题名汉语口语系统的鲁棒性语义分析器研究
作者刘蓓
学位类别博士
答辩日期2004
授予单位中国科学院声学研究所
授予地点中国科学院声学研究所
关键词口语对话系统 自然语言理解 语义分析 知识表示 鲁棒性
中文摘要我们开展汉语口语对话系统研究的目标是:构建以人机语音交互为基础面向实际应用的智能化人机接口系统,同时建立通用的系统开发方法和应用的工程化规范。在系统构建方面,最重要的是提高系统各功能模块的通用性以提高口语对话系统的可移植性;在提高对话成功率方面,最重要的是提高语义分析器对于灵活多变的用户口语输入的适应能力。因此本文针对面向任务口语对话系统的特点,围绕提高汉语口语对话系统中语义分析鲁棒性这个核心问题展开研究,主要完成了以下工作:1.建立基于语义概念层次模型(SCHM)的同构知识库,提出完整的口语对话行为表达方法本文针对系统各模块知识处理的特点,提出建立基于概念层次模型(SCHM)的同构知识库的方法,利用SC服构建用于分析表层语义关系的语义框架网络,用于分析深层语义格角色的语义格框架,用于语言生成的框架规则,以及用于实现事务处理自动推理结点格框架。同构知识库的建立,实现了口语对话系统中各模块语言处理方法的统一化及标准化。基于口语对话行为是一种包含语义及语用双重功能的信息结构的认识,本文提出了一种完整的口语对话行为表示方法。该方法为融合了C_S(Class_Subclass)语用表示方法以及面向任务语义信息表示方法的用户对话行为表达方法C_S_F(Class_Subclass_Frame)。C_S_FF方法与系统对话行为表示C_S_L(Class_Subclass_Logic)方法一起,构成了完整的口语对话行为表达体系,并进一步提出了从C_S_F到C_S_L的映射规则。2.设计并实现了面向任务口语对话系统的期待模型(EM),完善了口语对话的语境处理期待模型是语境的重要构成内容之一,本文根据任务结构与对话结构的内在联系提出了面向任务口语对话系统中期待模型的构建方法,以及期待队列的产生及优先级排序算法。其优势是可以随对话过程的进行,对下回合的用户对话行为产生合理的预测,并与对话历史记录相结合,构造出动态变化的语境,使系统具备利用语境推导用户意向,及修复用户对话行为的推理功能,从而大大提高语义分析的鲁棒性及对话成功率。该期待模型还可用于对话过程中的话题转移检测,方便系统决策.实验结果表明,加入期待模型后BEST系统的性能指标有了较大幅度的改善,其中最显著的是句子分析准确率,从68.98%提高到85.68%,改善率达24.21%;相应地,事务处理成功率也得到明显提高,从78.24%提高到92.15%,改善率达17.78%。去除不良数据的影响,赋值域内的统计数据表明系统的性能提高更加明显,句子分析正确率从75.76%提高到88.34%。3.提出语音语义分析与关系结构消歧相结合的鲁棒性语义分析方法首先,为了提高语义分析器的鲁棒性,本文用拼音格替代词图或汉语句子作为语义分析器的输入,最大程度地保留用户输入中的有效信息,减少识别错误的传递。然后,在语义分析预处理过程中,直接从拼音格中提取基本语义概念,并通过计算概念分析路径的分值来确定候选路径,充分发挥了语义框架语言模型对识别过程的指导和约束作用。在航班及铁路两个任务领域内进行语音输入模态的口语对话测试,从测试结果可以看出,该语音概念分析模式可有效降低识别错误对语义分析性能的不良影响,概念理解效率分别达到803%和81.1%。最后,针对语义分析时常见的两种歧义类型一关系歧义及结构歧义,本文分别提出了基于期待模型的关系歧义消歧策略及基于语义PcFG口robabilistic Conte沉沈-FreeGr~aro模型的结构歧义消歧策略。实验结果表明,在基线系统BEST(Beijing Railway Station Ticket Information SDS)实验平台上综合应用上述两种消歧策略后产生了良好的效果。特别是句子分析准确率有了大幅度的提高,从原来的75.7%上升到92.8%。标志语义单元理解率的三项指标,正确率,召回率和精度也平均提高了10%。4.提出了基于语义格框架的自动推理模型,提高了事务处理模块的通用性本文针对系统任务的共同特点建立了通用的面向任务的事务处理框架。并根据系统语义表示与数据库领域知识表示之间的内在联系,通过建立结点格框架来表示语义格与数据库字段之间的映射及推理关系,从而建立起事务处理过程的自动推理模型。并在此基础上提出了自动推理算法,使整个事务处理过程成为一个具有计划能力并且可进行自动推理的通用过程。5.基于口语对话系统开发平台,设计并实现了首都机场航班信息口语对话系统(C AFI)本文给出了通用的面向任务的口语对话系统的体系结构,并通过在口语对话系统平台上建立CAFI系统,进一步验证了本文研究成果的通用性及可移植性。最后通过cAFI与BEsT系统的融合实验,给出了实现领域融合的解决方法。
英文摘要The aim of making research on Chinese Spoken Dialogue Systems (SDS's) is to build intelligent systems integrating modern speech and language processing technology to solve specific problems in task-oriented area by man-machine interactions, and to realize harmonious man-machine interfaces by engineering development methods. In the aspect of system architecture, it is most important to improve the independences between modules to convenience system transplanting. In order to improve the dialogue success rate, it is most important to enhance the robustness of semantic analysis on flexible user utterances. Therefore, aiming at the characteristics of task-oriented SDS's, this thesis's research surrounds improving the robustness of semantic analysis. And the main contributions of this thesis are: 1. Building the homogeneous knowledge base based on the Semantic Concepts Hierarchical Model (SCHM), and putting forth complete representation form of dialogue acts Basing on SCHM, we can