Self-Emotion Blended Dialogue Generation
in Social Simulation Agents

Each speaker agent has its profile generated by GPT-4. A profile contains information about the speaker’s basic information such as name, age, gender, etc. Besides, each profile of an agent contains a “description” field providing information of the past experience (See Table 6). This is helpful for further generation of events and analysis of self-emotion status.

Based on their current self-emotional states and the ongoing conversation context, agents are prompted to choose the most appropriate dialogue strategies for their next actions. Dialogue strategies are selected from a pre-defined strategy pool that contains 11 dialogue strategies adapted from the taxonomy of empathetic response intents Welivita and Pu (2020). A full list of the strategies can be found in Table 8.

In the context of empathetic dialogue models and datasets, it is common to represent emotions using discrete labels Li et al. (2017); Hsu et al. (2018); Rashkin et al. (2019). During a conversation, speakers are randomly assigned one emotion label from a predefined pool, such as those used in the EmpatheticDialogues (ED) dataset Rashkin et al. (2019), as their self-emotion. We utilize labels from the ED dataset because they offer fine-grained distinctions between similar emotions. The self-emotion is directly represented as a sentence of “feeling <label>”. For example, if the emotional label “excited” is selected, the self-emotion might be represented as “<name> is feeling excited right now.”

Individuals’ self-emotion may be influenced by some random events that happen to them. To capture this, we represent self-emotion as an emotional label accompanied by an associated event. For example, “My promotion has been approved.” is an event that could evoke the emotion of “excited”. The self-emotion of this event could be represented as “I’m feeling excited because my promotion has been approved.” This approach allows us to incorporate more causal information into self-emotion, enabling speakers to potentially leverage this information in their future actions.

People with different personalities and past experiences may generate different self-emotions for identical events. For instance, a person with acrophobia may feel “fear” when riding a roller coaster, while others may feel “excited.” Therefore, we also consider a method of representing self-emotion using events related to the profiles of each speaker, referred to as “profile events.” Table 1 provides examples of self-emotion represented in three different ways.

Each conversation in the ED dataset includes two speakers. To ensure our agents maintain consistent personal backgrounds for both speakers, the original conversations in the dataset are provided to GPT-4 when generating agent profiles. The LLM is tasked with generating profiles of two individuals who could plausibly have the provided conversation. Figure 7 illustrates the prompt used for generating these profiles.

When having a conversation, each speaker talks according to their own profile as well as the first 2 or 3 utterances (depending on the number of utterances) at the beginning of each dialogue in the ED dataset. The speaker agents are tasked with two objectives simultaneously: 1) selecting the best strategies from a given strategy pool, and 2) generating the future conversation based on the selected strategies. This prompt is shown in Figure 6.

In this case, each speaker has their own self-emotion before engaging in a conversation. Self-emotions are generated by prompting different LLMs. Conversations are then generated similarly to the method used without self-emotion, except that self-emotion is included as part of the input, prepended to the beginning of the dialogue context. Figure 10 presents the prompt the speaker agents use to generate conversations with self-emotion. We utilize Chain-of-Thought Wei et al. (2022) technique in all the prompts, as it performs well in text classification tasks and is therefore useful for generating the best strategies.

Five language models are used as the backend of the speaker agents in this experiment: Mistral-7B-Instruct Jiang et al. (2023), Llama-2-7B-Chat Touvron et al. (2023), Gemma-2B-It Team et al. (2024), gpt-3.5-turbo and gpt-4 ²²2We use the gpt-3.5-turbo-0125 for gpt-3.5-turbo and gpt-4-0125-preview for gpt-4..

The experiment is conducted on the test set of the ED dataset, resulting in the generation of 2547 conversations for each self-emotion representation approach. Human annotations are collected as the ground truth, and we define the strategy accuracy as the cosine similarity between the model-predicted strategy and the human strategy:

$$Acc=\frac{S_{m}\cdot S_{h}}{\left\|S_{m}\right\|\left\|S_{h}\right\|}$$

(1)

Here, $S_{m}$ represents the list of strategies chosen by the model and $S_{h}$ is the list of strategies annotated by humans.

Table 2 presents the results of strategy accuracy for different representations of self-emotion. We are able to observe that within the same dialogue context, LLMs exhibit improved strategy selection when prompted with self-emotion. The random event self-emotion yields the highest performance, outperforming profile events. Additionally, among all models examined, GPT-4 demonstrates the most effective performance.

Figure 2 illustrates the relationship between the most frequent self-emotions and corresponding strategies. It shows that for negative self-emotions such as “anxious” and “nervous,” the models tend to express more pessimistic strategies such as “expressing concern” and “sympathizing.” Conversely, for positive self-emotions like “proud” and “joyful,” the models lean towards more optimistic strategies such as “encouraging.” Additionally, neutral strategies such as “sharing own thoughts” and “sharing experience” are commonly employed across both positive and negative self-emotions as the most frequently used strategies.

We employ the same workflow as described in Section 4 to generate conversations both with and without self-emotion using GPT-4. These generated conversations will then be used as training data to train the small scale models. Different from the previous experiment, we generate using only the random event (as it demonstrates the highest strategy accuracy) on the full ED dataset, resulting in a final train/val/test split of 14,274/2,762/3,569 after filtering invalid cases with incorrect formats. Table 7 shows an example of the generated conversation.

