基于平行推理机制的隐式篇章关系检测研究

发布时间：2018-02-11 03:59

本文关键词： 隐式篇章关系平行推理检索驱动关系消歧双语协作　出处：《苏州大学》2013年硕士论文　论文类型：学位论文

【摘要】：篇章关系分析是自然语言处理的一个核心问题。在篇章研究领域中，篇章是指一系列连续子句、句子或语段构成的语言整体；而篇章关系是指同一篇章中，不同论元之间的语义逻辑关系（如因果关系、转折关系等）。其中，论元是一种具有完整语义并表述独立观点的文字片段。篇章关系分析任务即是对两个不同论元（也称一组论元对）之间的具体篇章关系类型进行识别与判定。根据论元之间是否存在显式连接词（也称作线索词，如“因为”、“但是”等），可将篇章关系分为显式关系和隐式关系，相应的论元即被称为显式论元与隐式论元。现有研究对显式关系分析已具有较好的推理效果，而对缺失连接词的隐式关系尚无法有效处理。本文针对两个毗邻且隐式相关的论元，提出了一种基于平行推理机制的隐式篇章关系检测方法。核心思想是根据“论元语义平行，则论元关系平行”的理论假设，即某显式论元对与隐式论元对在句子结构以及语义层面上均相近，也就是说语义平行，，则这两个论元对的语义连接关系应相同，并借助信息检索技术，从大规模数据集中挖掘平行的语义论元集合及其论元关系的显式线索，从而建立无指导的隐式篇章关系检测系统。主要研究内容包括以下三个方面： 1）检索驱动的隐式关系平行推理机制借助大规模Web信息，实现平行论元的自动挖掘和平行关系的自主推理。一方面，在挖掘平行论元过程中，重点依赖信息检索技术，通过构建高质量查询关键词，从搜索引擎中挖掘语义表述相近的平行论元及线索；另一方面，构建三种篇章关系推理模型，从各方面对挖掘出的平行论元及语义线索进行质量评估，最终选择质量最优的N个平行论元，借助高质量平行论元中语义线索与篇章关系的映射体系实现隐式篇章关系推理。 2）平行推理中歧义线索消歧在基于平行推理机制实现隐式篇章关系检测任务中，推理线索存在两方面的歧义：（1）线索连接词本身的一词多义现象；（2）混淆连接关系的伪线索词存在。两类歧义都将影响最终隐式篇章关系检测性能。因此，本文从基于局部最优的连接词词义消歧以及基于潜在连接词识别的伪连接词过滤两个方面，对平行推理过程中的歧义线索进行消歧，以提高隐式篇章关系检测性能。 3）基于双语协作的平行推理优化在检索驱动挖掘平行论元及语义线索的过程中，由于部分隐式论元的词汇信息有限或搜索引擎检索性能限制，导致挖掘的平行论元数量少甚至为零，不利于平行论元与隐式论元之间平行关系的正确推理。因此，本文尝试通过跨语言信息检索技术，从搜索引擎中挖掘另一种语言的平行论元及推理线索，以此来扩充平行资源，使得平行关系推理依据更充分更丰富，从而进一步优化隐式篇章关系检测。
[Abstract]:Text relation analysis is a core problem of natural language processing . In the field of chapter research , the text refers to a series of continuous clauses , sentences or language whole ; and the chapter relation refers to the semantic logical relation ( such as causality , conversion relationship , etc . ) between different arguments in the same chapter . On the basis of whether there is explicit connection between elements ( also known as clue words , such as " because " , " but " , etc . ) , it can be divided into explicit relation and implicit relationship . In this paper , we propose an implicit text relationship detection method based on parallel inference mechanism for two adjacent and implicit theory elements . The core idea is that the semantic connection relation of the two theory elements should be the same , that is , the semantic parallelism is similar to that of the implicit theory element in the structure of sentence structure and the semantic level , that is , the semantic parallelism , then the explicit clues of the parallel semantic theory element set and its theory element relation are excavated from the large - scale data set , and the hidden chapter relation detection system without guidance is established . 1 ) Parallel Inference Mechanism of Implicit Relations Driven by Retrieval On the one hand , based on the large - scale Web information , the autonomous reasoning of automatic mining and parallel relation of the parallel theory elements is realized . On the one hand , in the process of mining the parallel theory , the emphases are on the information retrieval technology . By constructing the high - quality query key words , the parallel theory elements and the clues that are close to the semantic expressions are excavated from the search engine . On the other hand , three chapters are constructed , and N parallel theory elements with optimal quality are constructed , and the implicit chapter relationship reasoning is realized by the mapping system of the semantic clues and the text relation in the high - quality parallel argument . 2 ) Disambiguity in Parallel Reasoning In the task of realizing implicit text relation detection based on parallel inference mechanism , two aspects of reasoning lead exist : ( 1 ) The word ambiguity of the thread connection word itself ; ( 2 ) the existence of a pseudo - clue word which confuses the connection relation . Both types of ambiguity will affect the performance of the final implicit text relationship detection . Therefore , this paper attempts to disprove the ambiguous clue in the process of parallel inference based on the local optimal connectives and the false connection word filtering based on the potential connection word recognition , so as to improve the detection performance of the implicit text relationship . 3 ) Parallel reasoning optimization based on bilingual collaboration In the process of retrieving and driving parallel argument and semantic clue , due to the limited vocabulary information of partial implicit theory or the search performance limitation of search engine , this paper attempts to expand the parallel resources by using the technique of cross - language information retrieval , mining the parallel theory element and reasoning clue of another language from the search engine , so that the parallel relationship reasoning is more abundant and richer , thus further optimizing the implicit text relation detection .

【学位授予单位】：苏州大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.1

【共引文献】