The proposed project aims at the conceptualisation and implementation of a system carrying out bi-directional sentiment composition. The term bi-directional is meant to denote the insight that the polarity (positive, negative or neutral) at the document level depends on, but at the same time restricts the polarities at the sentence level. Positive and negative evaluations at the text level do not alternate all of a sudden, rather they are rhetorically indicated (e.g. 'but', 'however'). So two neighbouring segments separated by a valency shifter (e.g. 'but') must have inverse polarities. This global constraint might induce a polarity shift in a top down manner forcing an otherwise bottom up operating sentiment composition to revise its decisions.
We propose a new architecture for sentiment composition. At the document level, a constraint-based optimisation approach is taken that allows for global constraints. The sentence- and phrase-level sentiment theory is formulated as a Stochastic Logic Program (SLP). This SLP has to be learned from preclassified data and has to be tuned by the parameter estimation part of a Relational Learner. It represents the reliability of the whole process of sentiment composition, which depends on the quality of the parse trees and the various lexical resources for sentiment analysis that we plan to integrate. We also introduce a measure for the polarity strength that a particular SLP derivation bears. This way, we can answer the question how good or bad an evaluation is, but also how sure we actually are about it.
Manfred Klenner 2008-10-14