state of the art in many areas of NLP, but our understanding of what is behind their success is still limited. This talk is a high-level summary of the "Primer in BERTology", a survey of over 150 studies of the popular BERT model, focusing on the key takeaways and unresolved questions.

Anna Rogers

Jeronim will give a brief history on AXA's journey, coming from a handcrafted chatbot with lots of rules and strict conversation flows, to enable our customer service experts to drive a great user experience.

Jeronim Morina

i2c inc. is a Payment Processing firm that additionally offers its clients Customer Services (Applications and Agents) as well as Automated Call Quality Assurance facilities. This presentation walks through i2c inc's unique use-case of Multi-Speaker Dialog Management using Call Transcriptions with Rasa. Ramsha will talk about i2c inc's overall pipeline for processing a single call, as well as some neat tricks like: Data Augmentation with Alternative Words and Semi-Supervised Learning, Adding Speaker Context in NLU, Frame Detection in Slot Tagging, and Improving Dialog Management with Ensembles. Ramsha will also talk about the problems i2c inc faced with carrying Rasa to Production given our previous ML-Ops pipeline, and how it helped us improve it for the better! My hope is that this presentation will inspire others working in Voice AI to experiment with Rasa in their respective use-cases.

Ramsha Ijaz Siddiqui

I’ve recently embarked upon a new adventure, an experiment - can I prototype an immersive audio game using voice and conversational AI that other people want to play? In this talk I’ll give you an insight into what I’m working on and why, highlighting some of the challenges of developing our first assistant (and central character), Freddy. I’ll take you through my journey so far - from hacking to adopting a framework to establishing a repeatable workflow, and share some of the lessons I’ve learnt along the way.

Ben Ackland

Learn from a Rasa living deployment project - how to incorporate DevOps practices into your virtual assistant. This presentation extends the previous "Deploy your Rasa Chatbots Like a Boss" by including a new workflow showcasing an integration between DevOps Platform + Git for Data using DVC and bringing GPUs into your CI/CD Rasa path to production pipeline.

William Galindez Arias

Arte Merritt

Scott Truitt

Scott Truit

We use Rasa X in combination with Airy Open Source Conversational Infrastructure to build a feedback loop between the model and human agents.
Sometimes users ask questions your model is not perfectly trained for yet, resulting in low confidence levels for the potential responses. We demonstrate how you can suggest potential responses to Human Support Agents via the Airy Inbox UI and have them select the correct response to be sent back to the user. We observe the chosen response and retrain the model with Rasa X to be able to fully automate the same request the next time.

Christoph Pröschel

This talk covers the basic concepts of algorithmic bias and dissects areas where bias can enter an algorithmic system. We will shift the methodology for accountability from a developer level to an organizational level. You’ll learn how individuals and teams within an organization can push for nuanced and meaningful discussions on fairness within production algorithms.
Takeaways
- Understand how Bias enters and impacts an algorithm
- Tips on discussing organizational AI accountability broadly
- What kinds of ethical conversations and actions you can implement

Su Lin Blodgett

Dr. Brandeis Marshall

Ayodele Odubela

Vincent Warmerdam

Design and development are pillars of conversational AI app creation, but it’s damn near impossible for designers and developers to collaborate efficiently. Design thinking must be expansive, whereas developers must consider and navigate the minutiae of technological constraints. In this talk, conversation designer Lauren Golembiewski and developer Matt Buck will explore their processes, tips, and tricks for building delightful conversational AI experiences. They’ll discuss how designers and developers should perform and communicate user research, collaborate on design artifacts, and test and maintain conversational apps.

Lauren Golembiewski

Matt Buck

Best practices in deploying speech technology at the enterprise, and avoiding common pitfalls. The goal of the talk is to give attendees the big picture with enough specifics so they feel comfortable to start exploring speech in their company.

Josh Meyer

In this lighting talk, I will introduce Chinese components created by and used by Rasa Chinese community. Another thing I want to share is that I will introduce an experimental channel connector for Wechat (the most popular instant message software in China, the Chinese version of facebook+twitter).

Xiaoquan Kong

In this talk, we look at whether a new language, explicitly built for conversations, can help make designing and implementing great conversational experiences easier.
We introduce Colang, a new language with a simple and intuitive syntax, making it easy to learn, but with powerful semantics enabling the modeling of complex scenarios. We showcase a first implementation of the language using RASA as the backing conversational AI framework.
Using the Colang playground, we build a conference assistant and look at how we can handle various common scenarios, starting with FAQs, regular intents, and knowledge graph queries. Next, look at how we can support more advanced patterns like contextual questions (i.e., “what about X”), digressions, repair paths, and ambiguity. Last but not least, we look at how we can integrate generative models like GPT-3 to generate contextual responses based on the user’s questions while still keeping control of the conversation flow.

Razvan Dinu

Dialogue generation with large neural models is a hot topic in the research community. Neural dialogue generation models are responsive and flexible; they handle a wider variety of user utterances more naturally, which is important for social conversations. But they are also harder to control and constrain; they often hallucinate and are inconsistent. In this talk we'll look at how Stanford's Alexa Prize team used state-of-the-art neural generation in tandem with rules-based dialogue to get the best of both worlds. Within a year we were able to bootstrap from scratch and score higher than entrenched teams leading to a second place finish. We hope that our learnings about neural generation and more broadly our modular chatbot design for social conversations, now open-sourced, will be instrumental in adding a social and personable component to all kinds of chatbots.

Ashwin Paranjape 

The aerospace industry relies on massive collections of complex and technical documents covering system descriptions, manuals, or procedures. This talk will focus on a question answering (QA) system that was built to help aircraft pilots access information in this documentation by naturally interacting with the system and asking questions in natural language.

