项目名称: 基于数据驱动的中文自然语言生成关键技术研究
项目编号: No.61202248
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 计算机科学学科
项目作者: 袁彩霞
作者单位: 北京邮电大学
项目金额: 25万元
中文摘要: 如何让计算机生成人类可以理解的语言,是一个重要的科学问题,同时也是实现人机自然交互、机器翻译、文本摘要等任务的重要技术手段。本课题以中文自然语言生成为具体任务,探索语言理解过程中的形式化理论及统计学习方法在语言生成中的应用,为提升现有语言生成技术的可重用性及鲁棒性、降低系统开发代价寻求有效途径。研究内容包括:1.基于上下文无关语法的生成空间描述,借鉴语言理解中的句法分析方法,自动构建以语言生成为目的的概念短语层级树;2.基于噪声信道模型的生成决策规划,进行深层结构及表层结构的统一实现,充分利用基于知识驱动的语言模型及基于动态规划的解码算法的概率特性及领域无关特性,提高系统的可重用性及鲁棒性;3.针对中文自然语言生成系统的评测数据及评测技术研究,研制并开放一套用于汉语语言生成技术评测的标准数据及基础工具,为汉语语言生成技术提供科学一致的评测方法。
中文关键词: 自然语言生成;概率上下文无关文法;决策森林;解码算法;
英文摘要: Making computers learn to generate natural language is a critical scientific problem, and also a key technique for solving tasks like man-machine dialogue, machine translation and text summarization. The aim of this project is to develop techniques for generating Chinese natural language. By utilizing the language formalization theory and statistical analysis method in natural language understanding, we provide an effective way reducing deveplopment cost of the current NLG system, meanwhile improving its scalability and robustness. The research content includes: 1. Towards domain independent NLG representation based on context-free grammar (CFG), starting from a straightforward CFG parsing result, automatically build the concept-phrase hierarchical tree for natural language generation engine, via which explicitly describe the generation space. 2. Towards noise-channel model for generation strategy planning, carry out deep and surface realization in a unified way, and enhance its technologies in terms of reusability, scalability and robustness by exploring the statistical advantages of knowledge-driven language model and dynamic decoding algorithm. 3. Towards evaluation techniques and data development for Chinese language generation, develop a suite of shared data and text processing tools, and thereby propose a
英文关键词: Natural language generation;probabilistic context-free grammar;decision forest;decoding algorithmn;