Can Sentiment Analysis Support Cognitive Discourse Analysis at Scale?

Language is a window into people's thoughts, attitudes, and worldviews, and analyses such as cognitive discourse analysis (CODA) can help reveal this. However, conducting CODA is often resource intensive, requiring considerable time and person-power. By contrast, natural language processing (NLP) approaches may potentially be able to replicate similar results using sentiment analysis for substantially reduced researcher effort. Here, we use respondent-validated data from Amazon product evaluations and Internet Movie Database (IMDB) movie ratings to contrast CODA against two NLP sentiment analysis tools (TextBlob and VADER). NLP approaches showed strong correlations with user-assigned sentiment scores in both data sets. For example, Textblob and VADER show Pearson correlations of $r$=0.4383 and 0.5188 respectively with the Amazon data and $r$ = 0.594 and $r$ = 0.4664 respectively with the IMDB data (all $p$<0.001), while CODA approaches yielded lower correlations (Amazon: $r$=0.1357; IMDB: $r$=0.1708 ; $p$<0.001). Our results show that NLP techniques achieved higher correlations with pre-established sentiment scores than CODA across the datasets tested. We argue that NLP can complement CODA by handling large-scale sentiment screening, while CODA remains better suited to the broader interpretive goals, such as mapping cognitive frames and discursive patterns, that fall outside this evaluation. To help this, we provide an open-source widget \href{https://gitlab.com/lleleena/sasi/-/blob/main/README.md}{Survey Analysis for Sentiment (SASi)}, which reduces the technical requirements of implementing NLP techniques. We note that CODA is a broader interpretive framework than sentiment scoring alone, and our comparison is necessarily limited to the sentiment dimension of CODA's output.