Expresses-an-opinion-about: using corpus statistics in an information extraction approach to opinion mining

We present a technique for identifying the sources and targets of opinions without actually identifying the opinions themselves. We are able to use an information extraction approach that treats opinion mining as relation mining; we identify instances of a binary "expresses-an-opinion-about" relation. We find that we can classify source-target pairs as belonging to the relation at a performance level significantly higher than two relevant baselines. This technique is particularly suited to emerging approaches in corpus-based social science which focus on aggregating interactions between sources to determine their effects on socio-economically significant targets. Our application is the analysis of information technology (IT) innovations. This is an example of a more general problem where opinion is expressed using either sub- or supersets of expressive words found in newswire. We present an annotation scheme and an SVM-based technique that uses the local context as well as the corpus-wide frequency of a source-target pair as data to determine membership in "expresses-an-opinion-about". While the presence of conventional subjectivity keywords appears significant in the success of this technique, we are able to find the most domain-relevant keywords without sacrificing recall.