Generative models for chatbots

Presentation was given during Data Natives 2017 in Berlin.

Transcript

1. Dr. Karol Przystalski Generative models for chatbots
2. Current chatbots challenges Context Management Sentiment Analysis Pattern Recognition Natural Language Understanding Designing Bots, Amir Shevat, O’Reilly 2017
3. Current approches Rule-based Retrieval-based Generative-based
4. Rule-based Phrases list Show status of recruitment. What is the weather in Berlin? Hi! Hire candidate <name>. No valid phrase found Answers list We have currently X candidates. It is <current weather>. Hi. How are you? Sent an email to <email>.
5. Rule-based Bottlenecks/Challanges ⟶ too simple for most cases ⟶ not really intelligent Advantages ✓ is predictable ✓ clear principles ✓ cheap
6. Retrieval-based
7. Retrieval-based Advantages ✓ identify the intent ✓ usually easy to train ✓ does not need too many questions/answers ✓ more intelligent than rule-based Bottlenecks/Challanges ⟶ limited to questions/answers
8. Generative Models
9. Generative Models
10. Generative Models
11. Generative-based Advantages ✓ generic, intelligent answers ✓ raw data Bottlenecks/Challanges ⟶ usually takes longer to train ⟶ need a data set that is usually huge ⟶ sometimes unpredictable
12. Comparison ⟶ don’t use a solution/approach because it’s cool! ⟶ in many cases classic machine learning methods are good enough and generative models does not need to be used ⟶ generative models are good if we are looking for a more generic solution and have a lot of data to be used for training
13. Where to go next ... To learn more about bots and how build one, please register for Developing bots primers workshop: workshops.codete.com. It’s free. Read Pattern recognition Primer, Springer 2018 book. It will be published in Spring 2018. Looking for ideas or advices? Find me at our booth at Data Natives conference.
14. Q&A