| With the development of information technology, great convenience has beenbrought for news communication. More and more news polarities are exposed to thepublic at the same time, which would influence the public opinion directly or indirectlyand bring new challenges for public sentiment monitor. The efficient analysis for newspolarity could help the public obtain the latest developments in society and the currentevents information. Also it would help supervision departments learn the newesttendency of public opinion so as to respond in time to reduce the propagation of negativenews in network, guide reasonably the public’s attention and create a harmonious cultureenvironment.In news area, there are some difficulties in text sentiment analysis. For examples,news emotional expression is relatively obscure and news contents are originated fromall aspects in society, meanwhile it is lacking of effective discourse analysis theories andtools. Since that not all the news sentences contain their corresponding sentiment, andsome sentences have little relationship with news content, the solution that full news textwas used for sentiment analysis might be biased. According to the structurecharacteristics in news text, this paper firstly proposed a topic sentence extraction modelbased on multi-feature fusion according to news text structure and then discussed newssentence sentiment analysis methods, integrated them into a combination finally for newstext sentiment analysis task. In addition, when a news special topic is derived fromcertain sudden news whose fact is negative, the generation of stage summarization for thetopic is explored. Thus the main researches in this paper can be concluded as thefollowing aspects:(1) A multi-feature fusion model is discussed for news topic sentence extraction. Analgorithm to detect high-frequency words in news text is firstly introduced in detail, andthen these words were endowed with different weights according to the location wherethey appeared so as to evaluate the importance of the sentence they were in. The role ofnews content expression was discussed for news headline and then a quantitativecalculation method is put forward to determine the similarity between news headline andnews sentence. According to the "inverted pyramid" text structure, news sentencelocation weight was described. Some tendentious-cue words were collected to find outthe potential emotional sentence in news text. Thus, combining the above four characteristics, new topic sentence extraction model based on multi-feature fusion wasconstructed.(2)Three sentiment analysis methods for news sentence are discussed. The first is amethod based on an emotional dictionary, which includes21175ordinary emotionalwords and1438news emotional vocabulary. At the same time, some quotative wordswere collected, too. Then according to the priority of three kinds of different words, amethod for news sentence sentiment analysis based on emotional dictionary wasproposed, which performed well in the evaluation task1on COAE2014, combining withthe news topic sentence extraction model described in part (1). The second method ismainly based on machine learning through the feature selection such as emotional-wordfeature, unigram feature and their combination. The third is an integrate method of thetwo former ones, with the combination of emotional dictionary and unigram feature, andis used to classify emotion for news subjective sentences.(3)A standard description about news text polarity analysis process is given basedon the above researches. The analysis task was separated into two subtasks, news topicsentence extraction and news sentence sentiment analysis. With the above topic sentenceextraction model and the sentiment analysis method for news sentence, emotionaldictionary&unigram feature, news text polarity analysis task was completed.(4)An algorithm for stage summarization generated from news special topic isproposed. Focusing on certain news special topic derived from a sudden negative-fact,the relationship was discussed between the phenomena of news sub-topic formation,continuation, disappearance and news topic stage summarization, then TDT technologywas used to generate stage summarization for news special topic by bidirectional clusterbased on time stream and secondary cluster among topic intersections after bidirectionalsteps, finally it’s showed that the algorithm could receive good recall rate byexperiments. |
PDF Full Text Request