A Framework to Extract Arguments in Opinion Texts

A Framework to Extract Arguments in Opinion Texts

María Paz García-Villalba (Institut de Recherche en Informatique de Toulouse, France) and Patrick Saint-Dizier (Institut de Recherche en Informatique de Toulouse, France)
DOI: 10.4018/jcini.2012070104


In this article, the authors present foundational elements related to argument extraction in opinion texts with the objective to design a model of how consumers develop argumentation in such texts. A second goal is to analyze and synthesize user preferences and therefore user value systems from these arguments. They show that (1) within the context of opinionated expressions, a number of evaluative expressions with a ‘heavy’ semantic load receive an argumentative interpretation, and (2) that the association of an evaluative expression with a discourse structure such as an elaboration, an illustration, or a reformulation must also be interpreted as an argument. The authors develop a conceptual semantics of these discourse structures and show how they are analyzed using the Dislog programming language, running on the platform, dedicated to discourse analysis.
Article Preview

Aims And Challenges

In the framework of opinion analysis, besides determining whether consumers or citizens are happy or unhappy with a given product or decision, it is of most importance to be able to identify why they like or dislike a product, or approve or disapprove of a decision. These reasons are mainly provided by arguments associated with the evaluations of product attributes. At the present stage of opinion analysis, there is no ‘deep’ analysis of the satisfaction or dissatisfaction of consumers or citizens that would provide such an insight on the arguments behind these evaluations. On a large set of texts, over a closely related range of products, this analysis would allow the induction of the main priorities, value systems or preferences of consumers or citizens. This is however a difficult task, as illustrated in Van Eemeren and Grootendorst (1984, 1992), among others. Argument extraction in texts in general and in opinion texts in particular is an emerging area. Applied to opinion analysis, to the best of our knowledge, and besides principles given in Fiedler and Horacek (2001) and the results presented in Delmonte and Pallotta (2012), this constitutes a relatively innovative step that should allow the capture of consumers underlying motivations. Besides the extraction of arguments, the open challenge remains the construction of a synthesis of the extracted arguments in natural language.

Furthermore, a system of consumer preferences can be induced for each feature from these arguments. Given a recommendation (e.g., I strongly recommend it, do not stay here, go there if no alternative, etc.) and a list of attributes which are positively evaluated at various degrees (e.g., very well located, quite cheap) as well as another list of attributes which are negatively evaluated (not so clean, small rooms, quite noisy), it is possible to elaborate a system of user preferences or system of values. For a given consumer, it is indeed possible to identify his preferences: if the overall evaluation is positive, this roughly means that the positive attributes are globally more important than the negative ones. For a given product, on a larger scale, considering the pros and cons and the recommendations over a set of consumer evaluations, it is then possible to infer that some features are more essential than others, e.g., that localization and fares are more important than comfort for young consumers. In general, a system of values or preferences must be elaborated by consumer categories, e.g., by age and aims, since priorities may be different.

From an application point of view, identifying arguments related to opinion expression and making a synthesis is of much importance for the service provider or decision-maker who can then focus on the criteria which need to be improved first.

In this article, from a linguistic and cognitive point of view, we show that (1) within the context of opinion expression, a number of evaluative expressions, under the form of attribute-value pairs, where the value has a ’heavy’ semantic load, are interpreted as arguments, and (2) that the association of an evaluative expression with a discourse structure such as an elaboration, an illustration, or a reformulation must also be interpreted as an argument. These constructions are central to identify the why behind the evaluated statement.

In this article we develop a global conceptual semantic representation for these constructions, since they need to be interpreted to infer user preferences and to construct a synthesis. We feel this is also an important step in this type of work, since a conceptual representation should allow the abstraction of natural language utterances, therefore allowing their interpretation, inference and synthesis. Finally, we show how an automatic recognition of these structures can be implemented in the Dislog programming language on the <TextCoop> platform, dedicated to discourse analysis (Saint-Dizier, 2012).

Complete Article List

Search this Journal:
Open Access Articles: Forthcoming
Volume 14: 4 Issues (2020): 2 Released, 2 Forthcoming
Volume 13: 4 Issues (2019)
Volume 12: 4 Issues (2018)
Volume 11: 4 Issues (2017)
Volume 10: 4 Issues (2016)
Volume 9: 4 Issues (2015)
Volume 8: 4 Issues (2014)
Volume 7: 4 Issues (2013)
Volume 6: 4 Issues (2012)
Volume 5: 4 Issues (2011)
Volume 4: 4 Issues (2010)
Volume 3: 4 Issues (2009)
Volume 2: 4 Issues (2008)
Volume 1: 4 Issues (2007)
View Complete Journal Contents Listing