Effects of Word Class and Text Position in Sentiment-based News Classification

摘要

In news domain, sentiments captured in the form of sentiment labels (emoticon) give a quick feedback of reactions towards the contents of the news. As these reactions are valuable indicators for social and political well beings, we are motivated to automate the classification of news texts based on these indicators, e.g. happy, sad, angry, amused etc. Unlike other review texts that contain more explicit words which can be interpreted directly for sentiment classification, news texts mostly report facts and figures. This resulted in needs to identify whether contents of news can be exploited for classification or otherwise. Hence, in this work, a study is conducted to analyze and determine the relevant key parts of news contents that can be to be used for sentiment-based classification. Two criteria, i.e. text Part of Speech and text position, which could possible influence the training of the classifier are studied. Evaluations are conducted on the collection of 250 English news texts labelled with sentiments from sentiment voting system. The results for sentiment-based category has recorded F score of 0.422 whereas for polarity-based category has recorded F score of 0.837. The study has shown that when finer categories (e.g. happy, sad etc.) are used, the inspected criteria are less effectively; however, when these categories are based on polarity orientations, the outcomes show potentials of the proposed criteria especially for text positioned at headlines and text using adjective words.