1 of 4

Multi-Agent Personalized Dialogue System

by 凯华 倪

2 of 4

Approaches

Five agents for personalized dialogue generation

Motivation 

1. Data scarcity and generalization issues

2. Need for more engaging and authentic interactions

  1. Interlocutors Information Agent (IIA)
  2. Dialogue History Agent (DHA)
  3. Topic Planning Agent (TPA)
  4. Dialogue Generation Agent (DGA)
  5. Feedback and Adjustment Agent (FAA)

Multi-aspect optimization

  1. Real-time profile and mood updates
  2. Coherent and engaging dialogue generation
  3. Continuous feedback and adjustment

3 of 4

Architecture of MAPDS In CPDC

4 of 4

Performance on CPDC dataset with different structure

Agent

Input

Output

1 . IIA

Profile

Persona

2 . DHA

Dialogue history

History summary

3 . DGA

1, 2

Dialogue

Agent

Input

Output

1 . IIA

Profile, Dialogue history

Persona, Language style, Emotional state

2 . DHA

Dialogue history

History summary

3 . DGA

1, 2

Dialogue

Agent

Input

Output

1 . IIA

Profile, Dialogue history

Persona, Language style, Emotional state

2 . DHA

Dialogue history

History summary, Topic transitions, Current topic

3 . DGA

1, 2

Dialogue

Agent

Input

Output

1 . IIA

Profile, Dialogue history

Persona, Language style, Emotional state

2 . DHA

Dialogue history

History summary, Topic transitions, Current topic

3. TPA

1, 2

Planned topic

4 . DGA

1, 2, 3

Dialogue

Agent

Input

Output

1 . IIA

Profile, Dialogue history

Persona, Language style, Emotional state

2 . DHA

Dialogue history

History summary, Topic transitions, Current topic

3. TPA

1, 2

Planned topic

4 . DGA

1, 2, 3, 5

Dialogue

5 . FAA

Profile, Dialogue history, 4

Feedback

Structure 1 -- 2.12

Structure 2 -- 2.73

Structure 3 -- 3.54

Structure 4 -- 4.06

Structure 5 -- 4.35