Sentiment & Language Detection

Q: What are the possible values of Sentiment?

POSITIVE, NEGATIVE, NEUTRAL, and MIXED. MIXED means the text contains genuinely conflicting sentiment, not that the model is unsure.

Q: Can DetectDominantLanguage return more than one language?

Yes. Languages is a list, and text that mixes languages or is very short can return multiple candidates - pick the highest Score if you only need one.

Q: Is there a bulk way to score sentiment for many documents?

Yes - StartSentimentDetectionJob runs the same model asynchronously across an S3 prefix of documents and writes scored results to another S3 prefix.

DetectSentiment and DetectDominantLanguage are two of Comprehend's simplest pretrained calls, and two of the most commonly combined. Sentiment analysis needs to know what language it is scoring, and language detection is the fastest way to find that out when you don't already know it.

This page covers both operations on their own, then the two-step pattern of detecting language first and passing the result into sentiment.

Recipe

Quick-reference recipe card - copy-paste ready.

# --- Python (boto3) ---
import boto3
 
comprehend = boto3.client("comprehend")
resp = comprehend.detect_sentiment(
    Text="The onboarding flow was confusing but support fixed it fast.",
    LanguageCode="en",
)
print(resp["Sentiment"], resp["SentimentScore"])

// --- TypeScript (AWS SDK v3) ---
import { ComprehendClient, DetectSentimentCommand } from "@aws-sdk/client-comprehend";
 
const comprehend = new ComprehendClient({});
const resp = await comprehend.send(new DetectSentimentCommand({
  Text: "The onboarding flow was confusing but support fixed it fast.",
  LanguageCode: "en",
}));
console.log(resp.Sentiment, resp.SentimentScore);

When to reach for this:

Scoring reviews, support tickets, or survey free-text responses at submission time.
Routing negative or mixed feedback to a human queue automatically.
Handling input where the language isn't known up front and must be detected first.
Building dashboards that track sentiment trends over time across many short texts.

Working Example

Detect the language first, then feed the result into sentiment detection - the pattern for any pipeline that doesn't already know the input language.

# --- Python (boto3) ---
import boto3
 
comprehend = boto3.client("comprehend")
text = "El soporte respondio muy rapido, pero el problema volvio a aparecer."
 
lang_resp = comprehend.detect_dominant_language(Text=text)
top_language = max(lang_resp["Languages"], key=lambda l: l["Score"])
code = top_language["LanguageCode"]
print("detected:", code, top_language["Score"])
 
sentiment_resp = comprehend.detect_sentiment(Text=text, LanguageCode=code)
print(sentiment_resp["Sentiment"], sentiment_resp["SentimentScore"])

// --- TypeScript (AWS SDK v3) ---
import { ComprehendClient, DetectDominantLanguageCommand, DetectSentimentCommand } from "@aws-sdk/client-comprehend";
 
const comprehend = new ComprehendClient({});
const text = "El soporte respondio muy rapido, pero el problema volvio a aparecer.";
 
const langResp = await comprehend.send(new DetectDominantLanguageCommand({ Text: text }));
const topLanguage = (langResp.Languages ?? []).reduce((a, b) => (b.Score! > a.Score! ? b : a));
console.log("detected:", topLanguage.LanguageCode, topLanguage.Score);
 
const sentimentResp = await comprehend.send(new DetectSentimentCommand({
  Text: text,
  LanguageCode: topLanguage.LanguageCode,
}));
console.log(sentimentResp.Sentiment, sentimentResp.SentimentScore);

What this demonstrates:

DetectDominantLanguage needs no language input - it is the one Detect call that determines that for you.
Languages can return multiple candidates, so pick the highest Score when you only need one.
The detected LanguageCode feeds directly into DetectSentiment's required LanguageCode parameter.
Two calls, not one - there is no single operation that both detects language and scores sentiment.

Deep Dive

Reading a Sentiment Result

DetectSentiment returns a top-level Sentiment label - POSITIVE, NEGATIVE, NEUTRAL, or MIXED - and a SentimentScore object with a confidence value for all four, always summing close to 1.0. The label is simply whichever score is highest; if your use case cares about degree, not just direction, read SentimentScore directly rather than branching only on Sentiment.

MIXED is a real, distinct outcome, not a fallback for "unsure." Comprehend assigns it when a passage contains genuinely conflicting sentiment - praise and complaint in the same sentence, for example - and it is common in product reviews and support tickets. Treat it as its own routing bucket (often "needs human review") rather than discarding it or forcing it into positive or negative.

Reading a Language Result

DetectDominantLanguage returns Languages, a list of {LanguageCode, Score} pairs, potentially more than one for text that mixes languages or is too short to classify confidently. There is no LanguageCode parameter on the request - detecting it is the entire point of the call, which is a common point of confusion for people used to every other Detect operation requiring one.

Why LanguageCode Accuracy Matters

Sentiment models are language-specific under the hood. Passing an incorrect LanguageCode - English text tagged as Spanish, for instance - does not usually raise an error; it silently returns a lower-quality or misleading sentiment score, because the wrong language model was applied. This is why the detect-then-score pattern in the working example matters whenever the input language isn't already known with certainty, rather than defaulting every call to "en".

