Research On A Method Of Calculating Sentence Similarity By Comprehending Multi-level Information

Many traditional methods about sentence similarity computing, which just incline to one side of sentence, are only practical in specific areas. If we change its field or focus, the portability or accuracy of these methods may have not good performance. Therefore, in order to compensate for the lack of traditional methods, it is urgent to use a new method to balance many terms of information. This paper aims to research on a new method of calculating sentence similarity by comprehending multi-level Information, the main research work as follows:1. A method which based on two semantic dictionaries to calculate similarity between words has been proposed. The two semantic dictionaries are CiLin and HowNet. Meanwhile, the antonymous words are specially considered, which is joined to calculate similarity between words.2. A method which based on many levels of information to calculate similarity between sentences has been proposed. First of all, based on the perspective of sentence semantics, this method put forward to use maximum weight matching algorithm to obtain the maximum similarity between two sentences. Secondly, a method based on tree-kernel algorithm to calculate the syntactic similarity of two sentences has been proposed. Meanwhile, for the case of having two pairs of antonyms between two sentences, a method based on double antisense relationship has been put forward. Finally, the final method based on a variety of above information are proposed.3. Do several experiments to verify the availability of above proposed methods. For similarity computing of words, some experiments about synonyms recognition, derogatory or commendatory word recognition, text clustering have been performed, the results of which show that the method has better accuracy, recall rate and F-measure rate. For sentence similarity calculating, two different strategy about experiments have been done to prove the effectiveness of the above method.