Overview
Sentiment Analysis runs two independent scorers over the `text` you POST. The first is lexicon-based: the text is tokenised, each token is looked up in the AFINN list of integer valences from -5 to +5, and the matches are summed into `sentiment`. `comparative` divides that sum by the token count. `sentimentText` maps the sum to a label with fixed cut-offs: 4 and above is very positive, 2 to 3 positive, -1 to 1 neutral, -2 negative, and -3 or below very negative. `isPositive` and `isNegative` are only the sign of the sum, so a score of 1 returns `isPositive: true` while the label still reads neutral.
On paid plans the text is also split into sentences at terminal punctuation and line breaks, and the first 50 sentences go to a transformer classifier trained on social media text. Each sentence gets positive, neutral and negative probabilities. `modelScore` is the mean positive probability minus the mean negative probability, rounded to two decimals, and `modelSentiment` is the label with the highest mean. `isMixed` is true when at least one sentence is labelled positive and another negative, each with confidence of 0.6 or more; `modelSentiment` then reads `mixed`.
`normalizedScore` is the AFINN sum divided by 10 and clamped to -1..1, so any text scoring 10 or more in either direction saturates. `magnitude` is the sum of absolute valences and grows with length: use it to tell strongly worded text from flat text, not to compare documents of different sizes. The free plan skips the model call entirely, so the model fields are null there.
Live Test Analyze Sentiment Source →
Formula
One function covers the whole catalog: the first argument names the source, the next are its inputs, and the field after them says which single value lands in the cell. Formulas covers the grammar that applies to all of them.
=VERVE("sentimentanalysis", A2, "comparative")=VERVE.CALL("sentimentanalysis", A2, "comparative")A2 holds the text you are looking up. Drag the formula down and each row resolves on its own.
With literal values
Nothing has to come from a cell — typed values work the same way.
=VERVE("sentimentanalysis", "I absolutely love this product! It exceeded all my expectations.", "comparative")=VERVE.CALL("sentimentanalysis", "I absolutely love this product! It exceeded all my expectations.", "comparative")Inputs
Required inputs are positional, in the order below, followed by the field. Everything else is passed as a "name", value pair after that — the same shape as SUMIFS. Premium inputs are accepted on every plan but only take effect on plans that include them.
| Input | Type | Where it goes | Description |
|---|---|---|---|
textRequired | string | argument 2 (A2) | The text to analyze the sentiment of length 0–50000 |
What lands in your sheet
The field — the argument after the inputs — is what makes the formula resolve to one cell. Its value is a dot-path from the table below, so swapping it is how you pull a different value — one formula per column.
Seeing everything at once
Leave the field out and the formula returns the whole response instead: two columns, field name on the left and value on the right, spilling from the cell you typed in.
=VERVE("sentimentanalysis", A2)=VERVE.CALL("sentimentanalysis", A2)The second row's table would land on top of the first's, and every cell after the first reads #REF! (#SPILL! in Excel). Keep one per sheet, or name a field, and leave the cells below and to the right empty.
Response fields
Paths are relative to data. Each one is exactly what the field argument accepts, so swapping it is how you pull a different value into a cell. On a plan without a Premium field the cell reads #PREMIUM(field) rather than a value, so a locked field never looks like a real answer.
