An authorship signature is a computed collection of textual features that represents how an author writes. It can characterize a known book or an unknown text for comparison. section 11.1 section 11.8
A signature can include measurements of word use and sentence structure:
- Average word length. section 11.8
- Different words divided by total words. section 11.5.10
- Words used exactly once divided by total words. section 11.5.10
- Average sentence length, measured in words per sentence. section 11.3.1
- Average sentence complexity, estimated from phrase separators such as commas, semicolons, and colons. section 11.3.1
Authorship identification builds known signatures from texts with known authors, creates an unknown signature for the mystery text, and chooses the closest known signature. The approach can misidentify an author or fail when the real author is not represented. section 11.1 section 11.2
A five-number signature
For the text “A pearl! Pearl! Lustrous pearl! Rare, what a nice find.”, make_signature returns [4.1, 0.7, 0.5, 2.5, 1.25]. The function combines the results of five separate feature functions rather than mixing all calculations together. section 11.5.10
Table 11.1 Calculating the difference between two signatures
|
Feature Number
|
Value of Feature in Signature 1
|
Value of Feature in Signature 2
|
Weight of Feature
|
Contribution of Feature
|
|---|---|---|---|---|
|
1
|
4.6
|
4.3
|
11
|
abs(4.6 – 4.3) × 11 = 3.3
|
|
2
|
0.1
|
0.1
|
33
|
abs(0.1 – 0.1) × 33 = 0
|
|
3
|
0.05
|
0.04
|
50
|
abs(0.05 – 0.04) × 50 = .5
|
|
4
|
10
|
16
|
0.4
|
abs(10 – 16) × 0.4 = 2.4
|
|
5
|
2
|
4
|
4
|
abs(2 – 4) × 4 = 8
|
|
Sum
|
14.2
|
Compare two signatures feature by feature, apply weights to their absolute differences, and sum the contributions into a comparison score. section 11.3.1
