NLP Metrics of Diversity in Fantasy Game of Thrones Polari

# Quantitative NLP Analysis of Lexical Diversity in Fantasy Novels

Comparison of the debut novels A Game of Thrones (IP, G.R.R. Martin, 1996) and Arrow, Coin, Spark (SMI, R. Surzhikov, 2016) using NLP tools reveals differences in lexical diversity. IP contains 240,796 words, SMI — 277,389. Average sentence length: 9.6 words in IP vs. 8.5 in SMI. The metrics examined include MATTR, MTLD, HD-D, and the author's KUL for an objective assessment of style.

Main Text Parameters

| Parameter | IP | SMI | Note |

|----------|----|-----|------|

| Pages | ~933 | ~1089 | formal |

| Sentences | 24,998 | 32,448 (+30%) | - |

| Words | 240,796 | 277,389 (+15%) | - |

| DET (emotional tone) | 3.5% | 10.5% (+200%) | !/? combinations |

| KUL | 13.50% | 16.72% (+24%) | per 50K words |

| MATTR | 70.58% | 72.16% | local diversity |

| HD-D | 91.93% | 92.97% | global diversity |

| IUFO | 63.6 | 63.5 | Flash-Oborenva |

| Dialogues | 37.6% | 35.6% | - |

KUL measures the cumulativity of unique lemmas: the ratio of unique forms to 50K words using multiple random samples. Accuracy ~99.5%, independent of text volume. MATTR focuses on local diversity in windows, HD-D — on global type density.

• Advantages of KUL: smooths nonlinearity, suitable for texts from 50K words.
• MATTR: sensitive to short sequences of repetitions.
• HD-D: accounts for all types but overestimates in large corpora.
• Hapax Index: dips in the middle for IP, at the end for SMI.

KUL Calculation Algorithm

KUL = (average unique lemmas over N samples of 50K words) / 50,000 × 100%.

• Lemmatize the text (using pymorphy2 or similar).
• Randomly select 50K word forms M times (M=1000 for accuracy).
• For each sample, count unique lemmas.
• Take the average and normalize.

This reflects the 'active vocabulary' — the author's ability to introduce new lemmas without depletion.

Comparison of Emotional Tone and Style

Share of emotional tone (DET): percentage of sentences with !, !!, ?!, etc. (excluding pure ?). SMI leads with 10.5% vs. 3.5% in IP. Comparison with classics:

• Tolstoy War and Peace: ~5.2%.
• Dostoevsky Crime and Punishment: ~18.2%.

Long sentences match in punctuation marks: 456 in both. Example from IP: description of a troop vanguard. From SMI: etiquette instructions.

Alphabet letter distribution is stable: 'o' ~10-11%, 'e'/'a' ~8%. 'Yo' is minimal. Zipf's law is not satisfied (a=-1.358 for IP, -1.296 for SMI).

Depth and Themes

Narrative depth: 3.87 (IP) vs. 3.72 (SMI). Themes by n-grams: nobility in both. Top names: Jon/Ned/Tyrion vs. Erwin/Mira/Harmon. Palindromes: 63 vs. 83 (+31%). SMI has 155 historical events, IP — zero chronology.

Bodily narrative is stronger in IP (+37%). Refrains: 'seventh hell' vs. 'darkness (devour)'.

What Matters

• KUL shows a larger lemma reserve for Surzhikov (16.72% vs. 13.50%), despite the authors' experience difference.
• MATTR/HD-D are close (72% vs. 70%, 93% vs. 92%), the metrics complement each other.
• Emotionality is 3 times higher in SMI, affecting the perception of dynamics.
• IUFO is identical (~63.5), readability at middle-grade level.
• Hapaxes indicate fatigue: middle in IP, end in SMI.

Practical Application in NLP

For prose analysis, combine metrics: KUL for global reserve, MATTR for local freshness. Python scripts with NLTK/spaCy for lemmatization. Test on 50K+ words for stability. Differences in fantasy highlight: Surzhikov's debut (29 years old, 2 awards) competes with Martin (48 years old, 5 awards) in diversity.

— Editorial Team