[Paper Review] Few-Shot Bot: Prompt-Based Learning for Dialogue Systems

Andrea Madotto, Zhaojiang Lin+2 more|arXiv (Cornell University)|Oct 15, 2021

Topic ModelingComputer Science103 references45 citations

TL;DR

Prompt-based few-shot learning with large language models can achieve competitive results to fully trained dialogue models across diverse tasks, and the Few-Shot Bot (FSB) automatically selects skills using a prompt-based Skill Selector without fine-tuning.

ABSTRACT

Learning to converse using only a few examples is a great challenge in conversational AI. The current best conversational models, which are either good chit-chatters (e.g., BlenderBot) or goal-oriented systems (e.g., MinTL), are language models (LMs) fine-tuned on large conversational datasets. Training these models is expensive, both in terms of computational resources and time, and it is hard to keep them up to date with new conversational skills. A simple yet unexplored solution is prompt-based few-shot learning (Brown et al. 2020) which does not require gradient-based fine-tuning but instead uses a few examples in the LM context as the only source of learning. In this paper, we explore prompt-based few-shot learning in dialogue tasks. We benchmark LMs of different sizes in nine response generation tasks, which include four knowledge-grounded tasks, a task-oriented generations task, three open-chat tasks, and controlled stylistic generation, and five conversational parsing tasks, which include dialogue state tracking, graph path generation, persona information extraction, document retrieval, and internet query generation. The current largest released LM (GPT-J-6B) using prompt-based few-shot learning, and thus requiring no training, achieves competitive performance to fully trained state-of-the-art models. Moreover, we propose a novel prompt-based few-shot classifier, that also does not require any fine-tuning, to select the most appropriate prompt given a dialogue history. Finally, by combining the power of prompt-based few-shot learning and a Skill Selector, we create an end-to-end chatbot named the Few-Shot Bot (FSB), which automatically selects the most appropriate conversational skill, queries different knowledge bases or the internet, and uses the retrieved knowledge to generate a human-like response, all using only few dialogue examples per skill.

Motivation & Objective

• Motivate reducing reliance on heavy gradient-based fine-tuning for dialogue systems by leveraging prompt-based learning.
• Benchmark prompt-based few-shot learning across a wide set of dialogue tasks and datasets (knowledge-grounded, task-oriented, open-chat, and parsing).
• Introduce a Skill Selector to map dialogue histories to the most suitable prompt/skill without training.
• Propose the Few-Shot Bot (FSB) that automates skill selection, knowledge retrieval, and response generation using only few-shot prompts per skill.

Proposed method

• Use a language model to compute P(Y|X, Prompt_k) via concatenated few-shot prompts (X=dialogue history, Y=response) without gradient updates.
• Design task-specific prompts (Prompt_k) with k-shot demonstrations to guide response generation, baseline by comparing with fine-tuning."
• Introduce the Prompt-Based Few-Shot Skill Selector to choose the most appropriate prompt given dialogue history by evaluating P(X|Prompt_k^y) for each candidate skill y.
• Implement the Few-Shot Bot (FSB) by combining the Skill Selector with multiple prompts and external knowledge sources (KBs, the Internet, KG) for knowledge-grounded responses.
• Use GPT-J (6B) as the primary LM for local deployments and GPT-Jurassic via API for access to larger models, with 1-shot prompts and 0/1/k-shot experiments across tasks.

Experimental results

Research questions

• RQ1 Can prompt-based few-shot learning with large LMs achieve competitive results to fully trained models on diverse dialogue tasks without fine-tuning?
• RQ2 Does a prompt-based Skill Selector effectively identify the most suitable prompt/skill for a given dialogue history?
• RQ3 How does model size and number of shots affect performance across response generation and conversational parsing tasks?
• RQ4 What is the comparative gap between prompt-based few-shot methods and state-of-the-art fully trained models in perplexity and evaluation metrics?
• RQ5 Can an end-to-end Few-Shot Bot leverage multiple knowledge sources to generate human-like, grounded responses without model training?

Key findings

• Larger language models (GPT-J 6B) consistently outperform smaller models across tasks, showing a linear relationship between size and performance.
• More shots do not always improve results; in some knowledge-grounded tasks the best performance occurs with fewer shots.
• Conversational parsing tasks are harder for prompt-based learning, with larger gaps to SOTA fully trained models that may benefit from decoding strategies like beam search.
• Perplexity gaps between FSB and SOTA are relatively small, suggesting potential gains from improved sampling strategies.
• The prompt-based Skill-Selector achieves competitive results to fine-tuned RoBERTa baselines while avoiding any training, enabling easy extension with new skills.
• The end-to-end Few-Shot Bot can access multiple knowledge sources, track user preferences, and generate grounded responses using only few-shot prompts per skill.