Is a Potts diagram just a raw frequency chart?

The normalized-likelihood version described here first adjusts for the amount of text within each rating category.

Does a high point identify a particular customer's emotion?

No. It describes aggregated word usage in a labeled corpus.

Can a word with no occurrences have this profile?

Not under the unsmoothed formula, because its normalization denominator is zero.

AI Glossary / Natural language and speech / Intermediate

Potts diagram

Definition

A Potts diagram is a visualization of a word's association with ordered rating categories, such as review scores. In the sentiment-analysis sense used here, it plots a normalized profile derived from the word's frequency within each category. The shape can reveal distinctions that a single positive-or-negative label would hide.

What it means

The horizontal axis represents rating categories in their proper order. The vertical axis represents a specified association score. In the normalized-likelihood construction described in Speech and Language Processing, the word's within-category frequency is divided by the sum of those frequencies across categories.

This is a corpus-level description of word usage. It does not assign an emotion to an individual writer or guarantee the sentiment of every sentence containing the word. The term belongs here to lexical sentiment analysis, rather than diagrams associated with physical Potts models.

How it works

Begin with texts that have rating labels and a consistent tokenization policy. Count the chosen word in each category and divide by the total token count for that category. This produces comparable within-category rates even when the categories contain different amounts of text.

Next, divide each rate by the sum of the rates across the selected categories, provided that sum is positive. Plot the resulting values in rating order. Keeping the calculation visible is essential: raw counts and normalized rates answer different questions.

For an original analysis project, also retain the underlying counts and document how word forms were combined. A smooth plotted line can hide sparse observations. The plot connects category values for readability; it does not create observations at ratings that the dataset never used. Investigate unexpected shapes by reading representative source passages and checking whether repeated templates, domain vocabulary, or a small number of contributors dominate the evidence.

Three practical examples

1. Review exploration: An analyst compares profiles for candidate descriptive words before deciding whether a simple sentiment label loses useful information.

2. Lexicon review: A language team investigates a word whose profile differs across two product domains, checking actual contexts before changing its sentiment resource.

3. Reporting quality: A dashboard designer displays the calculation and supporting counts so readers can distinguish a normalized association profile from raw review volume.

Think of it this way

Think of the diagram as a word's usage silhouette across a rating scale. Two words can lean toward the same end yet have different silhouettes. The analogy has limits: the outline depends on the sampled texts and the scoring method, and it is not an intrinsic, permanent measurement of what the word means everywhere.

Worked example

Consider an original toy corpus with three rating categories: low, middle, and high. Their total token counts are 1,000, 2,000, and 4,000. An invented target term occurs 10, 20, and 40 times respectively.

The raw counts increase across categories, but the within-category rates are identical: 10/1,000=0.01, 20/2,000=0.01, and 40/4,000=0.01. Their sum is 0.03, so the normalized profile is [1/3,1/3,1/3]. The resulting diagram is flat. The apparent upward pattern in raw counts came entirely from differing category sizes.

Now consider a second term with counts 2, 8, and 32. Its rates are [0.002,0.004,0.008], summing to 0.014. Its normalized profile is [1/7,2/7,4/7], approximately [0.143,0.286,0.571]. That profile rises toward the high category.

Both calculations are independently checked. They are invented examples, not reported language findings. If a term has zero occurrences everywhere, this unsmoothed calculation divides by zero and provides no profile; the correct response is to mark insufficient evidence or explicitly choose another estimation procedure.

Illustrative case study

Fictional case: Harbor Appliances plots word counts from customer reviews and sees many terms rising toward the highest score. Its analyst notices that high-score reviews also contribute much more text. The team recomputes within-category rates before normalizing each word's profile.

Some apparent trends flatten, prompting a review of the original dashboard labels. The team keeps counts beside the profiles and reads examples before making claims about particular words. This case illustrates a denominator check; it does not claim that the invented corpus represents real customer behavior or that normalization resolves every sampling problem.

Common misconceptions and mistakes

Do not label the normalized likelihood profile as the observed probability that a review has a particular rating given a word. Those quantities generally differ when category base rates differ. In the first toy example, the term's occurrences are split 10,20,40 across categories, while its normalized rate profile is flat.

Also avoid treating the shape as a sentence-level sentiment decision. Negation, quotation, and the subject being evaluated can change an individual use. The profile summarizes an association under a defined sampling and counting procedure; it does not explain why the association arose.

Business use

For business use, the diagram is an exploratory aid for examining language in rated feedback. State the rating scale, time period, tokenization, and normalization alongside it. Preserve enough supporting evidence to investigate a pattern, and separate exploratory observations from decisions about customers or changes to an operational sentiment model.

Check your understanding

Counts are 10,20,40 in categories with 1,000,2,000,4,000 tokens. Does the normalized profile rise?

Answer: No. Each within-category rate is 0.01, so all three normalized values are one third. Increasing raw counts alone do not establish increasing association.

Related terms

Rating scale
Conditional probability
Normalization
Corpus
Sentiment analysis
Sentiment lexicon
word frequency
lexical semantics

Learn next

Learn the difference between counts and rates, then reproduce the toy calculation with unequal category sizes. Next compare a word's normalized likelihood profile with its observed distribution of occurrences across categories. Finally, inspect real contexts before interpreting a curve as evidence about sentiment.

Sources

Jurafsky and Martin, Speech and Language Processing, January 2025 chapter 22, section 22.5 and equation 22.6, consulted for the specific term and normalization: https://web.stanford.edu/~jurafsky/slp3/old_jan25/22.pdf

The textbook attributes the visualization to Christopher Potts (2011). That original workshop presentation was not independently inspected. All numerical examples, diagnostic suggestions, and the company case here are original teaching material.

Previous: Potential scale reduction factor Next: Potts model