基于Restriction-Centered Theory的自然语言模糊语义理论研究及应用

Human knowledge and natural language across a wide of array of domains have a big deal of imprecision and fuzziness. Therefore, as typical ill-posed problem, from theoretical, methodological and applicative aspects, researches on exploring the mathematical formalization and precisiation of semantics based on Restriction-Centered Theory on sentence level have been increasingly involved in Natural Language Understanding (NLU). Particularly, the elasticity of meaning cannot be ignored. To bridge the gap between human-understanding and machine-readable, especially, on common sense of language expression that almost imply fuzzy semantics, this project will focus on the following concrete points. i.to distinguish and clarify the differences between fuzziness and polysemy, ambiguity, metaphor both from lexical and pragmatics perspectives; ii.to study formalization and quantitative analysis of fuzzy semantics in Chinese and English phenomena under the machine-computable scheme in specific domains, namely Computing With Words or Computation with Natural Language problems; iii.to probe into the elasticity of semantics amid daily reasoning language expressions, then put forward the method of precisiation of fuzziness meanings through treating such expressions as fuzzy propositions with linguistic variables and generalized semantic constraints based on Restriction-Centered Theory; iv.to apply the expected studying results of this project to information retrieval with linguistic hedges which always play a role of syntactical function in sentence, and canonical forms and descriptions for construction of common sense knowledge which primarily impact the performance of automatic question and answering system.Furthermore, to support the deeply parsing logic form of natural language. This project will expand typical fields on fuzzy semantics and logic semantics for Chinese community, enrich academic disciplines in NLU, and promote intellectualization applications in above realms. The main tenor does provide more or less positive effect for those aspiring to contribute to semantic computing in the area of through the lenses of rationalism. Hence,it has the highly theoretical significances, practical prospects, and academic challenges.

人类知识和自然语言普遍存在不准确与模糊性。探索作为不适定问题的自然语言处理中，句子级语义的抽象表示与精确化是当前语义计算的难题，其中弹性语义的精准表达与计算分析不容被忽视。为了跨越人机在蕴含模糊语义的常识类语言命题理解上的鸿沟，本项目基于Restriction-Centered Theory（RCT）深入开展自然语言模糊语义的理论研究与应用验证。具体包括模糊性与歧义、多义及比喻手法的厘清与界定；在特定域开展计算机可计算视角下的、以汉语、英语为主的自然语言模糊性语义量化对比研究，基于RCT探索并提出符合汉语特点的包括弹性语义在内的模糊语义精确化表达与语义计算分析方法，并在信息检索、自动问答常识知识推理等应用中进行验证，为更深层次自然语言逻辑形式分析提供直接支持。本项目所提出的分析理论、研究方法及工程应用解决方案将进一步扩展模糊语义学和模糊逻辑的经典框架，具有很高的学术价值和实用意义与挑战性。

机器理解自然语言的能力受限于语义自动分析处理的水平。语义表达与模糊语义的精确计算是目前的难点问题，探索真实语言工程应用中弹性语义空间下句子级语义计算与语义模糊性有效度量具有重要的研究价值与实用意义。本项目基于中心限制理论（Restriction-Centered Theory）深入开展了自然语言模糊语义的理论研究、关键技术探索及多应用场景验证。.在自然语言（包含语义在内）自身特征挖掘方面，率先设计实现了藏语短语结构树-依存树自动转换算法，首次提出了自底向上的多组块粒度藏语短语树-依存树转换方法；创新性地提出了融合越南语音调、发音相关性特征的汉越SMT模型等；首次提出了基于众包模式构建蒙语口语语料库的解决方案，相关成果在提升稀缺资源型语言智能处理系统性能、改善研究资源生态方面发挥了积极作用。.特定问题域抽象语义表示与精确化方面，本项目基于中心限制理论探索了不同场景的弹性语义空间中解释性集合ED的产生模式及语义关系表示。提出了结合外部情感知识的情绪原因识别模型、基于富知识的隐式情绪及原因联合抽取模型；面向多文档阅读理解的粗-细粒度文档排序方法及答案补全策略；多维度群体情绪分析及可视化；基于加权聚合器的开放知识图谱补全等模型与方法，通过语义约束关系量化句子级模糊语义，成功实现了多个领域的应用验证。.在核心应用场景机器翻译领域，提出了融合目标端语义结构的串-树模型、融合词性特征的平行RNN语言模型、融合隐式树的神经网络翻译模型、基于Pre-training的多策略神经翻译模型、基于组块的汉法神经翻译模型及面向低资源神经网络机器翻译的数据增强方法等一系列关键技术，显著提升了相关领域的翻译系统性能。.本项目发表论文21篇，出版专著2部；申请发明专利8项（授权2项），获授权软件著作权1项；获省部级科技进步一等奖1项，先后培养研究生29名。部分成果已在国家安全部门及解放军某部部署应用。