A spatial approach to perception identification in editorials enhanced with anaphora resolution

A method is proposed to annotate editorials and news articles for sentences that most accurately represent the opinion of the speaker towards the issue. The speaker´s point of view or level of discernment of the issue is whittled out. A list of informative and related keywords is extracted from the document based on their frequency of occurrence. Subsequently, pronominal anaphora is resolved at the document level based on minimum distance measures. Normalized Google distance (NGD) is employed to resolve appositive instances. The vector space model is then applied to the anaphora resolved document. Based on distance measures in high dimensional space, the cohesion between target word pairs is observed. Strongly cohesive word clusters qualify as opinion words. Results from a pilot collection of short opinion editorials and critical reviews from newspapers and manual annotation of perspective of up to 710 such sentences indicate an agreement of 76% in opinion annotation.