Attribute Controlled Dialogue Prompting

“Prompt-tuning has become an increasingly popular parameter-efficient method for adapting large pretrained language models to downstream tasks. However, both discrete prompting and continuous prompting assume fixed prompts for all data samples within a task, neglecting the fact that inputs vary greatly in some tasks such as open-domain dialogue generation. In this paper, we present a novel, instance-specific prompt-tuning algorithm for dialogue generation. Specifically, we generate prompts based on instance-level control code, rather than the conversation history, to explore their impact on controlled dialogue generation.

In this work, we design a lightweight prompting module for adapting pretrained language models for attribute controlled dialogue generation. More precisely, for each attribute such as persona, intention, emotion etc. we only save an additional prompt module. Since the prompting module is a fraction of the size of the pretrained dialogue model, this allows many controlled dialogue systems to be stored on a device without too much overhead. We present results on both intent and persona-controlled dialogue.

Continuous prompts extend prompt selection to the entire space of embeddings, including vector embeddings that do not correspond to any human-interpretable natural language tokens. Hence, soft prompts are more expressive than discrete prompts.

However, both deep prompts and shallow prompts assume a static prompt / task-level prompt for all samples within a task, neglecting the fact that samples might vary greatly, especially in the field of conversation generation.

In contrast to previous work, we propose Controlled DialogPrompt for applying prompt-tuning in controlled dialogue generation, which optimizes prompts based on provided control codes rather than the previous conversation history and we further explore the controllability of prompts at the instance level.”

In summary, we presented a novel prompting technique, conditioned on a dialogue attribute (persona or intent), for controlled dialogue generation. The prompting module requires only 5%-6% of the total number of parameters, which allows the storage of several fined-tuned prompting modules for different dialogue generation tasks at a fraction of the cost of a full dialogue model.

However, Controlled DialogPrompt currently studies conditioning on simple control attribute sentences like the user’s persona and the work can be extended to more extensive and complex sentences such as background knowledge documents to further evaluate the controlled prompt’s encoding capabilities. Additionally, combining multiple Controlled DialogPrompts on several control attributes and automatically triggering various dialogue skills is an interesting and unexplored direction.