Confidence Is Not Certainty

Both operations return probabilistic scores, not ground truth. Very short text (a few words), sarcasm, and domain-specific slang all reduce reliability. For high-stakes routing decisions, treat a low top score - on either the language or the sentiment call - as a signal to fall back to a default path or human review rather than trusting a marginal result.

Gotchas

Defaulting every call to LanguageCode: "en" - text in another language gets scored through the wrong model with no error raised. Fix: run DetectDominantLanguage first whenever the input language isn't already known.
Ignoring MIXED as if it were an error - it is a valid, meaningful sentiment outcome. Fix: route MIXED results to a dedicated handling path instead of coercing them to positive or negative.
Reading only the Sentiment label and discarding SentimentScore - you lose the confidence margin between labels. Fix: check SentimentScore when a close call (for example Positive 0.52 vs Neutral 0.45) matters to your logic.
Trusting a single-word or very short DetectDominantLanguage result - confidence drops sharply on short strings. Fix: concatenate more context before detecting language, or treat low scores as "unknown" rather than authoritative.
Scoring sentiment per sentence and averaging - this can wash out a MIXED signal that only appears across the whole passage. Fix: score the full passage in one call when sentiment depends on overall context.

Alternatives

Alternative	Use When	Don't Use When
`DetectSentiment` / `DetectDominantLanguage` (this page)	Free text where the pretrained model's language coverage fits	You need a domain-specific sentiment scale (for example 1-5 stars mapped from text)
`StartSentimentDetectionJob` (batch)	Scoring a large archive of documents from S3	You need a result for one document immediately
Comprehend Custom classifier	Sentiment or category labels specific to your business, not the built-in four labels	The built-in Positive/Negative/Neutral/Mixed already fits
Manual keyword/rule scoring	Extremely narrow, well-known vocabulary	General free text with varied phrasing

FAQs

What are the possible values of Sentiment?

POSITIVE, NEGATIVE, NEUTRAL, and MIXED. MIXED means the text contains genuinely conflicting sentiment, not that the model is unsure.

Why does DetectSentiment need a LanguageCode but DetectDominantLanguage doesn't?

Sentiment scoring uses a language-specific model, so it needs to know which one to apply. Detecting the language is the entire purpose of DetectDominantLanguage, so it takes no language input.

What happens if I pass the wrong LanguageCode to DetectSentiment?

The call typically still returns a result, but accuracy degrades because the wrong language model was applied - there is usually no explicit error, which makes this an easy mistake to miss.

Can DetectDominantLanguage return more than one language?

Yes. Languages is a list, and text that mixes languages or is very short can return multiple candidates - pick the highest Score if you only need one.

Should I trust sentiment scores on very short text?

Be cautious. A few words gives the model less context, and both sentiment and language detection scores tend to be less reliable on very short strings.

Is there a bulk way to score sentiment for many documents?

Yes - StartSentimentDetectionJob runs the same model asynchronously across an S3 prefix of documents and writes scored results to another S3 prefix.

Does Comprehend support more than a handful of languages for sentiment?

Sentiment and language detection support a range of languages via LanguageCode, but coverage varies by operation - check the current service documentation for the exact supported list before depending on a specific language.

Comprehend Basics - the first sentiment and language calls.
Entity & Key Phrase Extraction - the other pretrained structured-signal APIs.
Comprehend's Managed NLP Model Surface - where these calls sit in the pretrained-vs-custom map.
Batch vs Real-Time Comprehend Jobs - scoring sentiment across an S3 archive.

Stack versions: This page was written for boto3 1.43.x (Python 3.10+) and the AWS SDK for JavaScript v3 (Node.js 18+).

This page covers both operations on their own, then the two-step pattern of detecting language first and passing the result into sentiment.

Recipe

Quick-reference recipe card - copy-paste ready.

# --- Python (boto3) ---
import boto3
 
comprehend = boto3.client("comprehend")
resp = comprehend.detect_sentiment(
    Text="The onboarding flow was confusing but support fixed it fast.",
    LanguageCode="en",
)
print(resp["Sentiment"], resp["SentimentScore"])

// --- TypeScript (AWS SDK v3) ---
import { ComprehendClient, DetectSentimentCommand } from "@aws-sdk/client-comprehend";
 
const comprehend = new ComprehendClient({});
const resp = await comprehend.send(new DetectSentimentCommand({
  Text: "The onboarding flow was confusing but support fixed it fast.",
  LanguageCode: "en",
}));
console.log(resp.Sentiment, resp.SentimentScore);

When to reach for this:

Scoring reviews, support tickets, or survey free-text responses at submission time.
Routing negative or mixed feedback to a human queue automatically.
Handling input where the language isn't known up front and must be detected first.
Building dashboards that track sentiment trends over time across many short texts.

Working Example

Detect the language first, then feed the result into sentiment detection - the pattern for any pipeline that doesn't already know the input language.

