Sentiment analysis is the machine determination of the evaluative colouring of a text towards a subject — usually as positive, neutral or negative. In media monitoring it makes visible in what tone an organisation, brand or topic is being reported on, and how that tone changes over time.
How machine sentiment analysis works
Early methods worked with dictionaries: terms carried fixed ratings that were summed into an overall value. That is transparent, but fails on context and negation. Today's systems use transformer-based language models that rate a word in the context of the sentence — so "not bad" is correctly not read as negative.
Such models are trained on rated example texts and then assign a sentiment to new items, often with a confidence value. For German-language media monitoring, models trained on the German language and its idiosyncrasies are needed — models ported directly from English deliver systematically poorer results.
Document vs. aspect level
Sentiment can be measured at different levels:
- Document level: a single value for the whole item. Fast, but coarse — an overall neutral article may contain a sharply negative passage.
- Aspect-based analysis: sentiment related to individual aspects or entities. This makes it possible to see that an item is positive about the product but negative about the price.
For robust media intelligence, the aspect-based, entity-related analysis is usually more meaningful, because it answers the question "positive about what?".
Limits & sources of error
Sentiment analysis is an estimate, not a measurement. The main pitfalls:
- Irony and sarcasm are frequently misclassified.
- Quotations often carry the sentiment of the person quoted, not of the outlet.
- Technical language may use seemingly negative terms in a neutral way (for example "loss" in financial reporting).
- Domain dependence: a model trained on product reviews errs on political coverage.
Traceability as a corrective
Because single values can err, every sentiment rating should be traceable back to the specific item. That way an outlier can be checked rather than blindly trusted.
Why the trend matters more than the single value
A single negative article says little. Sentiment becomes meaningful as a trend across many items: a curve turning negative over weeks is a signal — even if each individual rating carries uncertainty, errors average out over volume. That is why sentiment in media monitoring is above all a trend indicator and early-warning signal, not an absolute verdict on a single text.
Frequently asked questions
How accurate is machine sentiment analysis?
Modern, transformer-based methods achieve high hit rates on clearly coloured texts, but reliably fail on irony, sarcasm and strongly context-dependent statements. That is why the sentiment trend across many items is more robust than the rating of a single text.
What does neutral sentiment mean?
Neutral means an item carries no clearly evaluative colouring towards the subject under observation — for instance a purely factual report. In media monitoring, neutral is the most common and often desired value; what matters above all are shifts into the negative or positive.
Should sentiment be measured at sentence or document level?
Both have their place. The document level gives a quick overall impression; aspect-based analysis at the sentence level shows what the sentiment actually refers to — for example positive about the product but negative about the price. For media monitoring, the finer resolution is usually more meaningful.