一般注記Textual emotion recognition (TER) is the process of automatically identifying emotional states in textual expressions. It is a more in-depth analysis than sentiment analysis. Owing to its significant academic and commercial potential, TER has become an essential topic in the field of NLP. Over the past few years, although considerable progress has been conducted in TER, there are still some difficulties and challenges because of the nature of human emotion complexity. This thesis explores emotional information by incorporating external knowledge, learning emotion correlation, and building effective TER architectures. The main contributions of this thesis are summarized as follows:
(1) To make up for the limitation of imbalanced training data, this thesis proposes a multi-stream neural network that incorporates background knowledge for text classification. To better fuse background knowledge into the basal network, different fusion strategies are employed among multi-streams. The experimental results demonstrate that, as the knowledge supplement, the background knowledge-based features can make up for the information neglected or absented in basal text classification network, especially for imbalance corpus.
(2) To realize contextual emotion learning, this thesis proposes a hierarchical network with label embedding. This network hierarchically encodes the given sentence based on its contextual information. Besides, an auxiliary label embedding matrix is trained for emotion correlation learning with an assembled training objective, contributing to final emotion correlation-based prediction. The experimental results show that the proposed method contributes to emotional feature learning and contextual emotion recognition.
(3) To realize multi-label emotion recognition and emotion correlation learning, this thesis proposed a Multiple-label Emotion Detection Architecture (MEDA). MEDA comprises two modules: Multi-Channel Emotion-Specified Feature Extractor (MC-ESFE) and Emotion Correlation Learner (ECorL). MEDA captures underlying emotion-specified features with MC-ESFE module in advance. With underlying features, emotion correlation learning is implemented through an emotion sequence predicter in ECorL module. Furthermore, to incorporate emotion correlation information into model training, multi-label focal loss is proposed for multi-label learning. The proposed model achieved satisfactory performance and outperformed state-of-the-art models on both RenCECps and NLPCC2018 datasets, demonstrating the effectiveness of the proposed method for multi-label emotion detection.
連携機関・データベース国立情報学研究所 : 学術機関リポジトリデータベース(IRDB)(機関リポジトリ)