毕业生知识库

       随着移动通信和网络技术的快速发展与普及应用，智能客服和智能助手的需求极大地上升，任务型对话系统作为其中最重要的一种技术类型，是当前自然语言处理领域的一大研究热点。任务型对话系统旨在以有限的对话轮数帮助用户完成特定的任务，在虚拟个人助手、智能家居和智能客服等领域都有着广泛应用。

       随着全球化进程的加速发展，跨语言交流变得愈发重要。然而，现有的任务型对话系统主要利用某一种语言的数据进行训练，在进行多语言迁移和扩展时，面临较大挑战：一方面，部分语种的对话数据较为稀缺，无法满足当前深度学习对话系统的数据要求；另一方面，倘若采用机器翻译系统充当跨语言对话系统的桥梁，则存在鲁棒性差、响应速度低和部署难度大等问题。为此，本文研究如何有效利用高资源语言的对话系统和相关的多语言资源，快速实现低资源语言场景下的对话系统构建。本文的主要贡献和创新归纳如下：

1.提出了一种面向多粒度翻译噪声的鲁棒对话系统构建方法

       针对“翻译-对话-回翻”三级管道式跨语言对话系统由于误差累积导致的对话系统鲁棒性差的问题，提出了一种面向多粒度翻译噪声的鲁棒对话系统构建方法。该方法首先基于机器翻译多粒度噪声知识，从词汇、短语和句子三种粒度构造包含翻译噪声的对抗样本，然后采用两种对抗学习策略增强对话系统的鲁棒性。具体地，基于语句的对抗学习策略利用多粒度对抗样本和干净数据进行对抗训练，指导对话系统学习与噪声无关的隐层表示；基于知识的引导方法则将多粒度噪声知识显式加入到噪声知识库中，从而增强对话系统的鲁棒性。实验表明，所提方法能够显著改善对话系统的鲁棒性，继而提升跨语言对话系统的性能。

2.提出了一种基于知识迁移的端到端的跨语言对话系统构建方法

       针对管道式跨语言对话系统存在的误差累积、响应延迟大和计算复杂度高等问题，提出了一种基于知识迁移的端到端的跨语言对话系统构建方法。该方法在不依赖目标语言任何数据资源的情况下，通过迁移资源丰富的源语言对话知识提升了目标语言对话系统的性能。具体地，所提方法通过对单语言对话、跨语言对话和机器翻译三类任务的联合学习，在翻译知识和源语言对话知识的指导下实现零资源目标语言的端到端的对话系统训练。实验表明，所提方法能够在不依赖任何目标语言对话数据的情况下，有效建立端到端的跨语言对话系统，相比于管道式跨语言对话系统具有模型参数小、响应速度快等优势。

3.提出了一种基于语义对齐的零样本跨语言对话系统扩展方法

       针对目前对话系统不能利用单一模型处理多种语言的问题，提出了一种基于语义对齐的零样本跨语言对话系统扩展方法。该方法在缺少目标语言对话数据的情况下，利用多语言预训练模型将单语言对话扩展为多语言对话。具体地，所提方法从跨语言对齐和对话任务学习两个角度出发建立了多项预训练辅助任务。其中，前向词汇翻译、逆向句子还原和跨语言语义相似度量任务旨在指导模型学习对齐的多语言语义表示；单语言和跨语言对话状态追踪则帮助模型学习对话相关的知识。实验表明，所提方法能够实现源语言单语对话模型的零样本跨语言扩展，使其兼具源语言和多种目标语言处理的能力。

      综上所述，本文针对低资源场景下跨语言对话系统的构建方法展开了深入研究，分别从对话系统的鲁棒性、跨语言对话系统的迁移和跨语言对话系统的扩展三个方面出发，提出了一系列为低资源语言构建对话系统的方法。最终实验证实了本文所提出的方法能够有效利用高资源语言的对话系统和相关的多语言资源快速构建跨语言对话系统。

With the rapid development and widespread application of mobile communication and network technology, the demand for intelligent customer service and intelligent assistants has greatly increased. As one of the most important technologies, task-oriented dialogue system is a hot research topic in natural language processing. Task-oriented dialogue system is designed to help users complete specific tasks within limited turns. It is widely used in virtual personal assistants, smart homes, intelligent customer service, and other fields.

