This document discusses using dialogue to automatically expand an AI assistant's knowledge base. It presents an approach where the assistant engages in a dialogue to clarify statements made by users in order to learn new information. This learned knowledge is then added to the assistant's knowledge base after being validated through techniques like voting across sources, interactive validation with users, logical consistency checks, and knowledge classification. The goal is to facilitate online learning through dialogue and experiments comparing different knowledge collection and filtering methods.
3. Dialogue as Knowledge + Inference
Alex: “Hi how are you?”
Sam: “I can’t stand my brother’s wife.”
4. Dialogue as Knowledge + Inference
Alex: “Hi how are you?”
Sam: “I can’t stand my brother’s wife.”
Alex: ???
5. Dialogue as Knowledge + Inference
Alex: “Hi how are you?”
Sam: “I can’t stand my brother’s wife.”
Alex: ???
> “Why?”
6. Dialogue as Knowledge + Inference
Alex: “Hi how are you?”
Sam: “I can’t stand my brother’s wife.”
Alex: ???
> “Why?”
> “Do you and your brother get along?”
7. Dialogue as Knowledge + Inference
Alex: “Hi how are you?”
Sam: “I can’t stand my brother’s wife.”
Alex: ???
> “Why?”
> “Do you and your brother get along?”
> “How do holidays go?”
8. Dialogue as Knowledge + Inference
Alex: “Hi how are you?”
Sam: “I can’t stand my brother’s wife.”
Alex: ???
> “Why?”
> “Do you and your brother get along?”
> “How do holidays go?”
Commonsense Knowledge:
● Your brother’s wife is
family
● You should get along
with your family
● Balance theory (people
you dislike get along with
others you dislike)
● Families usually get
together on holidays
● Being around people you
dislike is stressful
10. Dialogue Management Approach (Background)
Alex Sam
greeted feels
?
brother
wife
dislikes family
YX
person
type
dislike
type
Z
cause
Alex: “Hi
how are you?”
Sam: “I can’t
stand ...”
11. Dialogue Management Approach (Background)
Alex Sam
greeted feels
?
brother
wife
dislikes family
YX
person
type
dislike
type
Z
cause
Alex: “Hi
how are you?”
Sam: “I can’t
stand ...”
Alex:
“Why?”
?
cause
12. Types of Knowledge
Positive
“The capital of the US is Washington D.C.”
“The election happened yesterday.”
Ontological
“People spend time with their family on holidays.”
“Dogs bark.”
YX
person
type
spend time
type
Z
during
family
holiday
type
US
country
type
D.C.
capital
city
type
14. 1. Interactive Learning
“I went to Miyazaki last year, that was pretty fun.”
“Is Miyazaki a city?”
“Yes.”
“Interesting, I’ve never heard of it. Is it in the US?”
“Japan.”
Related: Chai et al. 2018
15. 2. Offline Crawl Learning
From Wikipedia:
Japan is an island country in East Asia located in the northwest Pacific Ocean.
Japan
location Pacific
Ocean
East
Asia
type
island
made_of
country
part_of
17. 4. Reflexive Learning
“What did you do today?”
“I just went for a run.”
“Are you tired?”
● In theory, the user is performing the
same steps as Emora to respond
● What reasoning did the user do to
make their response?
Related: Otani et al. 2016
19. 1. Voting
● Count how many sources mention a predicate
● More sources mentioning a predicate makes it more reliable
● Mitigates for noise but not systematic errors in interpretation
Related: Otani et al. 2016
20. 2. Interactive Validation
Knowledge Candidate: Dogs annoy people
“I took a walk with my dog today.”
“Oh I’m sorry, that must be annoying.”
“What? No, it was good.”
22. 4. Logical Surprisal Minimization
● Good knowledge will be logically accordant with established knowledge
● Interpret static data with and without knowledge candidate
● Run inference step
● Evaluate contradictions and correct predictions made by inference
Related: Wu et al. 2018
23. Planned Contributions
● Novel dialogue management approach that facilitates online learning
● Use of dialogue as an environment to test candidate knowledge validity
● Experiments between different knowledge collection & filtering approaches
● Automatic expansion of chat-relevant knowledge base
24. References
Bosselut, Antoine, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz, and Yejin Choi. 2019. “COMET: Commonsense
Transformers for Automatic Knowledge Graph Construction.” In Proceedings of the 57th Annual Meeting of the Association for Computational
Linguistics, 4762–4779. Florence, Italy: Association for Computational Linguistics. https://doi.org/10.18653/v1/P19-1470.
Chai, J. Y., Qiaozi Gao, Lanbo She, Shaohua Yang, Sari Saba-Sadiya, and Guangyue Xu. 2018. “Language to Action: Towards Interactive Task
Learning with Physical Agents.” In IJCAI. https://doi.org/10.24963/ijcai.2018/1.
Nguyen, Dai Quoc, Thanh Vu, Tu Dinh Nguyen, Dat Quoc Nguyen, and Dinh Phung. 2019. “A Capsule Network-Based Embedding Model for
Knowledge Graph Completion and Search Personalization.” In Proceedings of the 2019 Conference of the North American Chapter of the
Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 2180–2189. Minneapolis,
Minnesota: Association for Computational Linguistics. https://doi.org/10.18653/v1/N19-1226.
Otani, Naoki, Daisuke Kawahara, Sadao Kurohashi, Nobuhiro Kaji, and Manabu Sassano. 2016. “Large-Scale Acquisition of Commonsense
Knowledge via a Quiz Game on a Dialogue System.” In Proceedings of the Open Knowledge Base and Question Answering Workshop (OKBQA
2016), 11–20. Osaka, Japan: The COLING 2016 Organizing Committee. https://www.aclweb.org/anthology/W16-4402.
Wu, Benjamin, Alessandra Russo, Mark Law, and Katsumi Inoue. 2018. “Learning Commonsense Knowledge through Interactive Dialogue.”
EasyChair Preprints. EasyChair Preprints. EasyChair. https://doi.org/10.29007/lrph.