The Affordances and Limitations of Nlp Tools in Supporting Discourse Analysis for News Articles

Discourse analysis and natural language processing appear to share fundamental priorities: the detailed and systematic examination of text. However, authors have argued that discourse analysis does not have a strong tradition in the field of applied natural language processing, mainly because of the computational challenges associated with the nature and characteristics of discursive resources [5]. In this study, I assess the current affordances and limitations of a specific set of NLP tools in supporting discourse analysis, in order to identify possible areas for further research and development. I conduct manual discourse analysis on a pool of 15 newspaper articles and identify recurring discursive strategies. I then explore the extent to which these strategies are supported by Stanford CoreNLP [11], a suite of tools that is considered to be foundational and comprehensive. My findings show that of the 12 recurring discursive strategies, four are fully supported by this tool; five are partly supported; and three are not. I conclude that many of the dimensions that are not addressed can nevertheless be supported by a combination of machine learning techniques and by other NLP-based resources, for example those that draw on rhetorical structure theory (RST) [9]. I conclude by suggesting areas for future exploration.