Heterogeneous Ensemble For Animation-oriented Information Extraction Of Chinese Short Messages

In the 1990 s, the technology of full process computer assistant animation auto-generation was proposed by Academician Lu Ruqian from CAS. In this technology, animation can automatically generate by the control of computer, at every step from the input of story described by restricted natural language to the final animation generation. In 2008, Researcher Zhang Songmao from CAS applied 3D auto-generated technology on mobile phone short messages, designed and implemented 3D Animation Auto-generation System on Mobile Phone, which aimed to send receiver auto-generated 3D animation according to sender’s message content.Auto-generate 3D animation according to sender’s message content, what we should do firstly is to extract information from message. Whether we can extract the key information or not, has a direct and obvious influence on the consistency of animation and messages. Up to now, we mainly use the rule-based method to extract information, however, the method, which is of relatively high accuracy, has a narrow coverage. To make up the limitation of the rule-based method, combine the range of content animation can show, we also use machine learning method to solve two problems: Chinese short message sentiment classification and topic classification. However, these methods improve the coverage, but reduce the accuracy. According to the statistics from March to May in 2014, there are 280 messages are tested in our system. Among them, 109 messages were classified correctly by sentiment classification with 39% accuracy, 162 by topic classification with 58% accuracy. Both have not reached the practical goal. For both message sentiment classification and topic classification use single classifier method, at present, we try to use heterogeneous ensemble of learners to improve the accuracy.The main work of this paper consists of the following two parts:Firstly, design and implement the heterogeneous ensemble system of Chinese short message sentiment classification. This system is used to analyze the emotional tendency of message content, classify the category of messages, and provide emotional information for subsequent animation plot planning. It is divided into two steps: The first is classification of subjective and objective messages, in this part experiment, NB, SVM, KNN and C4.5 are used as base classifiers to make up different combinations, as the heterogeneous ensemble classifier to execute experiment. The second is multi-label sentiment classification of subjective messages, in this part experiment, RAKEL, CC, MLKNN and BRKNN are used as base classifiers to make up different combinations, as the heterogeneous ensemble classifier to execute experiment.Secondly, design and implement the heterogeneous ensemble system of Chinese short message topic classification. The aim of this system is to analyze the topic of message content and guide the following animation plot planning. In this part experiment, NB, SVM, KNN and C4.5 are used as base classifiers to make up different combinations, as heterogeneous ensemble classifier to execute experiment.Through the use of 11 kinds of combination, perform sentiment classification experiment on 9600 messages, topic classification experiment on 17035 messages respectively. The experiments show that: compared with the single classifier, heterogeneous ensemble approach has a better effect on classification. The highest accuracy of sentiment classification is 63% with ensemble of RAKEL, MLKNN and BRKNN, and increase by 23 percentage points compared with the previous version. The highest accuracy of sentiment classification is 89% with ensemble of SVM, KNN and C4.5, and increased by 31 percentage points compared with the previous version. These two parts have reached practical goal. 3D Animation Auto-generation System on Mobile Phone can use the functions of these two parts.The main work of this paper make 3D Animation Auto-generation System on Mobile Phone can use rule-based method and machine learning method simultaneously, to extract information complementary. These two methods improve coverage and accuracy of information extraction, and make the final animation consistent with the content of messages. The future work contains: add more multi-label corpus in the corpus of experiments of multi-label classification, try more classification algorithms, and use more heterogeneous ensemble approaches.