The purposes of training a small-scale model are to enhance deployment convenience and to explore how effectively the capabilities of LLMs in understanding self-emotion can be transferred to a smaller-scale model. To do this, we fine-tune a FLAN-t5-large model Chung et al. (2024) on the collected datasets. Given the seq2seq architecture of the model, each conversation in the dataset is split into multiple turns between the two speakers. For each turn, the utterance of the first speaker serves as the input, and the utterance of the other speaker is treated as the label. The task instruction is then prepended to form a training instance. For instance, an example of the input in a training instance without self-emotion is:

The corresponding label is: “me: <utterance_4>. <eos_token>.” For models with self-emotion, the self-emotion is included in the task instruction:

The model training process was implemented using the HuggingFace framework³³3Model link: https://huggingface.co/google/flan-t5-large. The models were trained on NVIDIA A100 GPUs for 72 hours with a learning rate of 3e-4. The maximum input length was set to 512 tokens, in consideration of the original base model’s length window. For inference generation, the temperature was set to 0.7.

The models are evaluated on ROUGE Lin (2004), BLEU Papineni et al. (2002) and BERT-score Zhang et al. (2019). Table 3 shows the automatic metrics of the models fine-tuned on our collected self-emotion datasets.

We follow the method of ACUTE-Eval Li et al. (2019) and assess the models across four axes: naturalness, empathy, interestingness, and humanness. Naturalness assesses the ability to provide smooth, natural responses. Similar to the ED dataset, we use empathy to represent the model’s ability to understand emotions. Interestingness reflects the ability to generate interesting and diverse responses, while humanness is used to evaluate the ability to choose human-like strategies in the conversation. For each model, 100 conversations are generated in a self-chat manner Li et al. (2016), where two models are programmed to talk to each other. Table 9 shows the questionnaire used for human evaluation.

Group member creation involves creating a profile for each member, including their roles, positions, and background, by inputting the description of the group into GPT-4. The role of the a member is either the “leader” or “member”, where the “leader” will serve as the host of the discussion by pushing the topic to next steps. Each “member” has their own position and background which are related to their occupation and past experience to trigger self-emotion. Figure 13 shows the prompt we use to create group members.

Topic generation is the process of generating the topics that group members engage in. To capture the decision-making process, each topic is divided into several steps. For example, the topic of “organizing a group trip to Italy with a limited budget of $1500 per person” can be broken down into steps such as choosing dates, selecting flights, deciding on attractions, choosing hotels, and so on. Figure 14 is the prompt we use to generate different topics.

Agents follow the steps of the topic and have discussions. The agents are required to reach an agreement before moving to the next step. The “leader” of the group judges whether an agreement has been reached by analyzing the discussion history. During a discussion, a hidden “manager” will decide the next speaker by analyzing the positions of the members and discussion context. For instance, if the “manager” decides that a structural engineer should pose an idea about the material, it will set the structural engineer as the next speaker. The “manager” does not participate in the discussion by raising its own opinions.

As shown in Figure 3, in order to facilitate the self-emotion, we simulate complete process of an agent encountering events, stimulating self-emotion, taking behaviors and participating in the group discussions by prompting LLMs. Each agent maintains its own goals and self-emotion. For example, in a discussion about “building a house and maximizing profits within a limited budget,” the structural engineer may aim to secure better materials while the landscape engineer may prioritize budget allocation for sustainability. This way, agents can develop rich discussion content by expressing their own ideas, which might be affected by their self-emotions.

We assess discussions on 5 topics: house building, hosting a charity event, planning a trip, organizing a welcome party, and developing a mobile app. For each topic, we generate a group with 6 members, where each member has its own role and position. We run 10 different discussions, and in each discussion, agents will encounter their own events which will cause the self-emotion. We then compare the decisions made in these discussions to those made in a discussion where the self-emotion of the agents is disabled.

For evaluation, we examine the percentages of decision changes after incorporating self-emotion. Specifically, we categorize these changes into six types:

• •

  Undecided change: discussions that shift from an agreement to delegation.

• •

  Decided change: discussions move from delegation to agreement.

• •

  Authority change: a decision made by voting changes to being made by a single agent.

• •

  Majority change: a decision made by a single agent changes to a majority vote.

• •

  Details change: the overall direction does not change, but specific details do (e.g., changing “spending $30 for dinner” to “$20”).

• •

  Compromise change: a decision shifts from full agreement by all agents to a compromised agreement where one or more agents make concessions.

Figure 4 shows the average percentage of different categories of decision changes influenced by positive and negative self-emotion across all topics. We observe that a significant portion of decisions are affected: around 66% by negative and 51% by positive self-emotion.

For different categories of changes, we find that negative self-emotion leads to more undecided, majority, and compromise changes. This suggests that agents with negative self-emotion tend to express their opinions more, resulting in delegation or compromise in decision-making, which aligns with the findings in Koch et al. (2013). In contrast, positive self-emotion tends to lead to more agreements, with most changes involving the details of plans without altering the main direction of the decision. A comprehensive table of the decision change rates across topics can be found in Table 10.

Additionally, an analysis of the average length and frequency of utterances indicates that agents with positive self-emotion tend to be more active. Discussions reach agreements more quickly when agents have negative self-emotion (Table 11).

Figure 5 shows a discussion on the topic of “APP development”. The self-emotion of the front-end developer (FD) influences the discussion and ultimately leads to a decision change from “using React Native and JavaScript as the development tools” to “Kotlin.” In this case, despite being more agreeable when no self-emotion is introduced, the FD, experiencing a “sad” self-emotion, adopts a more objective stance and proposes a different idea. Similar patterns emerge in other topics, where members with negative self-emotion tend to express more objections.