Peregrin Semantic notes

Abstract: Dialogue interaction provides an extremely challenging and grounded evalu- ation of natural language understanding. Natural language understanding has greatly improved over the last 20 years, but what level of understanding is needed to respond appropriately in dialogue in a particular context? How can natural language genera- tion produce these contextually appropriate responses? How can it scale? Should a conversational assistant have a persona or style? Should it entrain to the user, take the initiative, understand greetings and other phatic acts, and use an explicit model of dialogue structure? What are some near term and some long term solutions towards more natural dialogue interaction?

problems scaling - How may I help you?

• problem

• 2004 moving from system initiative and larger vocabulary SDS
• 2004 future - embodied, personalized

• ELVIS dialogue system for retrieving email

• interesting - dialogues for something that people wants - e.g. for chocolate - but online e.g. game

• Can people tell if SDS is customized for them?
• icids: Generate Dialogue Variants from Monologue
• tunable by reinforcent learning
  • speed, prosody, volume
  • syntax, word choice
  • what to say
• ALI - hezky vysvetleni - jako kdyz se dela karikatura
  • odlisit avg norm - and example
• if you have large collection of speaking styles, you can find the one of the original speaker
  • idea
• people - cuteness -> willing to teach it? social helps to learn facts new facts

Ramesh - dai interpretation Yanchao Yu - learn knowledge from scratch

Chelsey Jurado

Tony

• dialport
• free-data

Bing Liu

• joint SLU and LM
• How to connect ASR (phonetic) and SLU and LM

One need to be interesting in the agent - must have personality How to measure drop-of?

How to detect that user is quitting because it?

Personality is important if I need something from user?