RESEARCH_REPORT.md

Italian — Full Ablation Study & Research Report

Detailed evaluation of all model variants trained on Italian Wikipedia data by Wikilangs.

👈 Back to README

📋 Repository Contents

Models & Assets

• Tokenizers (8k, 16k, 32k, 64k)
• N-gram models (2, 3, 4, 5-gram)
• Markov chains (context of 1, 2, 3, 4 and 5)
• Subword N-gram and Markov chains
• Embeddings in various sizes and dimensions (aligned and unaligned)
• Language Vocabulary
• Language Statistics

Analysis and Evaluation

• 1. Tokenizer Evaluation
• 2. N-gram Model Evaluation
• 3. Markov Chain Evaluation
• 4. Vocabulary Analysis
• 5. Word Embeddings Evaluation
• 6. Morphological Analysis (Experimental)
• 7. Summary & Recommendations
• Metrics Glossary
• Visualizations Index

---

1. Tokenizer Evaluation

Results

Vocab Size	Compression	Avg Token Len	UNK Rate	Total Tokens
8k	3.856x	3.86	0.1569%	3,369,266
16k	4.248x	4.25	0.1729%	3,057,865
32k	4.578x	4.58	0.1863%	2,837,619
64k	4.817x 🏆	4.82	0.1960%	2,696,638

Tokenization Examples

Below are sample sentences tokenized with each vocabulary size:

Sample 1: Eventi, invenzioni e scoperte Viene inventato il Lapis Benjamin Franklin inventa...

Vocab	Tokens	Count
8k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁viene ▁inv entato ▁il ▁la ... (+29 more)	39
16k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁viene ▁inventato ▁il ▁la pis ... (+21 more)	31
32k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁viene ▁inventato ▁il ▁la pis ... (+17 more)	27
64k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁viene ▁inventato ▁il ▁la pis ... (+17 more)	27

Sample 2: Eventi, invenzioni e scoperte Roma - Inaugurazione del Colosseo Personaggi 81 Ro...

Vocab	Tokens	Count
8k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁roma ▁- ▁inaugu razione ▁del ... (+19 more)	29
16k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁roma ▁- ▁inaugu razione ▁del ... (+19 more)	29
32k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁roma ▁- ▁inaugurazione ▁del ▁colo ... (+18 more)	28
64k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁roma ▁- ▁inaugurazione ▁del ▁colosseo ... (+16 more)	26

Sample 3: Eventi, invenzioni e scoperte Fine della cattività avignonese A Vicenza venne sp...

Vocab	Tokens	Count
8k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁fine ▁della ▁ca tti vità ... (+23 more)	33
16k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁fine ▁della ▁ca ttività ▁avi ... (+22 more)	32
32k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁fine ▁della ▁ca ttività ▁avi ... (+22 more)	32
64k	▁eventi , ▁invenzioni ▁e ▁scoperte ▁fine ▁della ▁cattività ▁avignon ese ... (+18 more)	28

Key Findings

• Best Compression: 64k achieves 4.817x compression
• Lowest UNK Rate: 8k with 0.1569% unknown tokens
• Trade-off: Larger vocabularies improve compression but increase model size
• Recommendation: 32k vocabulary provides optimal balance for production use

---

2. N-gram Model Evaluation

Results

N-gram	Variant	Perplexity	Entropy	Unique N-grams	Top-100 Coverage	Top-1000 Coverage
2-gram	Word	204,245	17.64	1,475,040	6.1%	16.8%
2-gram	Subword	214 🏆	7.74	16,385	74.4%	99.4%
3-gram	Word	980,193	19.90	3,253,104	3.2%	7.9%
3-gram	Subword	1,722	10.75	125,759	29.1%	78.9%
4-gram	Word	1,937,953	20.89	4,769,877	3.5%	7.0%
4-gram	Subword	10,064	13.30	693,084	14.1%	42.3%
5-gram	Word	1,090,157	20.06	2,714,110	5.0%	9.7%
5-gram	Subword	43,596	15.41	2,255,172	7.7%	24.4%

Top 5 N-grams by Size

2-grams (Word):

