Personality-affected Emotion Generation in Dialog Systems

TL;DR

This work introduces Personality-affected Emotion Generation, a task that generates dialog responses with emotions aligned to a given personality. It frames mood transitions in the Valence-Arousal-Dominance space and conditions transitions on the Big Five traits, using a two-module model: Mood Transition Regression and Emotion Generation, with affective information extracted via a multimodal-attentive mechanism. The authors release the PELD dataset, a 6,510-table dataset of daily-dialog triples from Friends with emotion and personality annotations, and show that incorporating mood-transition modeling and personality improves emotion-generation performance, especially for minority emotions, compared to a BERT-base baseline. They also analyze correlations between mood transitions and personality and discuss how different personalities influence mood dynamics in conversation. The work lays groundwork for personality-aware affect in dialog systems and points to future multi-modality extensions and deeper analyses of content semantics in responses.

Abstract

Generating appropriate emotions for responses is essential for dialog systems to provide human-like interaction in various application scenarios. Most previous dialog systems tried to achieve this goal by learning empathetic manners from anonymous conversational data. However, emotional responses generated by those methods may be inconsistent, which will decrease user engagement and service quality. Psychological findings suggest that the emotional expressions of humans are rooted in personality traits. Therefore, we propose a new task, Personality-affected Emotion Generation, to generate emotion based on the personality given to the dialog system and further investigate a solution through the personality-affected mood transition. Specifically, we first construct a daily dialog dataset, Personality EmotionLines Dataset (PELD), with emotion and personality annotations. Subsequently, we analyze the challenges in this task, i.e., (1) heterogeneously integrating personality and emotional factors and (2) extracting multi-granularity emotional information in the dialog context. Finally, we propose to model the personality as the transition weight by simulating the mood transition process in the dialog system and solve the challenges above. We conduct extensive experiments on PELD for evaluation. Results suggest that by adopting our method, the emotion generation performance is improved by 13% in macro-F1 and 5% in weighted-F1 from the BERT-base model.

Figures (9)

• Figure 1: A dialog snippet from a real conversation with a chatbot. The chatbot first optimistically encourages the user with a Joy emotion. But after the encouragement, the chatbot pessimistically shows empathy to the user in a Sadness emotion even the user expressed a positive attitude. A more appropriate emotional response here should still be in Joy so that consistent with the previous response.
• Figure 2: Mood domains and emotions in the VAD space. We only illustrate the plane of Valence (Pleasure-displeasure) and the Arousal (Degree of arousal) axis here because the mood domains are split only by these two dimensions itoh2009mood. $M_1$, $M_2$, $M_3$, and $M_4$ are different quadrants representing different mood states.
• Figure 3: The Model Illustration. The upper-left part is the mood transition regression, where the initial mood state $M_i$ changes to the $M_r$ in the upcoming response. The personality trait $P$ is transformed to the transition weights $\mathcal{P_V, P_A, P_D}$, and the extracted affective information from the dialog context is changed to the transition variables $\Delta V, \Delta A, \Delta D$. Then, the upper-right part is the emotion generation module, where the response $E_r$ is generated through integrating $M_r$, $P$, and the dialog context $R_c$. The lower part illustrates how we extract affective information by the Attention Layer aligning the emotion annotations, word-level VAD embeddings, and the semantic representations from the input dialog context.
• Figure 4: A triple example in PELD. The dyadic conversation between Rachel and Monica (two main roles in Friends, $P$ is the personality of Rachel. In this example, the dialog system is set as Rachel and generate the emotion Anger in the response $U_3$ to the user set as Monica.
• Figure 5: Mood transition matrixes of the six main roles in PELD. Each row in a matrix shows the ratios of the current mood state $M_i$ is transferred to the next mood state $M_r$.