Affective Natural Language Generation of Event Descriptions through Fine-grained Appraisal Conditions

Yarik Menchaca Resendiz,Roman Klinger
2023-07-26
Abstract:Models for affective text generation have shown a remarkable progress, but they commonly rely only on basic emotion theories or valance/arousal values as conditions. This is appropriate when the goal is to create explicit emotion statements ("The kid is happy."). Emotions are, however, commonly communicated implicitly. For instance, the emotional interpretation of an event ("Their dog died.") does often not require an explicit emotion statement. In psychology, appraisal theories explain the link between a cognitive evaluation of an event and the potentially developed emotion. They put the assessment of the situation on the spot, for instance regarding the own control or the responsibility for what happens. We hypothesize and subsequently show that including appraisal variables as conditions in a generation framework comes with two advantages. (1) The generation model is informed in greater detail about what makes a specific emotion and what properties it has. This leads to text generation that better fulfills the condition. (2) The variables of appraisal allow a user to perform a more fine-grained control of the generated text, by stating properties of a situation instead of only providing the emotion category. Our Bart and T5-based experiments with 7 emotions (Anger, Disgust, Fear, Guilt, Joy, Sadness, Shame), and 7 appraisals (Attention, Responsibility, Control, Circumstance, Pleasantness, Effort, Certainty) show that (1) adding appraisals during training improves the accurateness of the generated texts by 10 pp in F1. Further, (2) the texts with appraisal variables are longer and contain more details. This exemplifies the greater control for users.
Computation and Language
What problem does this paper attempt to address?
The paper attempts to address the issue of how to improve the quality of generated text and achieve fine-grained control over the generation process in affective natural language generation by introducing fine-grained appraisal conditions. Specifically, existing affective text generation models typically rely only on basic emotion theories or emotion values (such as happiness, sadness, etc.) as conditions. This is effective for generating sentences that explicitly express emotions but performs poorly in scenarios where emotions are implicitly expressed. For example, the emotions in event descriptions often do not need to be directly stated (such as "their dog died") but rather inferred by the interlocutor. The authors believe that appraisal theory in psychology can explain the connection between cognitive evaluations and the specific emotions that may arise. These appraisal variables include the ability to control the situation, the sense of responsibility for what happened, and so on. Therefore, they hypothesize that using appraisal variables as conditions in the generation framework can bring two advantages: 1. The generation model can obtain more detailed information about specific emotions, thereby generating text that better meets the conditions. 2. Users can achieve finer control over the generated text by specifying the attributes of the situation (rather than just providing emotion categories). To verify this hypothesis, the authors conducted experiments based on Bart and T5 models, using 7 emotions (anger, disgust, fear, guilt, joy, sadness, shame) and 7 appraisal variables (attention, responsibility, control, situation, pleasantness, effort, certainty). The experimental results show that after adding appraisal variables, the accuracy of the generated text increased by 10 percentage points (F1 score), and the text containing appraisal variables was longer and more detailed. This proves that the introduction of appraisal variables indeed improves the generation quality and provides users with finer control.