CASIA OpenIR  > 模式识别国家重点实验室  > 自然语言处理
任务型对话系统中口语理解模块方法研究
白赫
Subtype硕士
Thesis Advisor周玉
2019-05-24
Degree Grantor中国科学院自动化研究所
Place of Conferral中国科学院自动化研究所
Degree Name工学硕士
Degree Discipline模式识别与智能系统
Keyword对话系统 口语理解 跨语言迁移 强化学习 多任务学习
Abstract

任务型对话系统是在某一特定领域(导航、订餐馆等),通过自然语言交互 的形式,辅助用户完成任务的人机交互系统。任务型对话系统需要具有如下三个 主要模块:口语理解、对话管理和对话生成。其中,口语理解用于在对话限定的 领域中,理解用户话语的含义。它通常包括领域分类、用户意图检测和语义槽填 充三个子任务。领域分类和意图检测属于句子分类任务,语义槽填充属于序列标 注任务。近年来,在学术界评测与工业界需求的推动下,口语理解逐渐成为了一 个活跃的研究领域,并取得了相当的进展,但过去的研究主要关注单语言、单轮 解析的口语理解方法,关于多语言、多轮口语理解的研究还很少。因此,本文主 要以这两方面为研究重点,主要研究内容归纳如下:

(1) 一种基于源端评判增强学习的口语理解跨语言迁移方法

如果要使对话系统能针对不同国家支持不同的语言,可以为每一种语言单 独收集、标注大量的训练数据,但这极其的费时费力,限制了对话系统的可扩展 性。因此,将针对一种语言开发的口语理解系统迁移到其他语言,具有较大的应 用价值与研究意义。本文提出的方法借助机器翻译模型把源语言训练数据翻译 成目标语言,并针对标注信息在翻译过程中易丢失的问题,设计了语义槽保留率 作为奖励指标,并通过强化学习对该指标进行优化,进而实现用源语言标注数据 加上少量的双语数据,完成口语理解模块的跨语言迁移。实验表明,相比于传统 方法,该方法可以显著提升目标语言口语理解系统的性能,并在实验数据集上取 得了最好的效果。

(2) 一种基于对话逻辑推理进行记忆强化的多轮口语理解方法

目前的口语理解大多对单轮话语进行解析,但是对话通常是人机多轮交互。 已有一些研究表明,对话历史对当前用户话语的解析有重要作用,但这些研究关 于对话历史的建模还很初步,没有充分挖掘对话历史信息。本文首先定义了对话 逻辑推理这一辅助任务,然后通过将对话逻辑推理任务与多轮口语理解任务共 享记忆编码与记忆检索模块,在多任务学习框架下进行记忆强化,从而有效的提 升模型的记忆编码与检索能力,进而获得更好的语境知识,提升多轮口语理解的 性能。

Other Abstract

Task-oriented dialogue system is a human-machine interaction system to finish tasks assigned by user in specific domains, such as navigation and booking a restaurant. A task-oriented dialogue system usually consists of spoken language understand- ing (SLU), dialogue management (DM) and language generation (LG). And SLU is a key technique to parse user’s utterance into semantic form. A typical pipeline of SLU includes domain classification, intent detection and slot filling. Traditionally, domain classification and intent detection are treated as classification tasks. Slot filling task is usually treated as a sequence labeling task. In recent years, SLU has been an active research topic and great progress has been made. However, previous work mainly focuses on monolingual single turn SLU methods. In this thesis, we explore approaches to language transferring for SLU and multi-turn SLU. The main contents are summarized as follows:

(1) Source Critical Reinforcement Learning for Transferring Spoken Language Understanding to a New Language

To make dialogue systems support multiple languages over different markets, collecting and annotating a large SLU training corpus per language are tedious and costly, hindering the scalability of these systems. Thus, it would be greatly helpful if the efforts taken to develop one SLU system could be reused for other languages. In this paper, we focus on translating the training data from source language into target language and we show how to translate annotation information with reinforcement learning (RL) method, where translated sentences with more proper slot tags receive higher rewards. The experimental results show that our proposed method achieves the highest SLU performance among several baseline methods.

(2) Memory Consolidation for Contextual Spoken Language Understanding with Dialogue Logistic Inference

Traditional SLU focuses on single turn utterance parsing, while conversation between human and chatbot usually consists of multiple turns and dialogue contexts are proven helpful in the SLU system. However, most of the previous models learn the context memory with only one objective to maximizing the SLU performance, leaving the context memory under-exploited. In this paper, we propose a new dialogue logistic inference (DLI) task to consolidate the context memory jointly with SLU in the multi-task framework, by sharing the memory encoder and retrieval mechanism with the SLU model. Thus, the memory encoding and retrieval abilities are enhanced with our proposed method, contributing to a better context knowledge and SLU performance.

Pages59
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23866
Collection模式识别国家重点实验室_自然语言处理
Corresponding Author白赫
Recommended Citation
GB/T 7714
白赫. 任务型对话系统中口语理解模块方法研究[D]. 中国科学院自动化研究所. 中国科学院自动化研究所,2019.
Files in This Item:
File Name/Size DocType Version Access License
Thesis.pdf(2958KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[白赫]'s Articles
Baidu academic
Similar articles in Baidu academic
[白赫]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[白赫]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.