With the accelerated development of globalization, cross-lingual communication has become more and more important. However, the existing task-oriented dialogue systems mainly use the data of a certain language for training, which face great challenges in multilingual transfer and extension. On the one hand, the dialogue training data of some languages is scarce, which cannot meet the data requirements of the current deep learning dialogue systems. On the other hand, if the machine translation systems are used as the bridge of cross-lingual dialogue systems, there are problems such as poor robustness, low response speed, and difficult deployment. To this end, this dissertation focuses on how to effectively utilize the dialogue system of a high-resource language and related multilingual resources to quickly realize the construction of dialogue systems in low-resource language scenarios. The main contributions of this dissertation are summarized as follows:

1. Robust Dialogue System Construction for Multi-granularity Translation Noise

The machine translation based cross-lingual dialogue system, which consists of translation step, dialogue step, and back-translation step, has the problem of poor robustness caused by cascading errors. To overcome the problem, a robust dialogue system construction method for multi-granularity translation noise is proposed. Firstly, based on multi-granularity noise in machine translation, this method constructs adversarial examples containing translation noise at the word, phrase, and sentence level. Then two different adversarial learning strategies are designed to enhance the robustness of the dialogue system. Specifically, the utterance-level adversarial learning adopts adversarial learning over multi-granularity adversarial examples and clean data, guiding the dialogue system to learn noise-independent hidden representations. The knowledge-level guided method explicitly adds multi-granularity noise into the noise knowledge base, thus enhancing the dialogue system's robustness. Experimental results show that the proposed method can significantly improve the robustness of the dialogue system, thus improving the performance of the cross-lingual dialogue system.

2. End-to-End Cross-Lingual Dialogue System Construction Based on Knowledge Transfer

The pipeline cross-lingual dialogue system suffers from error accumulation, large response delay, and high computational complexity. To address the problem, an end-to-end cross-lingual dialogue system construction method based on knowledge transfer is proposed. This method improves the performance of the target language dialogue system by transferring the dialogue knowledge in data-rich source language without relying on any data resources of the target language. Specifically, the method realizes the end-to-end dialogue system training of zero-resource target language under the guidance of translation knowledge and source language dialogue knowledge through the joint learning of three tasks: monolingual dialogue, cross-lingual dialogue, and machine translation. Experimental results show that the proposed method can effectively build an end-to-end cross-lingual dialogue system without relying on any dialogue data of the target language. Compared with the pipeline cross-lingual dialogue system, it has the advantages of smaller model parameters and faster response speed.

3. Zero-Shot Cross-Lingual Dialogue System Extension Based on Semantic Alignment

Towards the problem that the current dialogue system cannot handle multiple languages with a single model, a zero-shot cross-lingual dialogue system extension method based on semantic alignment is proposed. In this method, the multilingual pre-trained model is used to extend the monolingual dialogue into multilingual dialogue in the absence of target language dialogue data. Specifically, the proposed method establishes several pre-training auxiliary tasks from the perspectives of cross-lingual semantic alignment and dialogue task learning. Forward word translation, reverse sentence recovery, and cross-lingual semantic similarity measurement tasks are designed to guide the model to learn aligned multilingual semantic representations. Monolingual and cross-lingual dialogue state tracking help the model learn the knowledge related to dialogue. Experimental results show that the proposed method can achieve zero-shot cross-lingual extension of monolingual dialogue model in the source language, enabling it to process both source language and multiple target languages.

To sum up, this dissertation conducts in-depth research on the construction method of cross-lingual dialogue system in low-resource scenarios and puts forward a series of methods to build a dialogue system for low-resource languages from three aspects: robustness, transfer, and extension of cross-lingual dialogue system. Finally, the experimental results show that the proposed method can effectively use the dialogue system of high-resource languages and related multilingual resources to build a cross-lingual dialogue system quickly.