| Abstract [eng] |
Now since almost every person has access to social network or hub, expression of opinion, statement or emotion became more visible. Such data set is a dream for the businesses and various analysts. It enables to measure impact of ad campaigns, success of new product or simply understand reaction to news or selected communication strategy. However, people tend to express opinions in complex ways. Rhetorical devices like sarcasm, irony, and implied meaning can mislead, thus complicating sentiment analysis task. In this work lexicon based semantic orientation (SO) algorithm SentiII was created which takes into account SO of past sentences and various contextual converters. By analyzing texts from popular English and Lithuanian "Facebook" pages, multiple strategies were tested how should previous text’s sentences affect ongoing SO score. Also good practices were listed how to handle various contextual elements (e.g. inversions, intensifiers or syntax). Moreover 10 semantic models were introduced (based on "Facebook" reactions) which later extended SO algorithm. SentiII was tested using 150 texts from popular Lithuanian "Facebook" pages comparing it with expert evaluations. |
