ChatPLUG: Open-Domain Generative Dialogue System with Internet-Augmented Instruction Tuning for Digital Human (2304.07849v3)

Abstract: In this paper, we present ChatPLUG, a Chinese open-domain dialogue system for digital human applications that instruction finetunes on a wide range of dialogue tasks in a unified internet-augmented format. Different from other open-domain dialogue models that focus on large-scale pre-training and scaling up model size or dialogue corpus, we aim to build a powerful and practical dialogue system for digital human with diverse skills and good multi-task generalization by internet-augmented instruction tuning. To this end, we first conduct large-scale pre-training on both common document corpus and dialogue data with curriculum learning, so as to inject various world knowledge and dialogue abilities into ChatPLUG. Then, we collect a wide range of dialogue tasks spanning diverse features of knowledge, personality, multi-turn memory, and empathy, on which we further instruction tune \modelname via unified natural language instruction templates. External knowledge from an internet search is also used during instruction finetuning for alleviating the problem of knowledge hallucinations. We show that \modelname outperforms state-of-the-art Chinese dialogue systems on both automatic and human evaluation, and demonstrates strong multi-task generalization on a variety of text understanding and generation tasks. In addition, we deploy \modelname to real-world applications such as Smart Speaker and Instant Message applications with fast inference. Our models and code will be made publicly available on ModelScope: https://modelscope.cn/models/damo/ChatPLUG-3.7B and Github: https://github.com/X-PLUG/ChatPLUG .

An Overview of ChatPLUG: Internet-Augmented Instruction Tuning for Open-Domain Dialogue Systems

The paper introduces ChatPLUG, a Chinese open-domain generative dialogue system explicitly engineered for digital human applications. Unlike conventional dialogue models that predominantly rely on scaling model size or the dialogue corpus, ChatPLUG employs internet-augmented instruction tuning to boost multi-task generalization and practicality in diverse scenarios, emphasizing open-world knowledge, distinct personality, and multi-turn memory.

Key Components and Methodology

The development of ChatPLUG comprises three pivotal stages:

1. Large-Scale Dialogue Pre-training: Initially, ChatPLUG undergoes extensive pre-training on vast collections of text from common document corpora and dialogue data. This stage hones ChatPLUG’s ability to assimilate wide-ranging world knowledge and dialogue proficiencies through curriculum learning. The document pre-training is executed using both denoising and prefix LM objectives, which lay a foundational comprehension of language.
2. Internet Knowledge Acquisition: Given the inevitable knowledge gaps in pre-trained models, ChatPLUG integrates a real-time internet search module. This is crucial for continuously updating the dialogue system’s knowledge base to address issues like knowledge hallucination. The process involves reformulating queries based on user interaction context, thus ensuring that strategically relevant external knowledge is incorporated into responses.
3. Internet-Augmented Instruction Tuning: In this critical phase, ChatPLUG’s ability to perform across diverse dialogue tasks is bolstered by deploying instruction tuning. Here, unified natural language instruction templates are used to fine-tune ChatPLUG across a variety of collected dialogue datasets, each containing unique task features.

Evaluation and Results

ChatPLUG’s performance is evaluated against existing state-of-the-art models in both automatic and human assessments. It consistently demonstrates superior capabilities in coherence, informativeness, persona alignment, safety, and reduced hallucinations compared to other Chinese dialogue models such as PLATO-XL, EVA 2.0, and ChatGLM.

In terms of automatic metrics, ChatPLUG outstrips its counterparts in ROUGE-L and BLEU scores, indicating its ability to generate more linguistically varied and comprehensible responses. Notably, the incorporation of internet-augmented knowledge ensures that ChatPLUG delivers factually accurate responses, mitigating the prevalent problem of knowledge hallucination in dialogue systems.

Human evaluations further corroborate these findings, illustrating ChatPLUG’s proficiency in maintaining engaging interactions while adhering to personality traits and user expectations in real-world applications. Moreover, its successful deployment in practical settings such as smart speakers and instant messaging platforms typifies its versatility and readiness for commercialization.

Implications and Future Directions

The development of ChatPLUG underscores the potential of leveraging internet-augmented instruction tuning for enhancing dialogue systems. By amalgamating extensive pre-trained knowledge bases with real-time internet information, ChatPLUG navigates the challenge of outdated or incorrect information which has traditionally hindered dialogue agents.

The architecture’s ability to customize dialogue style and character traits offers expansive opportunities in personalized user interactions, paving the way for more natural and human-like digital aids. As future work, these insights open avenues for further exploring the efficiencies of instruction tuning in other languages and task domains and amalgamating feedback-based learning for safety and user alignment in broader AI systems.

ChatPLUG exemplifies an innovative stride in building practical, robust dialogue systems that bridge the gap between technological capabilities and user demands. By seamlessly integrating state-of-the-art model fine-tuning with dynamic, internet-augmented information curation, ChatPLUG sets a new benchmark for multi-task generalization and applicability across AI-driven communication platforms.