---
url: https://findutils.com/blog/why-not-happy-can-still-read-as-joy-how-keyword-tone-and-sentiment-analysis-works
title: "Why \"Not Happy\" Can Still Read as Joy: How Keyword Tone and Sentiment Analysis Works"
description: "Keyword-based tone and sentiment tools count feeling words from a lexicon. Here is how that counting works, why negation and sarcasm trip it up, what rule-based tools like VADER add, and how to read the scores."
category: text_tools
content_type: blog
cluster: text
locale: en
read_time: 7
status: published
author: "olgunozoktas"
published_at: 2026-11-13T08:22:00Z
excerpt: "Most quick tone checkers work the same way: they look up each word in a list of feeling words and count. That makes them fast, private and easy to explain, and it is also why \"I am not happy\" can score as joy. This post walks through how word-list scoring works, which public lexicons exist, what rules like negation and intensifiers add, and how to read a tone score without being fooled by it."
tag_ids: ["text-tools", "sentiment-analysis", "tone-analysis", "nlp", "ai"]
tags: ["Text Tools", "Sentiment Analysis", "Tone Analysis", "NLP", "AI"]
primary_keyword: "how sentiment analysis works"
secondary_keywords: ["lexicon based sentiment analysis", "keyword tone analysis", "sentiment analysis negation", "why sentiment analysis fails on sarcasm", "vader sentiment explained", "rule based sentiment analysis"]
tool_tag: "tone-analyzer"
related_tool: "tone-analyzer"
related_tools: ["tone-analyzer", "readability-calculator", "text-summarizer", "word-counter"]
updated_at: "2026-11-13T08:22:00Z"
og_image: "/images/content/blog/why-not-happy-can-still-read-as-joy-how-keyword-tone-and-sentiment-analysis-works-cover-20261113.webp"
image_alt: "A row of rounded yellow and teal tiles on a dark background, with a red tile showing a no symbol placed just before a yellow tile with a heart, which glows under a magnifying glass."
---

A keyword tone checker decides how a text feels by looking up each word in a list of feeling words, called a **lexicon**, and counting the matches. "Happy" and "love" count towards joy, "furious" and "hate" towards anger, "maybe" and "might" towards uncertainty. The tone with the most matches wins. This is fast, private and easy to explain, and it is also why **"I am not happy and I do not love it"** can come out as joy: the counter sees "happy" and "love" and has no idea that "not" changed their meaning.

## Step One: A List of Feeling Words

Every lexicon-based tool starts with word lists. Some well-known public ones:

| Lexicon | Size | What each word gets |
| --- | --- | --- |
| Bing Liu's Opinion Lexicon | Around 6,800 words | Positive or negative |
| NRC Emotion Lexicon (EmoLex) | 14,182 words | Any of eight emotions (anger, fear, anticipation, trust, surprise, sadness, joy, disgust) plus positive or negative, labelled by crowdsourcing |
| VADER | Just over 7,500 entries, including slang and emoticons | A score from −4 to +4, the average of ten human raters |

The difference matters. A positive/negative list only tells you the direction. An emotion list such as EmoLex separates anger from fear from sadness. A scored list such as VADER's knows that "great" (3.1) is stronger than "good" (1.9), which is stronger than "okay" (0.9).

## Step Two: Count, Then Compare

The simplest scorer, sometimes called a **bag of words**, ignores word order completely. Take this support message:

> "Unfortunately the shipment is late and we are disappointed with the delay. Thank you for your patience."

| Word list | Matches | Count |
| --- | --- | --- |
| Sadness | unfortunately, disappointed | 2 |
| Polite | thank | 1 |
| Anger, fear, joy | none | 0 |

Sadness wins, so the dominant tone is sadness. Many tools then show each tone as a share of the strongest one: sadness 100%, polite 50%. For an overall sentiment, positive-leaning matches are weighed against negative-leaning ones. Here two negative words against one positive word gives a negative result.

For ordinary business writing, this works better than you might expect. People who are unhappy tend to use unhappy words. A reply full of "regret" and "unfortunately" really does read as bad news, and a proposal with "maybe" in every sentence really does sound unsure.

## Where Counting Breaks

Counting fails whenever the meaning comes from how words combine rather than from the words themselves:

| Text | What a word counter sees | What a reader sees |
| --- | --- | --- |
| "I am not happy with this." | "happy": joy | Unhappy |
| "Oh great, another meeting that could have been an email." | "great": joy | Annoyed |
| "Not bad at all." | "bad": negative, or nothing if "bad" is not listed | Mildly positive |
| "The food was great but the service was awful." | One positive, one negative: neutral | Mostly negative, because the "but" clause usually carries the verdict |
| "I am VERY disappointed!!!" | "disappointed": sadness, same as without caps | Much stronger than plain "disappointed" |

Two smaller problems come from how words are matched. If a tool matches word parts to catch "thanks" and "thankful", it can also catch "wonder" inside "wonderful" or "sure" inside "pressure". And a word list is written in one language: text in another language usually comes out neutral.

