Compared with computational language resources in English, word knowledge in Japanese for sentiment analysis is relatively under-developed. Sentiment axes have been used in Japanese psychological questionnaires for many years, and in this paper, we employ these axes for specific antonym relations in a word-relation network. The initial network is trained using target documents for sentiment analysis. We evaluate this trained word-relation network in several ways. First, we compare our results with a major sentiment-polarity dictionary. Then, we analyze the ex-tracted sentiment axes in each domain subjectively, along with the network's performance in predicting the grading of the reviews.
A recurrent neural network which its structure and link weights between units are appropriately determined has the ability to adequately predict for time-series data. It is difficult to find an optimal structure of the recurrent neural network minimizing both of the training and the generalization errors. In this paper, we propose an optimization method for structural design of neural networks for time-series data by employing tabu search with efficient neigh-borhood search. To demonstrate the efficiency and effectiveness of the proposed method, we apply our method to several benchmark problems of time-series data and it delivers superior or equal performance to other existing struc-tural optimization methods.