n-grams language model aims to use text feature combined of some words to train classifier. But it contains many redundancy words, and a lot of sparse data will be generated when n-grams matches or quantifies the test data, which badly influences the classification precision and limites its application. Therefore, an improved language model named W-POS (Word-Parts of Speech) was proposed based on n-grams language model. After words segmentation, parts of speeches were used to replace the words that rarely appeared and were redundant, then the W-POS language model was composed of words and parts of speeches. The selection rules, selecting algorithm and matching algorithm of W-POS language model were also put forward. The experimental results in Fudan University Chinese Corpus and 20Newsgroups show that the W-POS language model can not only inherit the advantages of n-grams including reducing amount of features and carrying parts of semantics, but also overcome the shortages of producing large sparse data and containing redundancy words. The experiments also verify the effectiveness and feasibility of the selecting and matching algorithms.
Traditional machine learning faces a problem: when the training data and test data no longer obey the same distribution, the classifier trained by training data can't classify test data accurately. To solve this problem, according to the transfer learning principle, the features were weighted according to the improved distribution similarity of source domain and target domain's intersection features. The semantic similarity and Term Frequency-Inverse Class Frequency (TF-ICF) were used to weight non-intersection features in source domain. Lots of labeled source domain data and a little labeled target domain were used to obtain the required features for building text classifier quickly. The experimental results on test dataset 20Newsgroups and non-text dataset UCI show that feature transfer weighting algorithm based on distribution and TF-ICF can transfer and weight features rapidly while guaranteeing precision.
Aiming at the sentiment classification for Chinese consumption comments, a method called two-dimensional coordinate mapping for sentiment classification based on corpus was constructed. According to the Chinese language characteristics, firstly, a more pertinent searching method based on corpus was proposed. Secondly, the rules of extracting the Chinese subjective phrases were defined. Thirdly, the choosing optimal seed words algorithm of the specific field was constructed. Finally, the two-dimensional coordinate mapping algorithm was constructed, which mapped the comment in two-dimensional Cartesian coordinates through calculating the coordinate values of the comment and decided the semantic orientation of it. Experiments were conducted on 1200 comments of milk (half of them are positive or negative comments) in Amazon. In the experiments, word “henhao-lou” was chosen as the optimal seed word by using choosing optimal seed words algorithm, then the sentiment orientation of it was decided according to two-dimensional coordinate mapping algorithm. The average F-measure of the proposed algorithm reached more than 85%. The result shows that the proposed algorithm can classify the sentiment of Chinese consumption comments.
A kind of topic tree detection method based on Latent Dirichlet Allocation (LDA) model was put forward, in order to solve the problems of nonstandard terms, randomness, uncertainty of reference and large number of network terms in microblog texts, which can not be solved in traditional detection method. Relevant microblogs were reorganized into a topic tree by increasing information entropy in Natural Language Processing (NLP), combining with the design idea that Dirichelet prior experience value α and experience value β vary with the topic number, then the contribution statistics of every word in the text was achieved using the specific dual probability statistical method of this model. Thus, the interference information would be disposed in advance and the influence of garbage data on topic detection was excluded. Using this contribution as the parameter value of the improved Vector Space Model (VSM), bursty topics were extracted through calculating the similarity between texts, in order to improve the detection precision of bursty topics. Experiments of the proposed detection method were made from two aspects: comparison of the value of F and the manual detection. The experimental data show that, this algorithm not only can detect the bursty topics, but also can improve the precision about 3% and 7% respectively compared with the HowNet model and the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm, and it is more in accordance with human's logic judgments than the traditional ones.
