学位論文: [徳島大学.先端技術科学教育部.システム創生工学専攻.知能情報システム工学コース]/[課程博士]/ウェブログの多重ラベル感情認識に関する研究/20120910/[任 福継]
(英) Recognition of Multi-label Emotion for Weblog
(英) With plenty of online resources constantly emerging (like product reviews, weblogs, emails, chatting messages, etc.), emotion analysis in text has become increasingly important in human-computer interaction. Moreover, textual emotion recognition can also facilitate the communication through other mediums such as facial expressions or speech. Many previous researches have been carried out to recognize single-label emotion in text. However, with the multiplicity of human being's affection and higher requirements for emotional intelligence, textual emotion detection goes on to a new stage of multi-label classification. This dissertation is dedicated to the recognition of multi-label emotion for weblog. Based on the Chinese emotion corpus Ren_CECps which contains weblog articles with detailed emotional tags, we model the problem as a multi-label text classification task. First, we compare different approaches to recognize word emotion in the dimension of eight emotion categories with corresponding intensities. In order to obtain emotion intensity, the algorithm of semantic similarity computation is revised for aiding emotion intensity computation. We propose the variant of semantic analysis method to compute word emotion vector, exploiting not only semantic relations but also morpheme characteristics. SVM model is also trained for word emotion classification in the comparison. As a result, a hybrid approach comes into being by integrating the two methods above. Next, we explore different features to recognize sentence emotion based on an ensemble multi-label algorithm and word emotion. The 8-dimension intensity representation for word emotion is compared to some traditional features. In order to deal with relatively complex sentences, additional features are examined to improve the performance. Grammatical characteristics of punctuations, disjunctive connectives, modification relations and negation are analyzed and integrated into the features. Interesting issues special for multi-label classification are presented in our work. Experiments are carried out for the proposed methods. The results show that: (1) The integration of morpheme characteristics and semantic relations can reduce the running time and improve the accuracy in emotion vector computation for a word. (2) The hybrid approach with two stages is effective for recognizing word emotion. (3) Multiple-dimension emotion representation for word emotion with grammatical features can efficiently classify sentence emotion in a multi-label problem.
|年月日||必須||2012年 9月 10日|