Emora: An Inquisitive Social Chatbot Who Cares For You

https://developer.amazon.com/alexaprize/challenges/past-challenges/challenge3

Transcript

1. Emora: An Inquisitive Social Chatbot Who Cares For you Emory University Presenters: Sarah E. Finch and James D. Finch
2. The Team 2 Ingyu (Jason) Choi, Ali Ahmadvand, Ruixiang Qi, Jinho Choi (advisor), Xiangjue Dong, Zihan Wang, Sarah Finch, Harshita Sahijwani, Sergey Volokhin, Zihao Wang, James Finch, Chenxi (Sonny) Xu
3. Our Goal • Chatbot that talks about many topics • Engage anyone in long conversations • More than just facts and information • Focus on sharing opinions, experiences, and feelings • Users should feel understood and appreciated 3
4. Information, Opinions, and Experiences 4 Information-Oriented Opinion-Oriented Experience-Oriented answering question sharing facts likes and dislikes reactions thoughts everyday life relationships sharing stories
5. Information, Opinions, and Experiences 5 Information-Oriented Opinion-Oriented Experience-Oriented answering question sharing facts likes and dislikes reactions thoughts everyday life relationships sharing stories
6. Information, Opinions, and Experiences 6 Information-Oriented Opinion-Oriented Experience-Oriented answering question sharing facts likes and dislikes reactions thoughts everyday life relationships sharing stories
7. 7 Empirical Evidence • Opinion oriented dialogue is more engaging • Experience oriented dialogue is even more engaging
8. 8 Technical Design of Emora
9. System Architecture User Table State Table Automatic Speech Recognition Text to Speech Profanity Detector Dialogue Manager Music Pets Sports Life ... Sentiment NER EVI Topic Entity linker Intent Natural Language Processing Pipeline EVI ElasticSearch RDS Informational Resources Human Speech Emora Speech
10. • Topic and Intent Classifier (Ahmadvand et al. 2018) • EVI Question Answering • SpaCy Named Entity Recognition • VADER Sentiment Analysis (Hutto and Gilbert 2014) 10 Natural Language Processing (NLP) Pipeline Sentiment NER EVI Topic Entity linker Intent NLP Pipeline
11. 11 Dialogue Manager (Finch and Choi, 2020)
12. 12 Dialogue Manager Precondition Postcondition #IF($movie) “Was $movie good?” [$movie = #MDB(movie)] $focus=$movie [{hi, hello, hey}] “Hi. So, have you seen any good movies?” Variable Value name “jane” movie “Cinderella” pet_type “cat” Memory Table Update Rule Table
13. 13 Dialogue Design
14. Personalization • Save information about user throughout talk • Talk about known user information • Helps user feel understood despite noisy NLU 14
15. 15 Current Event Awareness Highlight indicates time period in which the Coronavirus was the opening topic High daily user rating observed for April Fools as opening topic
16. Character Creation • Reciprocity is a major aspect of human dialogue • People are curious about Emora • Developed Emora as a character with opinions, experiences, and dreams • Improved conversation variety and piqued people’s interest 16
17. • User psychology matters • Starting conversation with personal topics didn’t work 17 Rapport Building
18. • Emora demonstrates untapped potential of social chatbots • Experience-based conversation tends to be more engaging • Knowing lots of facts is not necessary for engaging conversation • Future work will improve domain-scalability of experience-oriented chat 18 Conclusion
19. Thanks for Watching! 19 Ingyu (Jason) Choi, Ali Ahmadvand, Ruixiang Qi, Jinho Choi (advisor), Xiangjue Dong, Zihan Wang, Sarah Finch, Harshita Sahijwani, Sergey Volokhin, Zihao Wang, James Finch, Chenxi (Sonny) Xu
20. Ali Ahmadvand, Ingyu Jason Choi, Harshita Sahijwani, Justus Schmidt, Mingyang Sun, Sergey Volokhin, Zihao Wang, and Eugene Agichtein. 2018. Emory irisbot: An open-domain conversational bot for personalized information access. Alexa Prize Proceedings. James D. Finch and Jinho D. Choi. 2020. Emora stdm: A versatile framework for innovative dialogue system development. In Proceedings of the Annual Conference of the ACL Special Interest Group on Discourse and Dialogue: System Demonstrations (SIGdial 2020). C. J. Hutto and Eric Gilbert. 2014. VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text. Eighth International Conference on Weblogs and Social Media (ICWSM-14). Ann Arbor, MI, June 2014. Credit: Demo stock video provided by Videezy.com 20 References