| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 91 === This thesis presents an emotion classification system based on semantic content analysis. So far the linguistics based emotion extraction systems usually deal with emotional keywords or manually defined emotional rules and also have some disadvantages. Firstly, though the utilization of emotional keywords is useful for the increase of the classification accuracy, it uses only part of the linguistic information. A system with only keyword information cannot take into account the deep information of linguistics, such as semantic meaning. Secondly, the rules defined manually strongly depend on the application domain. As domain changes, rules have to be re-defined by some domain experts manually. This makes the construction of the system time consuming and un-portable. The system proposed in this thesis tends to solve these two problems.
We first reduce the basic rules that can explain the essential of emotion generation from the literatures in emotion psychology. According to the linguistic definition, such as HowNet and CKIP, we define some domain independent semantic labels. With the help of HowNet, we can extract the semantic meaning of an input sentence by these semantic labels and transfer it to semantic transactions. The data mining technique is then applied to generate the emotional rules from these semantic transactions automatically. Finally, compared to the emotional rules, the semantic transactions are mapped to a vector space and used to train the SVM classification models.
In the experiment, the classification accuracy was achieved at 80% for the original collected corpus. It can also achieved about 60% for other corpus. From the experimental results, the performance of semantic label and automatic rule generation is promising. And the system can essentially satisfy the requirement for domain portability.
|
|---|
