Concerning the problem that it is complex to detect the surface feature of continuous casting slab, a surface trait extraction method based on complex Contourlet decomposition was developed. Compared to conventional method, the new one has the characteristics of shift invariance, excellent directional selectivity and higher retrieval rate. The image was decomposed in Contourlet domain and then the image's sub-bands directional matrices were extracted by using directional filter bands, then the feature vector was constructed by the energy value, stand deviation and skewness. The support vector machine was trained by the feature vectors, to classify images. Industrial test result shows that the accurate rate of classifying surface features is about 90%, and it can be used in image feature extraction and slab flaw detection.

Taking into account that many existing name disambiguation approaches are not suitable for documents with few features, this paper put forward a method that can get more features related to documents through search engine extension with the help of rich Internet resources. Firstly, four types of queries were constructed according to the characteristics of the search engine rules, and then search was done based on those query rules and the first k documents were returned. Lastly, a feature selection method of Document Frequency (DF) was used in these documents and the selected features were added to the original documents. The experimental results show that the query expansion approach can improve the performance of name disambiguation and make F-measure increase from 76% to 81%.

Direct towards anonymity and high dynamic of online user in P2P TV, a P2P TV crawler, called TVCrawler was developed and deployed which enabled to launch an active measurement on several popular large scale P2P TV systems. The authors conducted a comparative research on time evolution and geographic distribution of online users in these different P2P TV systems. First, while intuitively researching the time evolution of online users in P2P TV channel, the method of MultiScale Entropy (MSE) analysis was introduced to investigate the complexity in time series of the number of online users. Second, the authors made a study on the regular pattern of online users' geographic distribution, and made a Google map-based visual representation about online users. Then, by analyzing the relationship between geographic distribution of online Chinese users and provincial economic development level of China, it is discovered that significant linear decreasing correlation exists between the two of them.

Compared to verbal predicates，the structure between nominal predicates and their roles in Semantic Role Labeling (SRL) is more flexible and complex. In this paper, some new word-related and syntactic features were explored from various nominal predicate-specific features to capture the structure information for nominal SRL. The experimental results show that the proposed nominal SRL system achieved the performance of 73.99 in F1-measure on gold parse trees and gold predicates, and outperformed the state-of-the-art nominal SRL. However, the performance dropped to 57.16 in F1-measure on automatic parse trees and automatic predicates. Finally, the training data were augmented with verbal SRL instances to examine whether nominal SRL could benefit from verbal instances. The experimental result show, however, adding verbal SRL instances does indeed improve the performance of nominal SRL, although the improvement is not statistically significant.