| With the rapid development of the Internet, people are in contact with theexplosive growth trend of information resource, at the same time, the way in whichpeople get information also presents the diversified development trend. How tocomprehensively, accurately extract the information they need from these intricateinformation, to help users to collect their interested information become a majorresearch focus in the current information technology.Data mining as a processing technique reveals patterns and relationships betweendata. It excavates internal new relationships or potential trend from the analysis ofmassive, complex data. And it provides a good solution for data extraction andknowledge discovery. Text classification as a common hot research direction in datamining labels the unlabeled text using trained classifier according to certain rules. It iswidely used in Artificial Intelligence, Natural Language Processing.Before the ninety’s of the twentieth Century, text classification is a purely manualoperation. Artificial methods require staff to have the appropriate expertise in theirrespective areas, and to classify document category according to the individual’sknowledge. This classification method is not only time consuming, but also a greatwaste of human resources. Due to subjective reasons, different people may be indifferent views on the same text, and inconsistency may be caused. So the artificialmethod is difficult to maintain the accuracy of the classification. With thecontinuously explore of machine learning, Machine learning gradually applied to textcategorization and get great development. This classification method preprocess forthe text to be classified, including removing stop words, root reduction and removingsome Webpage marker in Webpage. Through the processing of this step, text elementwhich will not work, or play a small role in the classification are removed. We usefeature extraction algorithm to extract the theme can represent document. We establish text representation model to train classifiers and use the classifier to test textand evaluate results. Compared to the manual method, text classification based onmachine learning speeds a little time but also reduces the amount of manpower,material output. It effectively improves the accuracy rate of classification result.But whether the feature extraction of text using information gain, mutualinformation or vector polymerization technology, it does not consider the impact ofimplicit relationship inside word text, words, sentence and other elements; andpresent information and data presented high-dimensional, nonlinear characteristics.Following the traditional feature extraction algorithm can not reach robustness andhigh accuracy requirements for the analysis of existing data.Aiming at some problems of the existing, new solutions put forward in the textdata processing in this paper, in the full analysis of synonym word Lin (extension)coding and the structure characteristics, the word similarity calculation method isapplied to the text data processing to strengthen the feature weights, on this basis, inthe full understanding about the structure of HowNet, we use HowNet tore-weighting feature vector assignment and combine Locally Linear Embedding innonlinear dimension algorithm to achieve dimensionality reduction.The main work includes:(1) The corpus that experiment used have some garbage data in training text setand test text, for example, some documents only have title, no real content; Somedocuments are empty document; duplicate documents are appeared, etc., thesedocuments have no effect on the classification, instead, to a certain extent, may affectthe result of the classification, we remove the useless data through the noise reductionprocessing and keep useful data for classification.(2) We can express a concept in different words in article. For example, whilethe expression "happy" meaning, some people will use "cheerful", some people willuse the "pleasure", which can indeed improve the article readability. But it bringssome feature extraction difficult. Computers don’t like a person can be easilydistinguishing synonyms and near synonyms. This may cause extracting the differentword which has the same concept into different themes, which will affect the classification result. To solve this problem, we use word similarity computationmethod of synonym word Lin (extension) combining the characteristics which thesimilarity reaches a certain threshold and strengthening the feature weights, thusreducing the impact on the classification results.(3) In order to dig out the implied relationship between text words elements, weuse word semantic relevance computation method of HowNet to re-weight featurevector assignment to improve the classification accuracy.(4) Compared with conventional feature extraction algorithm, Local linearembedding algorithm can dig out the low-dimensional manifold embedded inhigh-dimensional feature space form, at the same time; maintain the existingmanifold structure, so it is a new feature extraction solution. |
