Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data...
Transcript of Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data...
![Page 1: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/1.jpg)
![Page 2: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/2.jpg)
Ian LaneCarnegie Mellon University
Conversational Systems in the Era of Deep Learning and Big Data
![Page 3: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/3.jpg)
End-to-End Trainable Neural Network Models for Task Oriented Dialog
Ian LaneCarnegie Mellon University
3
User Agent
![Page 4: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/4.jpg)
Trainable End-to-End Models for Dialog Can we learn to perform task-oriented dialog by modeling
human conversations?
4
User Agent
![Page 5: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/5.jpg)
Trainable End-to-End Models for Dialog Can we learn to perform task-oriented dialog by modeling
human conversations?
5
User I’m looking for an Italian Restaurant
for Dinner.
Agent
![Page 6: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/6.jpg)
Trainable End-to-End Models for Dialog Can we learn to perform task-oriented dialog by modeling
human conversations?
6
User I’m looking for an Italian Restaurant
for Dinner.
Agent
Sure I can help you with that.
Whereabouts are you thinking?
![Page 7: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/7.jpg)
Trainable End-to-End Models for Dialog Can we learn to perform task-oriented dialog by modeling
human conversations?
7
User I’m looking for an Italian Restaurant
for Dinner.
Somewhere in Menlo Park or Palo Alto.
Agent
Sure I can help you with that.
Whereabouts are you thinking?
![Page 8: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/8.jpg)
Trainable End-to-End Models for Dialog Can we learn to perform task-oriented dialog by modeling
human conversations?
8
User I’m looking for an Italian Restaurant
for Dinner.
Somewhere in Menlo Park or Palo Alto.
Well there is a couple of good
places …
Agent
Sure I can help you with that.
Whereabouts are you thinking?
![Page 9: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/9.jpg)
Trainable End-to-End Models for Dialog Can we learn to perform task-oriented dialog by modeling
human conversations?
9
User Agent Model
Somewhere in Menlo Park or Palo Alto.
Output 2
Output 1
Agent
I’m looking for an Italian Restaurant
for Dinner.
Knowledge base
![Page 10: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/10.jpg)
End-to-End Models for Dialog
10
![Page 11: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/11.jpg)
End-to-End Models for Dialog
11
![Page 12: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/12.jpg)
Trainable Task-Oriented Models for Dialog Agent
• Understand user’s input in context• Take Action to take given user input, dialog and task history
• Respond to user?• Perform API call / database look up?• Perform other action?• Or combination of above.
• Update dialog state
User• Estimate of user goal / goal change
12
![Page 13: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/13.jpg)
An Neural Network Model for Dialog Agents Agent Model
• Perform de-lexicalization on input (i.e. Named Entity Recognition)• Update dialog state (via. LSTM model)• Estimate beliefs state • Action: Issue API Call and process results (if required)• Action: Update reference to results (if required)• Action: Response generation
• Estimate de-lexicalized response that best matches current state from all possible responses
13
![Page 14: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/14.jpg)
End-to-End Neural Network Models for Dialog
14
![Page 15: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/15.jpg)
Training - Approach Manual Annotation of Training Data Not Required
• Assumes NLU for delexicalization and named entity detection• Train on dialogs from human-to-human chat interaction• Assumes turn-based interaction can be extend to also model timing
Supervized Training• Treat as a classification task per turn Select best output given history
Reinforcement Learning• Optimize based on end-of-dialog reward
15
![Page 16: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/16.jpg)
Training – Results and Challenges Prediction Accuracy
Issues• Even if action taken by agent at specific turn does not match training
“training data” it may still be appropriate• “training data” is fixed, User does not respond based on output of Model
Can we simulate complete USER-AGENT interactions to improve dialog?
16
![Page 17: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/17.jpg)
Joint Modeling of Users and Agents
User
17
AgentI’m looking for an Italian Restaurant
for Dinner.
Somewhere in Menlo Park or Palo Alto.
Well there is a couple of good
places …
Sure I can help you with that.
Whereabouts are you thinking?
User Goal
![Page 18: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/18.jpg)
Joint Modeling of Users and Agents
User Model
18
Agent Model
AGENT
I’m looking for an Italian Restaurant.
Somewhere in Menlo Park or Palo Alto.
Well there is a couple of good
places …
Sure I can help you with that.
Whereabouts are you thinking?
USER
Knowledge base
User
Goal
Can we jointly model task-oriented dialog to simulate human conversations?
![Page 19: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/19.jpg)
Joint Modeling of Users and AgentsUser Model
19
Agent Model
![Page 20: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/20.jpg)
Joint Modeling of Users and AgentsUser Model
20
Agent Model
![Page 21: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/21.jpg)
Joint Modeling of Users and AgentsUser Model
21
Agent Model
![Page 22: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/22.jpg)
Joint Modeling of Users and Agents
22
![Page 23: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/23.jpg)
Dialog Success Rate
• Significantly higher dialog success rate using jointly optimized models (65%) than optimizing AGENT only (50%), or baseline trained via supervised learning (35%)
Results (DSTC 2 Task)
23
![Page 24: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/24.jpg)
24
![Page 25: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/25.jpg)
Conclusions and Challenges• Training task oriented dialog systems directly from human
conversations seems plausible• End-To-End trainable give logs of dialog + API calls• Well suited for task-oriented dialogs that include social, task-orientated dialog
and external knowledge base look-up or actions• i.e. Product and service support (Call-Centers / Messaging Interfaces)
• Ability to model both Agent and User in a conversation• With data can model personas or individual Agents or Users
• Exploring boot-strapping dialog systems ”on-the-fly” • Generate suggested response … see if Agent follows suggestion or not • Update models for next utterance / dialog
25
![Page 26: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/26.jpg)
![Page 27: Conversational Systems in the Era of Deep Learning …...Era of Deep Learning and Big Data End-to-End Trainable Neural Network Models for Task Oriented Dialog Ian Lane Carnegie Mellon](https://reader030.fdocuments.in/reader030/viewer/2022040620/5f30abb6f47def18d9438ed7/html5/thumbnails/27.jpg)