Sentiment Analysis in Communication

Lexicons (e.g., VADER, NRC, LIWC affect) need no training data, are transparent, and are reproducible, but they struggle with context, negation, and domain-specific language. Supervised and transformer-based classifiers can be far more accurate in a given domain but require labeled training data and risk poor transfer outside it. For communication research the decisive step either way is validation: report how well the chosen method agrees with human coding on a sample from your corpus.

Because tone is a substantive construct, not just an engineering target. Communication's content-analysis tradition treats any automated measure as a coding instrument that must demonstrate reliability and validity before its outputs support claims. Validating the classifier against a human-coded gold standard — and reporting accuracy and agreement (e.g., Krippendorff's alpha) — guards against drawing conclusions from systematically biased or domain-mismatched sentiment scores.

Generally not well. Sarcasm inverts surface valence ('Great, another scandal'), which lexicon methods miss entirely and even strong classifiers handle inconsistently without sarcasm-specific training. In genres where irony is common, researchers either flag sarcasm separately, restrict claims to aggregate trends where errors partly cancel, or fall back to human coding for the affected subset. Acknowledging this limitation is part of responsible reporting.