Rank	N-gram	Count
1	per la	97,286
2	è un	96,027
3	di un	94,634
4	e il	87,633
5	altri progetti	84,533

3-grams (Word):

Rank	N-gram	Count
1	altri progetti collegamenti	60,396
2	progetti collegamenti esterni	60,395
3	è un comune	49,422
4	note altri progetti	43,464
5	società evoluzione demografica	42,434

4-grams (Word):

Rank	N-gram	Count
1	altri progetti collegamenti esterni	60,395
2	è un comune francese	33,569
3	abitanti situato nel dipartimento	33,506
4	un comune francese di	33,266
5	società evoluzione demografica note	32,814

5-grams (Word):

Rank	N-gram	Count
1	è un comune francese di	33,067
2	evoluzione demografica note altri progetti	32,558
3	società evoluzione demografica note altri	32,545
4	note altri progetti collegamenti esterni	27,862
5	demografica note altri progetti collegamenti	18,549

2-grams (Subword):

Rank	N-gram	Count
1	e _	14,124,940
2	a _	13,040,624
3	i _	10,977,304
4	o _	10,608,684
5	_ d	9,830,426

3-grams (Subword):

Rank	N-gram	Count
1	_ d i	3,986,152
2	_ d e	3,607,817
3	l a _	3,337,560
4	d i _	3,263,476
5	_ c o	3,099,394

4-grams (Subword):

Rank	N-gram	Count
1	_ d i _	3,072,491
2	_ d e l	2,506,551
3	l l a _	1,668,082
4	_ i l _	1,519,530
5	d e l l	1,488,470

5-grams (Subword):

Rank	N-gram	Count
1	_ d e l l	1,459,429
2	e l l a _	1,154,387
3	i o n e _	1,144,610
4	_ d e l _	1,020,510
5	z i o n e	936,001

Key Findings

• Best Perplexity: 2-gram (subword) with 214
• Entropy Trend: Decreases with larger n-grams (more predictable)
• Coverage: Top-1000 patterns cover ~24% of corpus
• Recommendation: 4-gram or 5-gram for best predictive performance

---

3. Markov Chain Evaluation

Results

Context	Variant	Avg Entropy	Perplexity	Branching Factor	Unique Contexts	Predictability
1	Word	1.1017	2.146	14.31	1,066,977	0.0%
1	Subword	1.1777	2.262	7.67	8,745	0.0%
2	Word	0.4680	1.383	2.78	15,250,485	53.2%
2	Subword	0.6777	1.600	4.48	67,098	32.2%
3	Word	0.2017	1.150	1.44	42,356,757	79.8%
3	Subword	0.7209	1.648	4.10	300,612	27.9%
4	Word	0.0737 🏆	1.052	1.12	61,123,669	92.6%
4	Subword	0.6799	1.602	3.41	1,231,483	32.0%

Generated Text Samples (Word-based)

Below are text samples generated from each word-based Markov chain model:

Context Size 1:

1. di assessore regionale del manto di pilotaggio e la frase di classificazione i dimostranti scendono ...
2. e della borgogna franca contea società evoluzione demografica note book apogeo del jkd è inoltre l
3. il numero di sacco viene riportato che all inizio con le descrizioni matematicamente da cui le

Context Size 2:

1. per la stirpe più distante e poi per varie statue a delfi in grecia per la prima
2. è un genere teatrale di rendere il suo simbolo è appunto quello di stern gerlach numero quantico
3. di un giovane nero floyd patterson mettendolo anche in alcuni casi come trimble contro gordon brown ...

Context Size 3:

1. altri progetti collegamenti esterni white teeth a conversation with cary grant che lo portò in testa...
2. è un comune francese di abitanti situato nel dipartimento della valle della politica di per il migli...
3. progetti collegamenti esterni t dell oceano pacifico meridionale polinesia con una superficie per co...

Context Size 4:

1. altri progetti collegamenti esterni topo gigio all ed sullivan show di elvis presley scatenò i teena...
2. è un comune francese di 75 abitanti situato nella comunità autonoma della navarra altri progetti del...
3. abitanti situato nel dipartimento dell eure nella regione della normandia società evoluzione demogra...