# --- Python (boto3) ---
import boto3
 
comprehend = boto3.client("comprehend")
text = "El soporte respondio muy rapido, pero el problema volvio a aparecer."
 
lang_resp = comprehend.detect_dominant_language(Text=text)
top_language = max(lang_resp["Languages"], key=lambda l: l["Score"])
code = top_language["LanguageCode"]
print("detected:", code, top_language["Score"])
 
sentiment_resp = comprehend.detect_sentiment(Text=text, LanguageCode=code)
print(sentiment_resp["Sentiment"], sentiment_resp["SentimentScore"])

// --- TypeScript (AWS SDK v3) ---
import { ComprehendClient, DetectDominantLanguageCommand, DetectSentimentCommand } from "@aws-sdk/client-comprehend";
 
const comprehend = new ComprehendClient({});
const text = "El soporte respondio muy rapido, pero el problema volvio a aparecer.";
 
const langResp = await comprehend.send(new DetectDominantLanguageCommand({ Text: text }));
const topLanguage = (langResp.Languages ?? []).reduce((a, b) => (b.Score! > a.Score! ? b : a));
console.log("detected:", topLanguage.LanguageCode, topLanguage.Score);
 
const sentimentResp = await comprehend.send(new DetectSentimentCommand({
  Text: text,
  LanguageCode: topLanguage.LanguageCode,
}));
console.log(sentimentResp.Sentiment, sentimentResp.SentimentScore);

What this demonstrates:

DetectDominantLanguage needs no language input - it is the one Detect call that determines that for you.
Languages can return multiple candidates, so pick the highest Score when you only need one.
The detected LanguageCode feeds directly into DetectSentiment's required LanguageCode parameter.
Two calls, not one - there is no single operation that both detects language and scores sentiment.

Deep Dive

Reading a Sentiment Result

Reading a Language Result

Why LanguageCode Accuracy Matters

Confidence Is Not Certainty

Gotchas

Defaulting every call to LanguageCode: "en" - text in another language gets scored through the wrong model with no error raised. Fix: run DetectDominantLanguage first whenever the input language isn't already known.
Ignoring MIXED as if it were an error - it is a valid, meaningful sentiment outcome. Fix: route MIXED results to a dedicated handling path instead of coercing them to positive or negative.
Reading only the Sentiment label and discarding SentimentScore - you lose the confidence margin between labels. Fix: check SentimentScore when a close call (for example Positive 0.52 vs Neutral 0.45) matters to your logic.
Trusting a single-word or very short DetectDominantLanguage result - confidence drops sharply on short strings. Fix: concatenate more context before detecting language, or treat low scores as "unknown" rather than authoritative.
Scoring sentiment per sentence and averaging - this can wash out a MIXED signal that only appears across the whole passage. Fix: score the full passage in one call when sentiment depends on overall context.

Alternatives

Alternative	Use When	Don't Use When
`DetectSentiment` / `DetectDominantLanguage` (this page)	Free text where the pretrained model's language coverage fits	You need a domain-specific sentiment scale (for example 1-5 stars mapped from text)
`StartSentimentDetectionJob` (batch)	Scoring a large archive of documents from S3	You need a result for one document immediately
Comprehend Custom classifier	Sentiment or category labels specific to your business, not the built-in four labels	The built-in Positive/Negative/Neutral/Mixed already fits
Manual keyword/rule scoring	Extremely narrow, well-known vocabulary	General free text with varied phrasing

FAQs

What are the possible values of Sentiment?

POSITIVE, NEGATIVE, NEUTRAL, and MIXED. MIXED means the text contains genuinely conflicting sentiment, not that the model is unsure.

Why does DetectSentiment need a LanguageCode but DetectDominantLanguage doesn't?

Sentiment scoring uses a language-specific model, so it needs to know which one to apply. Detecting the language is the entire purpose of DetectDominantLanguage, so it takes no language input.

What happens if I pass the wrong LanguageCode to DetectSentiment?

The call typically still returns a result, but accuracy degrades because the wrong language model was applied - there is usually no explicit error, which makes this an easy mistake to miss.

Can DetectDominantLanguage return more than one language?

Yes. Languages is a list, and text that mixes languages or is very short can return multiple candidates - pick the highest Score if you only need one.

Should I trust sentiment scores on very short text?

Be cautious. A few words gives the model less context, and both sentiment and language detection scores tend to be less reliable on very short strings.

Is there a bulk way to score sentiment for many documents?

Yes - StartSentimentDetectionJob runs the same model asynchronously across an S3 prefix of documents and writes scored results to another S3 prefix.

Does Comprehend support more than a handful of languages for sentiment?

Comprehend Basics - the first sentiment and language calls.
Entity & Key Phrase Extraction - the other pretrained structured-signal APIs.
Comprehend's Managed NLP Model Surface - where these calls sit in the pretrained-vs-custom map.
Batch vs Real-Time Comprehend Jobs - scoring sentiment across an S3 archive.

Stack versions: This page was written for boto3 1.43.x (Python 3.10+) and the AWS SDK for JavaScript v3 (Node.js 18+).

Sentiment & Language Detection

Search Documentation

Sentiment & Language Detection