Attitude in Reported and Non-reported News: A Critique of Sentiment Analysis in Corpus Pragmatics

Abstract This study uses natural language processing (NLP) tools to examine a large Canadian English-language news corpus with respect to quotation and positive/negative sentiment. Specifically, we analyse sentiment in reported/quoted speech in comparison to non-quoted speech, testing the hypothesis that quoted speech contains negative sentiment and is more subjective. Our study explores whether NLP tools that simplify pragmatically complex concepts (such as attitude/evaluation/stance) can be used to test hypotheses that derive from discourse analytic or pragmatic studies of news discourse. We show that sentiment analysis is not suitable for accurate analysis of the news values of Positivity and Negativity and cannot be used to test hypotheses that derive from assumptions about these news values. At the same time, some of the insights from the sentiment analysis confirm our hypotheses (and are in line with other corpus studies), and sentiment results can be a starting point for additional qualitative analysis. Finally, we suggest a range of possible developments for sentiment analysis which draw on linguistic considerations.