Emotional Support Conversation FADO: Feedback-Aware Double COntrolling Network for Emotional Support Conversation

[5] Emotion Support Conversation Recent Developments Emotion Support Conversation

Today I would like to share with you a paper on emotional dialogue in KBS. The main idea is to use the user's emotional feedback information to make strategic choices and help generate supportive responses.

Relevant Emotional Support Papers Summary and Guidance- > Click here

FADO:

FADO: Feedback-Aware Double COntrolling Network for Emotional Support Conversation

code link

It is divided into the following four parts:

• Motivation
• Challenges & Contributations
• Model
• Experiment
• Discussion

motivation

• Previous work often considers the use of contextual information when making policy selection. In emotional support dialogue, emotional information is a key factor, but these works ignore the importance of user's emotion to policy selection.
• Previous work selects strategies based on contextual information, which is unidirectional. However, policy information can also in turn help the system to filter the context, making the model more able to pay attention to the context of policy constraints and play a role in filtering noise.

challenge

• What is the relationship between emotion and strategy? How to integrate emotional information into strategy selection?
• How to establish a two-way flow of information between policy information and context information?

solution

• In order to make more accurate and user-related strategy selection, this paper proposes a Dual-Level Feedback Strategy Selector (DFS) that incorporates user emotional feedback to select an appropriate strategy.
• In order to pay attention to the context information of policy constraints, this paper proposes Double Control Reader (DCR) to model the information flow between the two.

Let's illustrate our work with an example:

Overall structure diagram of the model

• Among them, the calculation of the encoding part:

• When the user shows positive emotions, we believe that the strategy and dialogue should be given priority. If the user shows negative emotions, the current strategy should be given lower priority in the future stage.
• The calculation of emotional feedback includes two aspects. The first aspect is the feedback calculation at the dialogue level, and the second aspect is the feedback calculation at the turn level. When the emotional feedback is positive, we will normally optimize the strategy selection; if the emotional feedback is negative, then we should try to avoid the currently selected strategy. Namely: On the negative side, other strategies should be considered more.

• We will control the updated context representation and policy representation through the two-way reader to obtain the updated representation information.
• The motivation of this part of the model is that the selected strategies in the emotional support dialogue are different, and the contextual information of attention should be inconsistent. For example, when the selected strategy is retelling, then we should carry out an average attention distribution on the context ( That is, all information should be considered); if the strategy is to ask questions, then we should pay more attention to the problems described by users.
  

Experimental results

artificial experiment

Ablation experiment

Visualization experiment

• The policy distribution output by our model is more similar to the policy distribution under the real scene.
• Compared with the current state-of-the-art models, our model has relatively few OTHER strategies at the end of the dialogue.
• Similarly, at the beginning of the dialogue, the question strategy probability is not too high.
  

case experiment

• In this part, we not only verified the quality of the generated replies, but also verified whether the generated replies and the strategies adopted were consistent. The dark blue represents the consistent situation, and the light blue represents the inconsistent situation.

• For the last sample, we found that all models are inconsistent, which may be a long-distance dependency problem caused by a long dialogue history.

We also counted the scores for the agreement between generated responses and policies. It is found that our model can greatly surpass the base model.

Summarize

• For better policy selection for emotional support, we propose an emotional feedback-aware bidirectional control network, which utilizes the user's emotional feedback information to help the system choose better policy information to generate policy-related responses.
• It is still worthwhile to consider other psychological states in the emotional support dialogue, such as cultural and educational background, personal portrait, personality, etc. At the same time, it is also a big challenge to combine external knowledge to learn and reason about the user's goals.
• For the interpretable process of emotion and intention understanding, how to customize specific evaluation indicators to measure whether the model really understands the user's emotion.