Generated Text Samples (Subword-based)

Below are text samples generated from each subword-based Markov chain model:

Context Size 1:

1. _ri_fi_ami_l'agi
2. i_po_mpr_pri_mpa
3. eo_ia,_ltavinafa

Context Size 2:

1. e_diatuas,_of_spe
2. a_inasa_quo_ca_ar
3. i_nal_re_e_ne_pie

Context Size 3:

1. _di_anni_di_gioria
2. _della_galle._mali
3. la_perfalcune_pelt

Context Size 4:

1. _di_fontempi_livini
2. _del_romanzo_è_anch
3. lla_classi_e_l'inse

Key Findings

• Best Predictability: Context-4 (word) with 92.6% predictability
• Branching Factor: Decreases with context size (more deterministic)
• Memory Trade-off: Larger contexts require more storage (1,231,483 contexts)
• Recommendation: Context-3 or Context-4 for text generation

---

4. Vocabulary Analysis

Statistics

Metric	Value
Vocabulary Size	511,837
Total Tokens	75,575,358
Mean Frequency	147.66
Median Frequency	4
Frequency Std Dev	7438.35

Most Common Words

Rank	Word	Frequency
1	di	3,083,677
2	e	1,824,641
3	il	1,551,962
4	la	1,467,509
5	in	1,133,909
6	a	1,025,836
7	del	1,022,058
8	che	801,243
9	un	799,178
10	della	790,420

Least Common Words (from vocabulary)

Rank	Word	Frequency
1	strensall	2
2	towthorpe	2
3	flamininus	2
4	etolici	2
5	riveros	2
6	kuntur	2
7	wachana	2
8	queñua	2
9	karabotas	2
10	tveitite	2

Zipf's Law Analysis

Metric	Value
Zipf Coefficient	1.0088
R² (Goodness of Fit)	0.996765
Adherence Quality	excellent

Coverage Analysis

Top N Words	Coverage
Top 100	39.3%
Top 1,000	60.2%
Top 5,000	76.8%
Top 10,000	83.4%

Key Findings

• Zipf Compliance: R²=0.9968 indicates excellent adherence to Zipf's law
• High Frequency Dominance: Top 100 words cover 39.3% of corpus
• Long Tail: 501,837 words needed for remaining 16.6% coverage

---

5. Word Embeddings Evaluation

5.1 Cross-Lingual Alignment

5.2 Model Comparison

Model	Dimension	Isotropy	Semantic Density	Alignment R@1	Alignment R@10
mono_32d	32	0.7834	0.3810	N/A	N/A
mono_64d	64	0.7465	0.3030	N/A	N/A
mono_128d	128	0.6690	0.2585	N/A	N/A
aligned_32d	32	0.7834 🏆	0.3788	0.3920	0.7480
aligned_64d	64	0.7465	0.3134	0.6060	0.8580
aligned_128d	128	0.6690	0.2626	0.6780	0.9340

Key Findings

• Best Isotropy: aligned_32d with 0.7834 (more uniform distribution)
• Semantic Density: Average pairwise similarity of 0.3162. Lower values indicate better semantic separation.
• Alignment Quality: Aligned models achieve up to 67.8% R@1 in cross-lingual retrieval.
• Recommendation: 128d aligned for best cross-lingual performance

---

6. Morphological Analysis (Experimental)

This section presents an automated morphological analysis derived from the statistical divergence between word-level and subword-level models. By analyzing where subword predictability spikes and where word-level coverage fails, we can infer linguistic structures without supervised data.

6.1 Productivity & Complexity

Metric	Value	Interpretation	Recommendation
Productivity Index	5.000	High morphological productivity	Reliable analysis
Idiomaticity Gap	-0.603	Low formulaic content	-

6.2 Affix Inventory (Productive Units)

These are the most productive prefixes and suffixes identified by sampling the vocabulary for global substitutability patterns. A unit is considered an affix if stripping it leaves a valid stem that appears in other contexts.

