| 作为天然说话处置的一个研究重点,语义剖析旨在将人类的天然说话转化为盘算性能够懂得的情势化说话。因为深层语义剖析的庞杂性,人们今朝更关怀浅层语义剖析,即剖析句子中谓词(可所以动词或名词等)的语义脚色成份,包含施事者、受事者、时光、所在等。作为浅层语义剖析的一种完成措施,语义脚色标注(Semantic Role Labeling,简称SRL)已被普遍应用于天然说话处置相干义务,如信息抽取、问答体系和机械翻译等。依据谓词词性的分歧,平日可以将SRL分为动词性谓词SRL和名词性谓词SRL。今朝主流的SRL研究集中于在给定句法树的前提下,运用各类统计机械进修技巧,采取基于特点向量或基于树核函数的办法,停止语义脚色的辨认和分类。最近几年来的研究注解,SRL的机能严重依附于句法剖析的机能,同时名词性谓词SRL机能远低于动词性谓词SRL机能。以上两个成绩在中文SRL研究中特别凸起,例如,在中文PropBank和中文NomBank的相干试验注解,基于准确句法树和准确谓词,动(名)词性谓词SRL机能F1值可以到达92(70),而基于主动句法剖析F1值降低为67(57)。本文以句法和语义的结合剖析为研究目的,研究新鲜的句法剖析模子和动/名词性谓词SRL,偏重点摸索二者之间的结合进修机制,推动SRL的适用化过程。重要研究内容包含1.句法剖析的研究。提出了条理句法剖析模子,为完成句法剖析和SRL的结合进修供给了强无力的基本。该模子将句法剖析分化为三个子义务词性标注、根本短语辨认和庞杂短语辨认,自底向长进行,其根本思惟是在每层处置进程中,优先辨认出轻易辨认的组块,如许就可以供给更丰硕的高低文信息停止庞杂组块辨认;未被归并的组块和新辨认发生的组块配合组成下步处置的输出,反复此进程直至辨认出根结点。2. SRL的研究。起首体系研究了中文动词性谓词SRL,重点摸索了若何从句法树中抽掏出各类立体特点和构造化特点。其次,深刻研究了中文名词性谓词SRL,从两个角度摸索了中文动词性谓词SRL对中文名词性谓词SRL的作用练习实例的扩大和动词性谓词SRL特点的应用,明显地进步了名词性谓词SRL机能。最初,研究了中文名词性谓词的主动辨认成绩。试验注解,本文获得的动(名)词谓词SRL机能优于其他同类型体系。3.句法剖析和SRL的结合进修机制研究。重要从两个条理摸索了句法剖析和SRL的结合进修第一,提出了一种结合进修计划,将SRL嵌入到基于条理句法剖析模子的句法剖析进程中,完成二者的同步履行;第二,将由SRL获得的语义信息集成到条理句法剖析模子中,更好地指点句法剖析。试验注解,该结合进修计划不只减缓了SRL对句法剖析成果的严重依附,并且可以或许进步二者的机能,特殊是SRL的机能。本文的立异点重要表示在提出了条理句法剖析模子,该模子不只获得较好的机能,并且具有优越的可扩大性,可以或许有用集成其他天然说话处置义务;提出了运用动词性谓词SRL生成的有用特点来帮助名词性谓词SRL;提出了一种有用的句法剖析和SRL的结合进修机制,削减SRL对句法剖析的依附。试验注解,这些研究年夜年夜进步了SRL的机能,加重了SRL对句法剖析的依附,对往后SRL的研究具有主要的参考价值。 Abstract: As a research focus of natural language processing, semantic analysis is aimed at transforming the nature of human beings into a situation that can be understood by computer. Because of deep semantic analysis of complex, people today more caring shallow semantic analysis, analysis of predicate in a sentence (so a verb or a noun semantic role ingredients, contains the agent, patient, time, location. As a shallow semantic analysis of a complete method and semantic role annotation semantic role labeling (SRL) has been widely used in natural language disposal relevant obligations, such as information extraction, question and answer system and machine translation and so on. According to the differences of the predicate part of speech, the SRL can be divided into verb predicate SRL and nominal predicate SRL. The current mainstream of SRL research concentrated in under the premise of a given syntactic tree, is widely used in various types of statistical machine learning skills and take feature vector based or tree kernel based approach to stop the semantic role identification and classification. In recent years, the function of SRL is heavily dependent on the function of syntax analysis, while the function of SRL is far lower than that of SRL. These two results are particularly salient in the Chinese SRL research. For example, the PropBank and NomBank of the Chinese are related to the test. Based on the exact syntax tree and the exact predicate, the F1 function SRL of the verb (name) can reach 92 (70), while the F1 value of the active parsing is reduced to 67 (57). In this paper, the combination of syntactic and semantic analysis for the purpose of the study, the new syntax analysis of the model and the dynamic / nominal predicate SRL, a combination of learning mechanism between the two points, to promote the application of SRL process. Important research content including 1 syntax analysis. The model of syntactic parsing is put forward, which provides the basic of the analysis of the combination of syntax and SRL. The mold will be parsing differentiate into three sub obligations of part of speech tagging, identify the fundamental phrase identification and complex phrases, from the bottom to the length, the fundamental thought is in each layer of the disposal process, give priority to identify easily recognizable chunks, such can provide more substantial level of the information stop complex block identification; not to be merged group block and newly identified the formation of block with the composition of the output of the disposal of the next step, repeated this process until identified the root node. 2 SRL research. First of all, the system of the Chinese verb predicate SRL, the focus of the study of how to draw out all kinds of three-dimensional features and structural characteristics. Secondly, it discusses the Chinese nominal predicate SRL, from the two perspectives, the Chinese verb predicate SRL is explored. The effect of SRL on the SRL of the nominal predicate SRL is used, and the function of the nominal predicate is obviously improved. Initially, the active recognition of Chinese nominal predicates is studied. Test notes, this paper obtains the dynamic (name) word predicate SRL function better than other same type system. 3 syntactic analysis and the study of the mechanism of SRL. Important from both structured and explore the syntactic parsing and SRL combine the advanced study first, presents a combination of study plan, SRL is embedded into based on structured syntactic parsing model of syntactic analysis process, the completion of the two synchronization performance; second, obtained by the SRL semantic information integrated into structured syntactic analysis model, better advice parsing. Test notes, the combination of learning programs not only slowed down the SRL syntax analysis of the results of a serious attachment, and can progress the function of the two, the special function of SRL. The innovation points of this thesis are important said in the structured syntactic analysis model, this model not only get the better performance, and has excellent scalability, may be useful to integrate with other natural language disposal obligations; the verb of SRL generated useful features to help the nominal SRL; put forward a useful syntactic parsing and SRL binding mechanism for further study and cut SRL of the reliance on syntactic analysis. Test notes, these studies have greatly improved the performance of the SRL, and increased the dependency of SRL on syntax analysis, and it has a great reference value for the future research of SRL. 目录: |