Gérard Dupont

Rebecca Evanhoe

Diana Deibel

Hans van Dam

Emily Lonetto

Casey Phillips

Joseph Suskin

Whether you’re creating your first conversational experience or you’re a veteran, broadening our approach for conversation design is a worthwhile business investment (Gartner, 2019). The future of conversation design is diverse and inclusive, and we need to drive rigor around acknowledging our own assumptions about language and culture as we design conversations. This presentation will show how we, at Salesforce, leverage sociolinguistics as a foundation for how we structure our design team, extend our services as a force multiplier across product areas, and craft intentional language experiences that authentically and inclusively represent an organization’s brand and values.

Greg Bennett

Alan Nichol

The chatbot value proposition is hard to argue with - the creation of faster, personalized self-service for customers and employees alike is something that every organization wants to adopt - but, current implementations have been lacking. This is because, simply, chatbots are hard - and it is very, very easy to get them wrong, even if you do everything right technically. In this keynote, Forrester Researcher Will McKeon-White will be going over some chatbot strategy best practices (and how to avoid common pitfalls) including:
<ul>
<li>How you should structure your 'chatbot team' to ensure success.</li>
<li>How to center your strategy around user experience.</li>
<li>How you can optimize your development efforts.</li>
<li>How you can scale your chatbot to new use-cases.</li>
<li>How (and which) metrics can help you determine success.</li>
</ul>

Will McKeon-White

Alex Weidauer

Sweta Patel

Dennis Yang

Vineet Malhotra

Samantha Gardner

Bruna Nayara Moreira Lima

Lindsey McCarthy

Heather Nolis

Akela Drissner

Mady Mantha

Working on natural language processing for real-world applications requires more than just developing model implementations and evaluating them on existing datasets, or memorizing various library APIs. Often, what's needed is an entirely different mindset: How can I break down complex business problems into machine learning components? How do I design my data to make the problem easier and get human experts involved? And how do I incorporate linguistic insights to find approaches that are more likely to succeed? In this talk, I'll share some lessons we've learned from commercial use cases of our software, spaCy and Prodigy, and suggest ways we can teach applied NLP thinking and ship more successful projects.

Ines Montani

The simple guidelines on how to manage your Virtual Machines and Kubernetes clusters locally on Microsoft Azure during Rasa-X deployment. Outlining the ways to secure the Azure Resources (Linux VMs, Azure DNS hosting and Kubernetes), running your shell scripts on Linux server, securing your Rasa X deployment, and storing your credentials.

Jacques Patricks Khisa

Last year at L3AI, I gave a talk entitled “Meta bots: why they’re probably not the solution to your problems”. Since then we’ve seen two big trends in the enterprise:
1. Growth in the complexity of enterprise scale, i.e. more end users, larger teams building the assistants, more user goals for the assistants etc.
2. A huge increase in demand for voice assistants, particularly in the IVR application.
I will talk about the Rasa approach for multi-bot orchestration to handle scale (spoiler alert: it’s not orchestration in the traditional sense), and what to consider when branching into the voice domain.

At PicPay, the chatbot needs to handle on average half a million messages daily. Our initial solution was not efficient enough, so we needed to make fine-tunings in the chatbot architecture and infrastructure to handle the waves of messages and avoid waste of resources. In this presentation, I will show to you each challenge that we had to overcome to achieve a good Chatbot solution.

Arthur Rocha Temporim de Lacerda

Building a virtual assistant is a challenging task, and adding voice support can make it further more complicated. Creating human-like conversations require speech recognition, language understanding and speech synthesis models to run at low latency while delivering high accuracy. Additionally, training and fine tuning models on custom data to increase domain-specific accuracy may need several iterations, and a deep understanding of the models.
In this talk, we will describe how NVIDIA Jarvis' NLP can be used in conjunction with Rasa to build a simple weather domain Virtual Assistant. We enhance it further by adding voice support by leveraging Jarvis' ASR and TTS capabilities. We will then discuss fine tuning the ASR and NLP models on your custom data using NVIDIA Transfer Learning Toolkit (TLT), a zero-coding training toolkit to train, adapt and optimize AI models.

Nikhil Srihari

Shashank Verma

Josh Cutler

The first lesson in chatbots that everyone learns regardless of their role is – customers will say and do things you never imagined, no matter how much planning you put into your features. This creates a natural dilemma. Who wins the top priority spot in the backlog? The newly identified customer pain points and opportunities, or the well-thought, thoroughly planned, curated backlog your stakeholders have their hearts set on?
Through weekly incremental model releases, continual data review, and a UX-focused scrum team, we have managed two parallel streams of work - one responding to our immediate opportunities to help our users, and one responding to the long-view backlog of our stakeholders.
In this talk, I will walk through how the AI @ T-Mobile team utilizes data to identify new use cases, refine a design, productionize it, and iterate on it to improve the customer experience to drive a delightful, functioning chatbot with incredible ROI.

Best practices in conversational AI" - Having seen 90 billion chatbot and voice messages over thousands of conversational interfaces and use cases, Arte will discuss best practices in developing conversational AI including how to get started building NLP models, build vs buy, the role of context and personalization, handling fallbacks, integrating with 3rd party systems, and testing, measurement, and iterating.

Entities are a fundamental concept in the dialog systems, however, they are often overlooked. In this presentation, we will introduce the benefits of using entities in dialog systems and what approaches are being taken - including machine learning models.

Mirelle Bueno

Merve will talk about how to build a conversational assistant depending on your task. Each have their own challenges in NLU but there are choices you can make in the understanding and dialogue management pipeline to overcome them.

Schedule