Robust, Incremental Parsing and Disambiguation for a Dialog Agent

本研究的目标是实现对有目的的口语对话的鲁棒解析和消歧，其中普通短语和句子结构可能被重写、自编辑、校正和澄清子对话破坏，实现鲁棒性和在线消歧的主要手段是（1）增量式“对话图解析器”从词和韵律层次到对话片段结构层次，全面整合各个层次的分析;以及（2）用于在线选择优选分析的“偏好权衡”技术，使用词汇、句法、语义语篇结构偏好被编码为数字势。将使用可用的文本语料库和在罗切斯特大学收集的TRAINS规划对话语料库对潜力进行统计“调整”。研究的意义在于，同时处理短语结构和对话结构，以及应对猖獗的歧义，是自然人机交流仍然面临的两个最重要的挑战。健壮的、偏好搜索对话解析器的出现将为各种智能代理中真正的、实用的对话理解提供基础，这些智能代理在规划、解决问题、危机管理和信息检索方面提供交互式帮助

The goal of this research is to achieve robust parsing and disambiguation of purposeful, spoken dialogs, where ordinary phrase and sentence structure may be disrupted by acknowledgements, self- editing, correction and clarification subdialogs, etc. The primary means for achieving robustness and on-line disambiguation will be (1) an incremental `dialog chart parser` that fully integrates all levels of analysis from the level of words and prosody to the level of dialog segment structure; and (2) a `preference trade-off` technique for selecting preferred analyses on-line, using lexical, syntatic, semantic, and discourse structure preferences encoded as numerical potentials. Potentials will be statistically `tuned` using available text corpora and the TRAINS planning dialog corpus collected at the University of Rochester. The significance of the research lies in the fact that dealing simultaneously with phrase structure and dialog structure, and coping with rampant ambiguity, are two of the most important challenges still facing natural human-machine communication. The advent of robust, preference-seekin g dialog parsers will provide a basis for genuine, practical dialog understanding in a wide variety of intelligent agents offering interactive assistance in planning, problem solving, crisis management, and information retrieval