Social Media

Sentiment Analysis

Also called Opinion mining, tone analysis

The classification of mentions as positive, negative or neutral so a brand can see tone of conversation, not just volume.

Quick facts: Sentiment Analysis

Category
Social Media
Also called
Opinion mining, tone analysis
Level
Advanced
Affects
Reputation monitoring, crisis response, product feedback
Where to see it
Social listening platforms, platform comment inboxes, review dashboards
In this article4
  1. How sentiment analysis works
  2. Why sentiment analysis matters
  3. Where sentiment analysis goes wrong
  4. How to act on it

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.

Do and do not

Do

  • Read a sample of labels before trusting the score
  • Track the direction of sentiment, not the absolute number
  • Split sentiment by theme, channel and language

Do not

  • Do not compare sentiment scores from different tools
  • Do not treat neutral as a meaningful group
  • Do not report a score without reading the mentions

Questions people ask about this

How accurate is sentiment analysis?

Accurate enough to show direction, not accurate enough to quote as fact. Tools handle plain, clearly worded English best and struggle with sarcasm, slang, mixed-language comments and sentences carrying both praise and complaint. Read a sample of the labels yourself before trusting a score, and treat every reading as an estimate that still needs a human check.

Can it handle Nepali or mixed-language comments?

Less well than English. Most tools are trained mainly on English text, so Devanagari Nepali, Romanised Nepali and comments that switch between the two are often mislabelled or dumped into neutral. If a large part of your audience writes that way, budget time for someone to read the mentions rather than relying on the dashboard figure.

What should I do when sentiment turns negative?

Find the cause before reacting to the number. Filter negative mentions by theme and by date to see whether one incident, one product or one channel is driving the change. Reply in public where a reply helps other readers, move the detail into private messages, and fix the underlying problem. The score follows the operation, not the other way round.

Related terms

Found this useful?

Share it, or ask an AI to summarise it

Back to the glossary

Knowing the term is the easy part

Applying it to your own site and budget is the work. Book a call and I will tell you what actually applies to you.