| Use as the field | Type | Example cell value | Description |
|---|---|---|---|
comparative | number | 0.25 | The sentiment score divided by the number of words, so long and short texts compare |
sentimentText | string | positive | The score read as a label: very positive, positive, neutral, negative or very negative |
sentiment | number | 3 | AFINN sentiment score: the summed valence of the matched words, negative for negative sentiment |
isPositive | boolean | true | Whether the sentiment is positive |
isNegative | boolean | false | Whether the sentiment is negative |
normalizedScorePremium | number | 0.3 | Sentiment score normalized to -1 to 1 range |
magnitudePremium | number | 3 | Total emotional strength: the summed absolute valence of every matched word, so strongly mixed text scores high even when sentiment nets to zero |
positiveWordsPremium | array | ["excited"] | The distinct words that pushed the AFINN score up |
negativeWordsPremium | array | [] | The distinct words that pulled the AFINN score down, including positive words negated by not or never |
modelSentimentPremium | string | positive | Overall label from a transformer sentiment model read sentence by sentence: positive, neutral, negative, or mixed when the text has a confident positive and a confident negative sentence; null if the model is unavailable |
modelScorePremium | number | 0.98 | Model score from -1 (negative) to 1 (positive): the mean positive minus the mean negative probability across sentences |
isMixedPremium | boolean | false | True when at least one sentence is confidently positive and another confidently negative (model confidence 0.6 or higher) |
sentencesPremium | array[2] | Per-sentence model results for the first 50 sentences | |
sentences.0.textPremium | string | I'm so excited that tomorrow is going to be sunny! | The sentence as split from the input |
sentences.0.labelPremium | string | positive | positive, neutral or negative |
sentences.0.confidencePremium | number | 0.992 | Model probability of the label, 0 to 1 |
sentences.0.scoresPremium | object | {…} | Probability of each label |
sentences.0.scores.positivePremium | number | 0.992 | Probability the sentence is positive |
sentences.0.scores.neutralPremium | number | 0.006 | Probability the sentence is neutral |
sentences.0.scores.negativePremium | number | 0.002 | Probability the sentence is negative |
Filling a whole column
Both platforms make one batched call for a column rather than one request per cell, and both cost the same — 10 credits per row looked up. How you write it differs, because the two runtimes batch at different points.
=VERVE("sentimentanalysis", A2:A100, "comparative")In Google Sheets, give the arguments ranges and the answers spill down from the cell you typed in. Custom functions there run synchronously and in isolation, so dragged cells cannot be coalesced — the range form is how a column becomes one call. The field is required here: a per-row two-column table has nowhere to spill. Use a full column per input (A2:A100, not A2) — a single-cell reference is read as row 1 only and the rest of the column comes back blank.
In Excel, type the formula once and fill down — no range form and none needed, because its runtime is asynchronous and the add-in coalesces the calls a fill produces into a single batched request. Do not hand VERVE.CALL a range: Excel flattens it into positional arguments, so A2:A100 arrives as a hundred separate arguments and the cell reads #ERR rather than filling.
Either way, column A holds your text values and the answers land beside them, one row each, blank rows skipped.
When a cell doesn't fill
A failed lookup returns readable text starting with #ERR rather than a spreadsheet error, because Excel cannot show a custom message on a real error — you would get a bare #VALUE! with no reason. The trade-off is that IFERROR() will not catch it; test with LEFT(cell, 4) = "#ERR" instead. Error handling covers the rest.
| Cell reads | What happened |
|---|---|
#ERR Not connected | No API key saved. Open the side panel and connect your account. |
#ERR Replace <name> with a cell or value | A preview formula was copied as-is. Swap the placeholder for a cell reference. |
#ERR No "x" in sentimentanalysis | An option name this source does not take. The message lists the ones it does. |
#ERR No data point "…" | The field path is not in the response. Check it against the table above. |
#PREMIUM(field) | The field exists but your plan does not include it. |
#ERR Out of credits | The cycle's credits are spent. Usage resets on your billing date. |
Before the formula works
The add-in reads the API key you saved once in the side panel — the key never goes in the formula, so a shared sheet does not leak it and a collaborator without access simply sees the last calculated values.
- Install the add-in for Google Sheets or Excel.
- Open the side panel and sign in, or paste a key from your dashboard.
- Type the formula above into any cell.
Use cases
- Support Ticket Escalation
- Route angry customer messages to senior support reps immediately when the sentiment score drops to very negative.
- Review Stream Filtering
- Product managers categorize incoming app store feedback into positive praise and negative complaints using the five-tier sentiment label.
- Social Media Sentiment Analysis
- When tracking brand tags on social networks, marketing teams flag sudden spikes in negative valence scores to address PR issues early.
- Survey Response Sorting
- Sort open-ended employee survey responses automatically by comparative sentiment to isolate recurring workplace grievances without manual reading.
Other ways to use Analyze Sentiment
Set up Analyze Sentiment on VerveSheets, or reach the same source a different way. Your VerveSheets account and credits work on all of them — one key, one balance.
Related
More in AI/Computer Vision: