Keyword sentiments represent sentiment score of the keyword computed from each individual article. For example, one can search the keyword "Facebook" and find the sentiment scores appeared at the top of each article, which shows the total or combined sentiment score of the keyword "Facebook" in that article (it could be a aggregation of a few instances of scoring on the same keyword in one article).
This keyword sentiment score is not the text score of the whole article, which is shown when no keyword search is performed. For example, when one goes to the news main page of the app, the score found already on each news article is a text sentiment score, which is computed from all sentences in the text instead, without any keywords involved.
Please also see Q16, Q28, Q35