Productive Prefixes

Prefix	Examples
-s	sansavi, sfocianti, santegidiese
-a	astrolabes, applicheremo, ancia
-ma	madonne, marclay, mastrantuono
-m	mistruzzi, medioriente, madonne
-c	camminerò, ceraio, coltellacci
-p	primaibidem, pohliana, poschiavini
-b	brozzo, berlette, bucherer
-t	tigerdirect, takahito, tintry

Productive Suffixes

Suffix	Examples
-e	medioriente, madonne, lasiocampidae
-o	takahito, ceraio, brozzo
-a	gialloviola, pohliana, dawa
-i	mistruzzi, coltellacci, creatrici
-s	astrolabes, glîrs, canids
-no	mastrantuono, corrispondano, leprino
-n	guédelon, esametilen, cupidon
-te	medioriente, berlette, supercorazzate

6.3 Bound Stems (Lexical Roots)

Bound stems are high-frequency subword units that are semantically cohesive but rarely appear as standalone words. These often correspond to the 'core' of a word that requires inflection or derivation to be valid.

Stem	Cohesion	Substitutability	Examples
rono	2.58x	96 contexts	grono, crono, ronon
nche	1.63x	202 contexts	anche, nchev, ponche
ogra	1.55x	226 contexts	fogra, sogra, dogra
nter	1.48x	287 contexts	enter, inter, nterr
lmen	1.98x	62 contexts	almen, ulmen, ilmen
izza	1.47x	191 contexts	vizza, mizza, nizza
stru	1.49x	158 contexts	strub, strup, strum
ostr	1.37x	174 contexts	costr, ostro, nostr
uest	1.64x	65 contexts	quest, fuest, guest
ntro	1.50x	94 contexts	antro, intro, entro
ontr	1.42x	115 contexts	contr, contrò, hontra
ggio	1.33x	157 contexts	iggio, aggio, eggio

6.4 Affix Compatibility (Co-occurrence)

This table shows which prefixes and suffixes most frequently co-occur on the same stems, revealing the 'stacking' rules of the language's morphology.

Prefix	Suffix	Frequency	Examples
-c	-o	141 words	collalto, connettivo
-c	-e	139 words	cappelline, clorotiche
-s	-a	138 words	stäfa, serratissima
-s	-e	136 words	shilke, sommette
-a	-e	132 words	acquaforte, accreditabile
-s	-o	128 words	sperandiofrancesco, sfido
-c	-i	116 words	caseifici, consistenticittadini
-c	-a	114 words	cabarga, carapinheira
-a	-o	111 words	arcagato, ammandorlato
-a	-i	110 words	appaltatrici, angiulli

6.5 Recursive Morpheme Segmentation

Using Recursive Hierarchical Substitutability, we decompose complex words into their constituent morphemes. This approach handles nested affixes (e.g., prefix-prefix-root-suffix).

Word	Suggested Split	Confidence	Stem
iannacone	ianna-co-ne	7.5	co
frantumata	frantum-a-ta	7.5	a
strigosus	strigo-s-us	7.5	s
roccatani	rocca-ta-ni	7.5	ta
scoppiato	scoppi-a-to	7.5	a
archedemo	arched-e-mo	7.5	e
pontificem	pontific-e-m	7.5	e
approvarono	approvar-o-no	7.5	o
millières	milliè-re-s	7.5	re
cercheremo	cercher-e-mo	7.5	e
sintaxina	sintax-i-na	7.5	i
ancorarono	ancora-ro-no	7.5	ro
contesero	conte-se-ro	7.5	se
wirelessman	wirelessm-a-n	7.5	a
granadini	granad-i-ni	7.5	i

6.6 Linguistic Interpretation

Automated Insight: The language Italian shows high morphological productivity. The subword models are significantly more efficient than word models, suggesting a rich system of affixation or compounding.

---

7. Summary & Recommendations

Production Recommendations

Component	Recommended	Rationale
Tokenizer	64k BPE	Best compression (4.82x)
N-gram	2-gram	Lowest perplexity (214)
Markov	Context-4	Highest predictability (92.6%)
Embeddings	100d	Balanced semantic capture and isotropy

---

Appendix: Metrics Glossary & Interpretation Guide

This section provides definitions, intuitions, and guidance for interpreting the metrics used throughout this report.

