How sentiment analysis works
A tool gathers mentions — posts, comments, reviews, forum threads — and gives each one a label. Older systems matched words against dictionaries of positive and negative terms, with rules for negation and emphasis. Current tools use language models trained on labelled text, which read context better and can also tag the emotion or the topic behind a mention. Either way the output is a guess with a confidence attached, rolled up into a score and a trend line.
That distinction matters more than any dashboard suggests: the label is a prediction, not a fact. Accuracy depends on the language, the subject and the material the model was trained on. Nepali written in Devanagari, Romanised Nepali and comments that switch between Nepali and English are much weaker ground than plain English, so a Kathmandu business reading an English-trained score is often reading noise dressed up as insight.
Why sentiment analysis matters
Volume on its own cannot tell a surge of praise from a surge of complaints. Mentions spike after a campaign, a price change or an outage, and until tone is attached you do not know whether to celebrate or to respond.
Used properly it does three jobs. It gives early warning when the negative share starts climbing, before anyone rings to tell you. It groups complaints by theme, so you can see whether the problem is delivery, price or support. And it lets you compare the tone of conversation about you with the tone about a competitor, which is a more honest reading than raw share of voice.
Where sentiment analysis goes wrong
Sarcasm, jokes, slang and mixed sentences defeat it regularly. A review saying the delivery was late but the staff were lovely contains both, and a single label loses the useful half. Neutral becomes a dustbin: news repeats, bot posts and anything the model could not read all land there, which makes the neutral share look meaningful when it is mostly leftovers.
Two reporting habits do the rest of the damage. Comparing scores between tools is meaningless, because none of them define the labels the same way. And steering by the aggregate hides the thing that matters most, since one influential complaint can do more harm than a large number of mild positives can undo.
How to act on it
Read a random sample of the labels yourself before you trust the trend. If the tool is misreading your customers, you will see it within a few minutes and can adjust the queries, the language filters or the tool. Then watch direction rather than the absolute number, because the number is only comparable with itself.
Break the results down by theme, channel and language, so the output points at something fixable. Route negative mentions to a named person with an agreed response time, and feed recurring complaints to whoever can change the product or the process — sentiment is a symptom, and the cure is operational. Where reputation is the business risk, treat social listening and structured reputation management as one workflow, not two dashboards.