Reading media sentiment the way a comms team does
"The company announced its quarterly results." A generic sentiment model often flags that as negative. A communications professional knows it's neutral reporting. That gap — between how a machine reads tone and how a comms team reads it — is where most sentiment tools fall down, and it's why so many teams quietly stop trusting the sentiment column in their reports. Once a column has cried wolf a few dozen times, people stop reading it, and a genuinely hostile week slips past unnoticed.
Why generic models miss
Consumer-review sentiment models are trained on opinions — a one-star rant, a five-star rave. Business and media copy is different: much of it is factual, and the tone that matters is the tone toward your brand specifically, not the overall mood of the article. A model built for product reviews sees words like "loss", "cut", or "falls" and reaches for negative, even when the sentence is a neutral statement of fact or is about someone else entirely. The vocabulary of business reporting is full of words that read as bad news in a review and as plain description in a market report, and a model that can't tell the two apart will mislabel a large share of perfectly routine coverage.
Neutral is not negative
Most media coverage is neither praise nor attack — it's reporting. "X reported a 4% decline in quarterly revenue" is a fact, not an insult. If your tool marks every mention of a hard number as negative, your sentiment trend becomes noise, and the one week that genuinely turns hostile gets lost in a sea of false negatives. Reading tone the way a comms team does starts with a simple rule: factual is neutral until something in the framing makes it otherwise. A model that can't tell reporting from criticism will always over-report negativity, and over-reported negativity is indistinguishable from no signal at all.
The cost of a noisy sentiment column
There's a real, compounding price to a sentiment score that's wrong a third of the time. Teams learn to ignore it, so it stops informing decisions. Worse, when someone does act on it — escalating to leadership over a "negative spike" that turns out to be routine earnings coverage — the whole metric loses credibility, and with it the genuinely useful early warnings it might have given. A sentiment column is only worth having if people trust it enough to act, and trust is earned by not being wrong about the easy cases.
Attributing sentiment to the right brand
An article can praise one company while criticising another in the same paragraph. Scoring the article rather than the brand means a rival's bad week can drag down your numbers for no reason. Sentiment worth reporting is attributed: it measures the tone directed at the brand in question, so a negative story about someone else doesn't land on your dashboard. This is the core of how InMedia's sentiment analysis is designed to work — the unit of measurement is your brand's treatment, not the article's overall weather.
Reading each language natively
Translating an article to English before scoring it strips out exactly the nuance that decides tone — idiom, hedging, the difference between a pointed verb and a neutral one. Coverage should be read in the language it was written in, so a subtle criticism in a regional-language paper isn't flattened into a neutral machine translation. Much of the tone in reporting lives in word choice that translation smooths away; by the time a critical piece has been round-tripped through English, the very signal you needed has often been lost.
Reputational risk vs a bad tone
Not every negative is equal. A mildly critical trade note is not the same as a coordinated wave of hostile national coverage. Separating everyday negative tone from genuine reputational risk is what lets a team react proportionately — a distinction that matters most when a story is starting to escalate, which is where crisis monitoring picks up. Treating every negative mention as a five-alarm fire is its own failure mode: it burns out the team and buries the one story that truly warranted the alarm.
When to trust the number, and when to read the article
Even a well-tuned model is a triage tool, not a final verdict. The right posture is to trust the aggregate trend — is the line moving, and in which direction — while still opening the individual articles behind any sharp move before you act. The score tells you where to look; your own reading tells you what it means. A sentiment system that encourages people to read the coverage it flags is doing its job; one that's treated as gospel is being misused. In practice the best-run teams use the aggregate to set the agenda for a weekly review and the underlying articles to decide what, if anything, to do — the number narrows the field, the reading makes the call.
The human-in-the-loop flywheel
The goal isn't a perfect score on a benchmark. It's a tone reading a comms team would actually agree with — and the fastest route there is correction. Every time a human overrides a classification, that judgement should feed back and sharpen the model for the next article. Over months, a system tuned this way stops arguing with your team and starts matching them, building a labelled record of your coverage and your judgements that no off-the-shelf model could have shipped with. That's the real test: not accuracy against an academic dataset, but agreement with the people who read this coverage for a living. And it compounds — every correction is an asset, so the longer a well-run programme operates, the further its sentiment reading pulls ahead of anything a generic classifier could offer out of the box. If that's the kind of measurement you want running continuously, our plans show where it fits.