## What Rule-Based Tools Add

Rule-based analyzers keep the lexicon but add rules about word order. VADER, published by Hutto and Gilbert in 2014, is the best-known example. Its rules cover:

- **Negation**: "not good", and contractions such as "wasn't very good", flip or reduce the score.
- **Intensifiers and dampeners**: "extremely good" scores higher than "good", "marginally good" lower. The paper measured the effect with human raters.
- **Emphasis**: ALL CAPS words and exclamation marks increase intensity.
- **Contrast**: after "but", the second clause gets more weight.
- **Slang and emoticons**: entries such as "sux", ":(" and ":D" are in the lexicon.

The VADER project shows the difference with its own examples. "Not bad at all" gets a compound score of 0.431, which is positive. "The plot was good, but the characters are uncompelling and the dialog is not great." gets −0.7042, negative, even though it contains "good" and "great". The compound score runs from −1 to +1, and the project suggests treating 0.05 or above as positive, −0.05 or below as negative, and anything in between as neutral.

Rules still have limits. Sarcasm such as "Oh great, another meeting" has no grammatical marker, so a rule set cannot see it. Detecting it reliably needs context: who is speaking, about what, and what usually happens next. That is the territory of trained machine-learning models, which learn patterns from large sets of labelled examples instead of fixed lists, at the cost of being harder to explain and usually needing more computing power.

## How to Read a Tone Score

1. **Treat it as a signal, not a verdict.** A score tells you which feeling words are present. It is a prompt to reread the sentence, not a judgment on the message.
2. **Check every sentence with "not", "never" or "no" yourself.** That is where a counting tool is most likely to be wrong.
3. **Look at the mix, not just the winner.** "Sadness plus polite" is the classic shape of a well-written apology. "Anger plus nothing else" is worth a second look before you hit send.
4. **Use it on drafts you wrote, in the language it was built for.** You know what you meant, so you can spot a false reading immediately.
5. **Prefer neutral for facts.** A status update that scores neutral is usually doing its job.

## Trying It on Your Own Text

The FindUtils [Tone Analyzer](/text/tone-analyzer/) is a keyword tool of exactly this kind: it uses an English lexicon and simple rules, runs in your browser and shows a dominant tone, an overall sentiment, a formal or casual writing style, and each detected tone as a share of the strongest one. Its method note says to read the scores as signals, not verdicts, and the examples above show why. The [Tone Analyzer guide](/guides/tone-analyzer-guide/) has tested examples of every result. For the rest of the editing pass, the [Readability Calculator](/text/readability-calculator/) checks how easy the text is to read and the [Text Summarizer](/text/text-summarizer/) shortens it.

---

Need help with a project? The team behind FindUtils takes on freelance work through [Upwork](https://www.upwork.com/freelancers/~010a5875eab37b4bd7), including [an AI feature in your web app: a chat assistant, document analysis or search](https://www.upwork.com/services/product/development-it-an-ai-feature-in-your-web-app-chat-assistant-document-analysis-or-search-2112348761417603851), for example sentiment tagging for customer feedback that understands negation and context.

## Related Reading

- [Tone Analyzer guide: check the tone and sentiment of your writing](/guides/tone-analyzer-guide/)
- [Extractive summarization explained: how a text summarizer picks sentences](/blog/extractive-summarization-explained-how-a-text-summarizer-picks-sentences/)
- [Readability Calculator guide](/guides/readability-calculator-guide/)
- [Word Counter guide](/guides/word-counter-guide/)

## Sources

- [Hutto and Gilbert, VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text (ICWSM 2014)](https://ojs.aaai.org/index.php/ICWSM/article/view/14550)
- [VADER Sentiment Analysis project (lexicon, rules, example scores and thresholds)](https://github.com/cjhutto/vaderSentiment)
- [NRC Word-Emotion Association Lexicon (EmoLex), National Research Council Canada](https://saifmohammad.com/WebPages/NRC-Emotion-Lexicon.htm)
- [Bing Liu, Opinion Mining, Sentiment Analysis, and Opinion Spam Detection (Opinion Lexicon)](https://www.cs.uic.edu/~liub/FBS/sentiment-analysis.html)
