Concept - Word Likelihood
Definition
The probability of seeing a word given that it occurs in a message of a certain class, (s.37-39). Estimated from a histogram of word counts of the class: count of the word divided by the total number of words in all messages of that class (s.40-41, 55-56). It is the per-word Concept - Likelihood used by Concept - Multinomial Naive Bayes.
Worked example (s.36-60)
Vocabulary: Dear, Friend, Lunch, Money.
| Word | Normal (total 17) | Spam (total 7) |
|---|---|---|
| Dear | 8/17 = 0.47 | 2/7 = 0.29 |
| Friend | 5/17 = 0.29 | 1/7 = 0.14 (slide 58 prints “0.29”, a typo) |
| Lunch | 3/17 = 0.18 | 0/7 = 0 |
| Money | 1/17 = 0.06 | 4/7 = 0.57 |
Each class is divided by its own total word count. The 0 for Lunch in spam illustrates the Concept - Zero Probability Problem.
Related
Concept - Prior Probability, Concept - Additive Smoothing, Concept - Naive Bayes, Concept - Bag of Words
Source chunk: 03 Text Classification - Chunk 02 (s.31-60)