Here’s a brief summary:
Automated sentiment analysis uses a combination of two approaches to determine sentiment: lexical analysis (which looks for emotionally-weighted words) and machine-learning (which relies on a “training corpus” of manually scored documents to predict the emotional content of new documents that it processes.) Continued human intervention (“training”) in the machine learning process may improve results, but makes it more or less meaningless to compare the results over time.
No two sentiment analysis tools on the market agree closely; with most disagreements occurring and around “sentiment-neutral” comments. Since these form the bulk of the content, there is much room for disagreement.
Sentiment analysis on short content (e.g. Tweets) lacks sufficient context for accurate judgments, whereas analysis of longer content often lacks sufficient relevance (e.g. the search term may only be mentioned in passing, and the sentiment score refer to a different )
Human-based/manual sentiment analysis also faces reliability challenges. There is often as much disagreement between two human analysts as there is between two automated systems. Worse still, research demonstrates that the same person – when presented with the same text on different occasions – may score it differently each time.
Conclusion: Sentiment Analysis may be used to guide customer service engagement, but should not be used as a KPI
If you’ve more time on your hands, here’s a slide deck I pulled together that mentions more research (my informal research, and others’ more rigorous.)