Tokenizer Metrics

Compression Ratio

Definition: The ratio of characters to tokens (chars/token). Measures how efficiently the tokenizer represents text.

Intuition: Higher compression means fewer tokens needed to represent the same text, reducing sequence lengths for downstream models. A 3x compression means ~3 characters per token on average.

What to seek: Higher is generally better for efficiency, but extremely high compression may indicate overly aggressive merging that loses morphological information.

Average Token Length (Fertility)

Definition: Mean number of characters per token produced by the tokenizer.

Intuition: Reflects the granularity of tokenization. Longer tokens capture more context but may struggle with rare words; shorter tokens are more flexible but increase sequence length.

What to seek: Balance between 2-5 characters for most languages. Arabic/morphologically-rich languages may benefit from slightly longer tokens.

Unknown Token Rate (OOV Rate)

Definition: Percentage of tokens that map to the unknown/UNK token, indicating words the tokenizer cannot represent.

Intuition: Lower OOV means better vocabulary coverage. High OOV indicates the tokenizer encounters many unseen character sequences.

What to seek: Below 1% is excellent; below 5% is acceptable. BPE tokenizers typically achieve very low OOV due to subword fallback.

N-gram Model Metrics

Perplexity

Definition: Measures how "surprised" the model is by test data. Mathematically: 2^(cross-entropy). Lower values indicate better prediction.

Intuition: If perplexity is 100, the model is as uncertain as if choosing uniformly among 100 options at each step. A perplexity of 10 means effectively choosing among 10 equally likely options.

What to seek: Lower is better. Perplexity decreases with larger n-grams (more context). Values vary widely by language and corpus size.

Entropy

Definition: Average information content (in bits) needed to encode the next token given the context. Related to perplexity: perplexity = 2^entropy.

Intuition: High entropy means high uncertainty/randomness; low entropy means predictable patterns. Natural language typically has entropy between 1-4 bits per character.

What to seek: Lower entropy indicates more predictable text patterns. Entropy should decrease as n-gram size increases.

Coverage (Top-K)

Definition: Percentage of corpus occurrences explained by the top K most frequent n-grams.

Intuition: High coverage with few patterns indicates repetitive/formulaic text; low coverage suggests diverse vocabulary usage.

What to seek: Depends on use case. For language modeling, moderate coverage (40-60% with top-1000) is typical for natural text.

Markov Chain Metrics

Average Entropy

Definition: Mean entropy across all contexts, measuring average uncertainty in next-word prediction.

Intuition: Lower entropy means the model is more confident about what comes next. Context-1 has high entropy (many possible next words); Context-4 has low entropy (few likely continuations).

What to seek: Decreasing entropy with larger context sizes. Very low entropy (<0.1) indicates highly deterministic transitions.

Branching Factor

Definition: Average number of unique next tokens observed for each context.

Intuition: High branching = many possible continuations (flexible but uncertain); low branching = few options (predictable but potentially repetitive).

What to seek: Branching factor should decrease with context size. Values near 1.0 indicate nearly deterministic chains.

Predictability

Definition: Derived metric: (1 - normalized_entropy) × 100%. Indicates how deterministic the model's predictions are.

Intuition: 100% predictability means the next word is always certain; 0% means completely random. Real text falls between these extremes.

What to seek: Higher predictability for text generation quality, but too high (>98%) may produce repetitive output.

Vocabulary & Zipf's Law Metrics

Zipf's Coefficient

Definition: The slope of the log-log plot of word frequency vs. rank. Zipf's law predicts this should be approximately -1.

Intuition: A coefficient near -1 indicates the corpus follows natural language patterns where a few words are very common and most words are rare.

What to seek: Values between -0.8 and -1.2 indicate healthy natural language distribution. Deviations may suggest domain-specific or artificial text.

R² (Coefficient of Determination)

Definition: Measures how well the linear fit explains the frequency-rank relationship. Ranges from 0 to 1.

