Computational approaches to sentiment analysis eschew a general theory of emotions and focus on extracting the affective content of a text from the detection of expressions of sentiment. These expressions are assigned scalar values, represent­ing a positive, a negative or neutral sentiment to­wards some topic. Using information retrieval, text mining and computational linguistic techniques to­gether with a set of dedicated linguistic resources, one can calculate opinions exploiting the detected "bag of sentiment words". Recently, new meth­ods aim to assign fine-grained affect labels based on various psychological theories-e.g., the MPQA project (Wiebe et al, 2005) based on literary the­ory and linguistics and work by (Read et al., 2007) based on the Appraisal framework (Martin and White, 2005).

We think there is still room for improvement in this field. To get an accurate appraisal of opin­ion in texts, NLP systems have to go beyond pos­itive/negative classification and to identify a wide range of opinion expressions, as well as how they are discursively related in the text. In this paper, we describe a preliminary study for a discourse based opinion categorization. We propose a new annotation schema for a fine-grained contextual

©2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li­cense (http://creativecommons.0rg/licenses/by-nc-sa/3.O/). Some rights reserved.

opinion analysis using discourse relations. This analysis is based on a lexical semantic analysis of a wide class of expressions coupled together with an analysis of how clauses involving these expres­sions are related to each other within a discourse. The aim of this paper is to establish the feasibil­ity and stability of our annotation scheme at the subsentential level and propose a way to use this scheme to calculate the overall opinion expressed in a text on a given topic.

We categorize opinion expressions using a typol­ogy of four top-level categories (see table 1): Re­porting expressions, which provide an evalu­ation of the degree of commitment of both the holder and the subject of the reporting verb, Judg­ment expressions, which express normative eval­uations of objects and actions, Advise expres­sions, which express an opinion on a course of ac­tion for the reader, and Sentiment expressions, which express feelings (for a more detailed de­scription of our categories see (Asher et al, 2008)).

Our approach to categorize opinions uses the lexical semantic research of (Wierzbicka, 1987), (Levin, 1993) and (Mathieu, 2004). From these classifications, we selected opinion verb classes and verbs which take opinion expressions within their scope and which reflect the holder's com­mitment on the opinion expressed. We removed some verb classes, modified others and merged re­lated classes into new ones. Subjective verbs were split into these new categories which were then ex­tended by adding nouns and adjectives.

Our classification is the same for French and En­glish. It differs from psychologically based classi­fications like Martin's Appraisal system : in ours the contents of the Judgment and Sentiment categories are quite different, and more detailed for sentiment descriptions with 14 sub-classes. Ours is also broader: the reporting and the ad­vise categories do not appear as such in the Ap­praisal system. In addition, we choose not to build our discourse based opinion categorization on the top of MPQA (Wiebe et al, 2005) for two reasons. First, we suggest a more detailed analysis of pri­vate states by defining additional sets of opinion classes such as hopes and recommendations. We think that refined categories are needed to build a more nuanced appraisal of opinion expressions in discourse. Second, text anchors which corre­spond to opinion in MPQA are not well defined since each annotator is free to identify expression boundaries. This is problematic if we want to in­tegrate rhetorical structure into opinion identifica­tion task. MPQA often groups discourse indica­tors (but, because, etc.) with opinion expressions leading to no guarantee that the text anchors will correspond to a well formed discourse unit.

Rhetorical structure is an important element in un­derstanding opinions conveyed by a text. The fol­lowing simple examples drawn from our French corpus show that discourse relations affect the strength of a given sentiment. SI: [I agree with you]a even if I was shocked and S2 : Buy the DVD, [you will not regret it]},. Opinions in SI and S2 are positive but the contrast introduced by even in 51 decreases the strength of the opinion expressed in (a) whereas the explanation provided by (b) in 52 increases the strength of the recommendation.

Using the discourse theory SDRT (Asher and Las-carides, 2003) as our formal framework, our four opinion categories are used to label opinion ex­pressions within a discourse segment. For exam­ple, there are three opinion segments in the sen­tence S3: [[It's poignant]d, [sad]e]g and at the same time [horrible]]

We use five types of rhetorical relations: Con­trast, Correction, Support, Result and Continuation (For a more detailed description see (Asher et al, 2008)). Within a discourse seg­ment, negations were treated as reversing the po­larities of the opinion expressions within their scope. Conditionals are hard to interpret because they affect the opinion expressed within the conse­quent of a conditional in different ways. For exam­ple, conditionals,expressions of advise can block the advice or reverse it. Thus if you want to waste you money, buy this movie will be annotated as a recommendation not to buy it. On the other hand, conditionals can also strengthen the recommenda­tion as in if you want to have good time, go and see this movie. We have left the treatment of con­ditionals as well as disjunctions for future work.

Experiments and Preliminary Results

We have analyzed the distribution of our categories in three different types of digital corpora, each with a distinctive style and audience : movie re­views, Letters to the Editor and news reports in English and in French. We randomly selected 150 articles for French corpora (around 50 articles for each genre). Two native French speakers anno­tated respectively around 546 and 589 segments. To check the cross linguistic feasability of gener­alisations made about the French data, we also an­notated opinion categories for English. We have annotated around 30 articles from movie reviews and letters. For news reports, the annotation in En­glish was considerably helped by using texts from the MUC 6 corpus (186 articles), which were an­notated independently with discourse structure by three annotators in the University of Texas's DIS-COR project (NSF grant, IIS-0535154); the anno­tation for our opinion expressions involved a col­lapsing of structures proposed in DISCOR.

The annotation methodology is described in (Asher et al, 2008). For each corpus, annotators first begin to annotate elementary discourse seg­ments, define its shallow representation and finally, connect the identified segments using the set of rhetorical relations we have identified. A segment is annotated only if it explicitly contains an opin­ion word that belong to our lexicon or if it bears a rhetorical relation to an opinion segment.