Article Preview
TopThere are several approaches in textual affect sensing (Shanahan, 2006), namely: 1) affective keyword spotting (Strapparava, 2007), 2) commonsense approach (Liu, 2003), 3) rule-based approach (Boucouvalas, 2003; Chaumartin, 2007), and 4) machine learning approach (Kim, 2006; Alm, 2008).
The affective keyword spotting is simple and dependent basically on defined affect-bearing words; however, it is inaccurate in many cases because it disregards syntactic and semantic information of text. The commonsense approach strongly relies on collected real-world knowledge which is sometimes encoded in a graph form. The advantage of this approach is to take account of contextual information, if a similar situation exists in the knowledge base; however, it is hard or impossible to cover all the possible situations. The machine learning approach uses large annotated corpus and is popular especially for classifying sentiments, i.e., negative/positive opinions. It is difficult, however, in the affect sensing, to formulate diverse set of features; as a result, it mostly disregards modifiers, negation and condition constructions, syntactic relations and semantic dependencies in sentences. It can be said that the machine learning approach is semantically weak and less accurate for sentence-level analysis at present, though it is promising. The rule-based approach can take into account the contextual information and work on sentence level. The performance can be improved by adding rules and extending affective lexicon. The weakness of this approach is that rules always have exceptions. Nevertheless, this is the most practical at present. However, linguistic analyses have been weak in many rule-based approaches. Our two approaches are basically rule-based ones, but have incorporated elaborated functions into their linguistic analyses.
For example, a so-called polarity shift problem occurs in the following sentences: