A recent survey on controllable text generation: A causal perspective.

Area of Science:

• Natural Language Processing
• Artificial Intelligence

Background:

• Controllable Text Generation (CTG) is crucial in natural language generation, focusing on text generation with added constraints.
• Existing CTG models primarily rely on statistical associations from training data, often neglecting causal relationships.

Purpose of the Study:

• To review recent Controllable Text Generation (CTG) approaches through a causal lens.
• To identify challenges arising from the lack of causality in current CTG models.
• To explore advancements addressing these causal limitations.

Main Methods:

• Review of foundational CTG models to understand their reliance on statistical associations.
• Identification and categorization of four key challenges stemming from the absence of causality.
• Analysis of recent CTG improvements in representation disentanglement, causal inference, knowledge enhancement, and multi-aspect CTG.
• Inspection of current evaluation methods for CTG, with a focus on causality assessment.

Main Results:

• Existing CTG models predominantly capture statistical associations, not causal underpinnings.
• Four primary challenges are identified due to the lack of causality in CTG.
• Improvements in CTG are emerging through disentangled representations, causal inference, knowledge integration, and multi-aspect controls.
• Current evaluation metrics for CTG often overlook causal considerations.

Conclusions:

• There is a significant need to incorporate causality into Controllable Text Generation (CTG) research.
• Future research should focus on developing CTG methods that explicitly model and leverage causal relationships.
• Enhanced evaluation methodologies are required to accurately assess the causal capabilities of CTG systems.