Intuition: R² near 1.0 means the data closely follows Zipf's law; lower values indicate deviation from expected word frequency patterns.

What to seek: R² > 0.95 is excellent; > 0.99 indicates near-perfect Zipf adherence typical of large natural corpora.

Vocabulary Coverage

Definition: Cumulative percentage of corpus tokens accounted for by the top N words.

Intuition: Shows how concentrated word usage is. If top-100 words cover 50% of text, the corpus relies heavily on common words.

What to seek: Top-100 covering 30-50% is typical. Higher coverage indicates more repetitive text; lower suggests richer vocabulary.

Word Embedding Metrics

Isotropy

Definition: Measures how uniformly distributed vectors are in the embedding space. Computed as the ratio of minimum to maximum singular values.

Intuition: High isotropy (near 1.0) means vectors spread evenly in all directions; low isotropy means vectors cluster in certain directions, reducing expressiveness.

What to seek: Higher isotropy generally indicates better-quality embeddings. Values > 0.1 are reasonable; > 0.3 is good. Lower-dimensional embeddings tend to have higher isotropy.

Average Norm

Definition: Mean magnitude (L2 norm) of word vectors in the embedding space.

Intuition: Indicates the typical "length" of vectors. Consistent norms suggest stable training; high variance may indicate some words are undertrained.

What to seek: Relatively consistent norms across models. The absolute value matters less than consistency (low std deviation).

Cosine Similarity

Definition: Measures angular similarity between vectors, ranging from -1 (opposite) to 1 (identical direction).

Intuition: Words with similar meanings should have high cosine similarity. This is the standard metric for semantic relatedness in embeddings.

What to seek: Semantically related words should score > 0.5; unrelated words should be near 0. Synonyms often score > 0.7.

t-SNE Visualization

Definition: t-Distributed Stochastic Neighbor Embedding - a dimensionality reduction technique that preserves local structure for visualization.

Intuition: Clusters in t-SNE plots indicate groups of semantically related words. Spread indicates vocabulary diversity; tight clusters suggest semantic coherence.

What to seek: Meaningful clusters (e.g., numbers together, verbs together). Avoid over-interpreting distances - t-SNE preserves local, not global, structure.

General Interpretation Guidelines

1. Compare within model families: Metrics are most meaningful when comparing models of the same type (e.g., 8k vs 64k tokenizer).
2. Consider trade-offs: Better performance on one metric often comes at the cost of another (e.g., compression vs. OOV rate).
3. Context matters: Optimal values depend on downstream tasks. Text generation may prioritize different metrics than classification.
4. Corpus influence: All metrics are influenced by corpus characteristics. Wikipedia text differs from social media or literature.
5. Language-specific patterns: Morphologically rich languages (like Arabic) may show different optimal ranges than analytic languages.

Visualizations Index

Visualization	Description
Tokenizer Compression	Compression ratios by vocabulary size
Tokenizer Fertility	Average token length by vocabulary
Tokenizer OOV	Unknown token rates
Tokenizer Total Tokens	Total tokens by vocabulary
N-gram Perplexity	Perplexity by n-gram size
N-gram Entropy	Entropy by n-gram size
N-gram Coverage	Top pattern coverage
N-gram Unique	Unique n-gram counts
Markov Entropy	Entropy by context size
Markov Branching	Branching factor by context
Markov Contexts	Unique context counts
Zipf's Law	Frequency-rank distribution with fit
Vocab Frequency	Word frequency distribution
Top 20 Words	Most frequent words
Vocab Coverage	Cumulative coverage curve
Embedding Isotropy	Vector space uniformity
Embedding Norms	Vector magnitude distribution
Embedding Similarity	Word similarity heatmap
Nearest Neighbors	Similar words for key terms
t-SNE Words	2D word embedding visualization
t-SNE Sentences	2D sentence embedding visualization
Position Encoding	Encoding method comparison
Model Sizes	Storage requirements
Performance Dashboard	Comprehensive performance overview

---

👈 Back to README

Generated by Wikilangs Pipeline · 2026-03-03 12:12:25