【作者】 范冬梅；

【导师】 张汝波；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2011， 博士

【摘要】 如何解决语言的歧义问题一直困扰着自然语言处理技术的研究人员。语言的歧义最重要的一个表现就是一词多义现象。如何根据歧义词出现的上下文语言环境识别出正确词义是词义识别要解决问题。在自然语言理解领域,词义识别是应用基础研究课题,也是自然语言理解的重点和难点之一。早期的词义识别研究以基于规则的方法为主,近些年来随着计算技术和存储技术的改进和提高,统计学习方法越来越受到广泛的欢迎,迅速成为了主流的研究方法。有监督的学习方法应用于词义识别可以获得较高的识别精度,但是该类方法需要有规模足够大的训练样本,这样的样本不是容易获得。无监督的方法不需要人工标注训练样本,但是词义识别相对效果不是十分理想。本文分析了统计词义识别研究需要解决的几个关键性问题。从词典资源和语料库资源的建设到词义识别问题的建模方法,以及词义分类问题的特征选择,对所涉及的问题都一一进行了讨论。在这些基础上,本文最后给出了词义类扩展思想,并探讨了如何将其应用于统计词义识别的研究上。本文研究成果和创新如下：1.从词义的可计算性角度探讨词义刻画与词义识别的关系,探讨如何通过科学控制词义刻画粒度来重新整合现有的词典资源,建设新的机读词典,更好地为词义识别的应用服务。实验分析说明词义刻画粒度的大小直接影响了词义识别的精度,适当控制词义粒度再不产生二义性的前提下可以提高词义识别的精度。本文提出在词义再刻画的基础上整合现有词典资源,建设面向词义计算的新的分类词典；2.在探讨面向词义识别的特征选择方法的基础上提出以信息增益改进贝叶斯模型的词义识别新方法。实验中以朴素贝叶斯模型、最大熵方法和支持向量机建立的词义分类器作为参照模型,讨论信息增益改进贝叶斯模型的效果。实验结果显示参照系中使用最大熵和支持向量机构建的词义分类器都比朴素贝叶斯模型强,其中基于支持向量机的词义分类器最好,而经过信息增益改进的贝叶斯模型在词义识别上表现更突出,实验结果比SVM还要高出1.4个百分点,获得了对比实验中最优的识别结果；3.从语料库资源难于大规模建设的角度出发,实验分析和探讨了人造歧义词技术的使用问题,并在此基础上提出替换词的概念,以及基于替换词技术的词义识别新方法。实验结果表明,人造歧义词技术可以帮助研究者缓解训练语料短缺的压力,源于此的替换词技术可以让研究者避开人工标注训练样本,实现一种无监督的词义识别方法。实验结果表明基于替换词技术的词义识别方法具有较高的识别精度；4.针对词义识别训练语料规模不足够大的问题,提出了词义类扩展思想和基于词义类扩展的词义识别新方法。该方法通过词义类扩展,可以在有限训练语料中获得更多的词义信息,提高训练的效率,改善词义识别的效果,此外词义类扩展技术可以从无词义标注(无词义标记等先验知识)的生语料中统计相关词语信息,以此来为小规模的训练样本提供补充。实验结果表明基于词义类扩展思想的方法提高了训练语料的使用效率,改善了有监督词义识别的效果,这为增强小规模训练样本的统计学习效果提供一个崭新的思路。综上所述,本文在资源建设、词义识别的建模、特征选择,以及如何突破训练语料规模的限制实现无监督词义识别和改善有监督的词义识别方法上都作了一些有益的尝试,取得了一些初步成果。随着词义识别研究的不断深入,会涌现出更多更好的新的解决思路和方案。更多还原

【Abstract】 How to solve the language ambiguity problem is the issue always plaguing the language processing technology researchers. Polysemous situation is thought to be the most important performance of language ambiguity. The problem solved by word sense discrimination (WSD) is how to identify the right meaning of the word by the context of language environment of the ambiguous words. In the field of natural language understanding, word recognition is the application of basic research issue, and also, one of the most important and difficult things of natural language understanding.Traditional word recognition is mostly studied by rule-based method, but in recent years, with the improvement and enhancement of computing and storage technology, statistical learning method is becoming more and more popular, and quickly became the mainstream research methods of word recognition. We can get a high identification accuracy when we apply the study method with supervision to word recognition, but such methods require a large enough scale of training sample which is not easy to gain. Although unsupervised method does not need training samples with manual tag, the relative effect of word sense discrimination is not very good.Some key issues which should be solved by the statistical word sense discrimination were analyzed in this paper. The questions involved were discussed one by one, not only the construction of dictionary and corpus resources, the modeling method of the problem of word recognition, but also the feature selection of the semantic classification. Basing on these issues, we got the thought of word-sense category extending at the end of this paper, and also, we discussed how to apply it in the research of the statistical word sense discrimination.This research and innovation are as follows:1. The relationship between word sense discrimination and word sense characterization were discussed from the perspective of semantic computability, and it was also studied how to make re-integration of existing dictionary resources and to construct a new machine readable dictionary by the scientific control of the semantic granularity, providing better service for word sense discrimination. Experimental results showed that the size of semantic granularity in the semantic characterization directly affected the accuracy of word sense discrimination. The accuracy of WSD could be increased by proper control the semantic granularity without ambiguity. It was proposed to integrate the existing dictionary resources, to build a new category dictionary for word sense discrimination;2. The improved Bayesian model by information gain for the feature selection of word sense discrimination was proposed. The word sense classifiers established by Naive Bayes model, Maximum Entropy method and Support Vector Machine were used as reference models in experiments which the effectiveness of Bayesian model improved by information gain. The results showed that the classifiers constructed by Maximum Entropy and Support Vector Machine are stronger than the Naive Bayes model, in which Support Vector Machine is the best in several of reference models. But the Bayesian model improved by information gain in word sense discrimination is more prominent comparing with the reference models, and its experimental results were much higher 1.4 percentage points than the SVM classifier, the Bayesian model improved by information gain obtained the best results in comparative experiments;3. The use of artificial ambiguous word technology was analyzed and discussed by experiments form the perspective of the difficulty in the construction of large-scale corpus resources, and the concept of vicarious words and the new method of word sense discrimination based on vicarious words were proposed. The results showed artificial ambiguous words technology could help researchers to relieve the pressure of the shortage of training data. The vicarious words technology from artificial ambiguous word allows researchers to achieve an unsupervised method for WSD avoiding the use of the training samples with manual tagging. Experimental results showed the WSD method based on vicarious words had high discrimination accuracy;4. The ideas of Word-sense Class Extending (WSE) and a new word sense discrimination method based on WSE were proposed for the problem which the size of training corpus was not sufficiently large. This new method could obtain more semantic information in limited training corpus to enhance training efficiency and improve the effect of word sense discrimination by WSE, in addition, the WSE technology could statistics related word s information in raw corpus (no prior knowledge of semantic tags, etc.) in order to provide additional training samples. Experimental results showed that the word sense discrimination method based on WSE improved the efficiency of the training corpus, made better the effectiveness of WSD. The WSE technology provided a new idea to enhance the effectiveness of statistical learning in small-scale training corpus.To sum up, this article had give some useful attempts in resource-building, word sense discrimination modeling, feature Selection, as well as on how to improve word sense discrimination with supervision, and we had achieved some initial results. With the further research of word sense discrimination, more and more new ideas and solutions will be emerged.更多还原