Aiming at the low accuracy caused by the phenomenon of rule weight weakening from iterations of machine learning when judging the sentence relationships by applying rules and machine learning methods, the method of strengthening the imported obvious rule characteristics in the process of combining rules and machine learning was proposed. Firstly, these specific characteristics that having obvious rules such as dependency vocabulary, syntax and semantics information were extracted; secondly, universal characteristics were extracted based on these words that could indicate relationships; then, the characteristics were written into the data vector that to be input, and another dimensional vector was added to store the obvious rule characteristics; Finally, rules and machine learning methods were combined with LIBSVM model to perform the experiment. The experimental results show that the accuracy rate is averagely 2% higher than that before strengthening the characteristics, and all kinds of relationships' accurate rate, recall rate and F1 value show good results as a whole, their average values achieved 82.02%, 88.95% and 84.76%. The experimental ideas and methods are important for studying the compactness of adjacent sentences.

In order to solve the problems in Chinese sentence grouping domain, including the lack of computational linguistics data and the joint makers in a discourse, this paper proposed an automatic Chinese sentence grouping method based on Multiple Discriminant Analysis (MDA). Moreover, sentences group was rarely considered as a grammar unit. An annotated evaluation corpus for Chinese sentence group was constructed based on Chinese sentence group theory. And then, a group of evaluation functions J was designed based on the MDA method to realize automatic Chinese sentence grouping. The experimental results show that the length of a segmented unit and one discourse's joint makers contribute to the performance of Chinese sentence group. And the Skip-Gram model has a better effect than the traditional Vector Space Model (VSM). The evaluation parameter P_μ reaches to 85.37% and WindowDiff reduces to 24.08% respectively. The proposed method has better grouping performance than that of the original MDA method.