Sentiment Analysis For Chinese Text In Continuous Valence-Arousal Space

Sentiment analysis refers to the use of computational linguistics to analyze, process, induce and deduce subjective texts with affective information. Compared to the categorical approach, dimensional approach can provide more fine-grained sentiment analysis. However, there are still several difficulties in applying dimensional sentiment analysis methods for Chinese text, such as cultural differences, the lack of dimensional affective lexicon and corpora, the coverage problem of existing lexicons, and the low accuracy of dimensional sentiment analysis. Designing a robust dimensional sentiment application for Chinese text is still a challenging work. To address these problems, this thesis focuses on sentiment analysis methods on both word- and text-level in continuous valence-arousal space. Main contents include:(1) For cross-lingual VA transformation of affective words, this thesis proposes a locally weighted method to improve linear regression. This method performs a regression around the point of interest using only training data that are "local" to that point, and thus can reduce the impact of noise from unrelated training data. It solves the under-fitting problems in linear regression. Experimental results show that the proposed method achieved a lower error rate and a higher correlation coefficient.(2) For the mono-lingual VA prediction, this thesis proposes a community-based weighted graph model that can select similar seeds with similar ratings (or the same polarity) to each unseen word to form a community (sub-graph) so that its VA ratings can be estimated from such high-quality seeds using a weighted propagation scheme. Experimental results show that the proposed method can effectively remove the noisy neighbors and improve the prediction performance.(3) On text-level, this thesis proposes a regional deep neural network to predict the VA ratings of texts. The proposed regional convolutional neural network (CNN) can extract the useful affective information in each region. Such regional information is sequentially integrated across regions using long-short term memory (LSTM) for VA prediction, which can improve the prediction accuracy of dimensional sentiment analysis. Experimental results show that the proposed method outperforms lexicon-based, regression-based, and NN-based methods proposed in previous studies.