Fig 1.
Schematic representation of the proposed architecture.
Agents are set up by providing relevant context and role-playing character information already integrated in a memory system. In interactions, user input gets stored into the memory system and triggers appraisal (i.e., explicit emotion expression) that is also stored in the memory system. Based on the current state of the memory system, agent output is generated.
Fig 2.
Example of model input and output for the three conditions.
The input of the No-Memory and Memory condition is the same for the first item. For the No-Memory condition, all following items only include the example question and answer, as well as the next question in the scale. The Memory condition includes all prior questions and generated answers. The Appraisal-Prompts condition is the same as the Memory condition, but the example answer is changed to include two steps: First, appraising the situation to generate emotions of the involved person and second, providing the answer.
Table 1.
STEU scores (out of 42) by condition.
Each STEU item can either be right (1) or wrong (0).
Fig 3.
Results of the comparison between conditions.
The Y-axis represents cumulative STEU score by item and the X-axis represents individual STEU items.
Fig 4.
Illustration of prompting strategies for all three conditions.
The No-Memory condition constructs a prompt out of the system instruction and the user input. The Memory condition constructs a prompt out of the system instruction, message history, and user input. The Chain-of-Emotion condition uses a separate LLM call to appraise the agent’s emotion for each message before creating a user response. It therefore constructs a prompt out of the system instruction, the message history, the history of the emotions generated through the appraisal step, and the user input.
Table 2.
Descriptive overview of LIWC variables per output sentence by condition for the fixed prompt responses with F and p values of the significance test.
Fig 5.
Screenshots of the conversational game “Wunderbar”.
The left screenshot shows dialog provided by the model. The user can click to continue each dialog line until the input field for a response appears (right screenshot).
Table 3.
Descriptive overview of user research variables per condition with F and p values of the significance test.
Each item has a minimum value of 0 and a maximum value of 6.
Table 4.
Descriptive overview of LIWC variables per participant by condition for all outputs generated in the user study with F and p value of the significance test.