set up the semaitic frame network for analyzing surface layer semantic relations, the semantic case frames for analyzing deep level semantic case roles, the frame rules for language generation, and the node case frames for transactions auto reasoning. The construction of homogeneous knowledge base realizes the unification and standardization of language processing methods for the system modules in SDS 's. Based on the notion that Dialogue act should be the information structure conveying both semantic and pragmatic information, the thesis puts forth the complete spoken dialogue act presentations C_S_F (Class_Subclass_Frame). C_S_F integrates C_S (Class_Subclass) pragmatic representation and task-oriented semantic information representation. C_S_F method together with C_S_L (Class_Subclass_Logic) method constitutes the complete representation system of dialogue acts. Then we designed the mapping rules for C_S_F to C_S_L transformation. 2. Designing and realizing the Expectation Model (EM) for task-oriented SDS's, and perfecting the Dialogue Context processing technique EM is one of the most important components of Dialogue Context. According to the inter-relations between task-structure and dialogue-structure, we puts forth the method of constructing EM in SDS's, also the expectation generation and ranking algorithms. EM can generate reasonable expectations of the user dialogue acts in the next turn, so as to construct dynamic context to endow the system the ability to reason user intentions and restore ill-formed user dialogue acts. Therefore, EM can help improve the robustness of semantic analysis and the dialogue success rate. Besides, EM can be used to detect topic transition during dialogues, which helps system make decisions. The experimental results show that when integrated with EM, the performance of BEST is greatly improved, in which the Sentence Accuracy (SA) is improved from 68.98% to 85.68%, the improve rate reaches 24.21%, so does the Task Success Rate (TS) which is improved from 78.24% to 92.15, by 17.78%. And when wiping off the influence of out-of-domain data, the within-domain statistics show even more evident improvement, in which the SA is improved from 75.76% to 88.34%. 3. Putting forth robust semantic analysis method which includes Speech Semantic Analysis and Relation Structural Disambiguation Strategies. To improve the robustness of the semantic analysis on low confidence speech recognition results, we replace the words graph or Chinese sentence input with the Pinyin Lattice input that can best reflect the character of real speech. Because Pinyin Lattice can effectively retain the valid information contained in user inputs, at the same time, restrain the transfer of recognition errors.
语种中文
公开日期2011-05-07
页码119
内容类型学位论文
源URL[http://159.226.59.140/handle/311008/814]  
专题声学研究所_声学所博硕士学位论文_1981-2009博硕士学位论文
推荐引用方式
GB/T 7714
刘蓓. 汉语口语系统的鲁棒性语义分析器研究[D]. 中国科学院声学研究所. 中国科学院声学研究所. 2004.
个性服务
查看访问统计
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。


©版权所有 ©2017 CSpace - Powered by CSpace