RESEARCH_REPORT.md

Hungarian — Full Ablation Study & Research Report

Detailed evaluation of all model variants trained on Hungarian 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	3.505x	3.51	0.1725%	3,328,421
16k	3.921x	3.92	0.1930%	2,974,904
32k	4.311x	4.31	0.2122%	2,705,982
64k	4.661x 🏆	4.66	0.2295%	2,502,482

Tokenization Examples

Below are sample sentences tokenized with each vocabulary size:

Sample 1: Elvonás, addiktológia Elvonás, a szóalkotás egy módja

Vocab	Tokens	Count
8k	▁elv on ás , ▁ad d ikt ológia ▁elv on ... (+9 more)	19
16k	▁elv on ás , ▁add ikt ológia ▁elv on ás ... (+8 more)	18
32k	▁elvon ás , ▁add ikt ológia ▁elvon ás , ▁a ... (+5 more)	15
64k	▁elvon ás , ▁add ikt ológia ▁elvon ás , ▁a ... (+4 more)	14

Sample 2: Memória (biológia) Memória (számítástechnika): Számítástechnikában használják Me...

Vocab	Tokens	Count
8k	▁mem ória ▁( bi ológia ) ▁mem ória ▁( szám ... (+23 more)	33
16k	▁mem ória ▁( bi ológia ) ▁mem ória ▁( számítás ... (+19 more)	29
32k	▁memória ▁( bi ológia ) ▁memória ▁( számítás technika ): ... (+12 more)	22
64k	▁memória ▁( biológia ) ▁memória ▁( számítás technika ): ▁számítástechn ... (+10 more)	20

Sample 3: Óe, japán családnév Óe, kisváros Japánban, Jamagata prefektúrában ÓE, az Óbudai ...

Vocab	Tokens	Count
8k	▁ó e , ▁japán ▁család név ▁ó e , ▁kis ... (+21 more)	31
16k	▁ó e , ▁japán ▁családnév ▁ó e , ▁kisváros ▁japánban ... (+16 more)	26
32k	▁ó e , ▁japán ▁családnév ▁ó e , ▁kisváros ▁japánban ... (+13 more)	23
64k	▁ó e , ▁japán ▁családnév ▁ó e , ▁kisváros ▁japánban ... (+11 more)	21

Key Findings

• Best Compression: 64k achieves 4.661x compression
• Lowest UNK Rate: 8k with 0.1725% 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	362,917	18.47	2,207,301	5.4%	14.2%
2-gram	Subword	429 🏆	8.74	22,141	54.6%	98.3%
3-gram	Word	1,261,909	20.27	3,510,384	2.1%	6.2%
3-gram	Subword	4,501	12.14	188,349	17.4%	56.2%
4-gram	Word	2,487,801	21.25	5,404,855	2.0%	5.1%
4-gram	Subword	29,749	14.86	1,214,042	7.6%	26.9%
5-gram	Word	1,806,602	20.78	3,707,959	2.3%	5.8%
5-gram	Subword	135,455	17.05	4,635,566	3.7%	15.6%

Top 5 N-grams by Size

2-grams (Word):

Rank	N-gram	Count
1	és a	381,304
2	hogy a	127,431
3	és az	111,129
4	a magyar	107,955
5	volt a	102,268

3-grams (Word):

Rank	N-gram	Count
1	személyek elhunyt személyek	25,908
2	született személyek elhunyt	25,567
3	madarai madarai madarai	21,522
4	jegyzetek további információk	19,348
5	kisbolygók listája jegyzetek	13,405

4-grams (Word):

Rank	N-gram	Count
1	született személyek elhunyt személyek	25,567
2	madarai madarai madarai madarai	17,618
3	listája jegyzetek naprendszer kisbolygói	13,129
4	kisbolygók listája jegyzetek naprendszer	13,128
5	kapcsolódó szócikkek kisbolygók listája	13,112

5-grams (Word):

Rank	N-gram	Count
1	madarai madarai madarai madarai madarai	14,553
2	kisbolygók listája jegyzetek naprendszer kisbolygói	13,128
3	kapcsolódó szócikkek kisbolygók listája jegyzetek	13,034
4	szócikkek kisbolygók listája jegyzetek naprendszer	12,763
5	a naprendszer kisbolygóövében található aszteroida	12,084

2-grams (Subword):

Rank	N-gram	Count
1	_ a	14,447,157
2	a _	13,088,817
3	s z	10,335,605
4	t _	8,923,690
5	e l	8,560,877

3-grams (Subword):

Rank	N-gram	Count
1	_ a _	7,569,874
2	_ s z	3,684,895
3	_ a z	2,740,812
4	é s _	2,565,914
5	s z e	2,500,891

4-grams (Subword):

Rank	N-gram	Count
1	_ a z _	2,372,557
2	_ é s _	2,184,467
3	_ e g y	1,442,169
4	_ m e g	1,293,813
5	. _ a _	1,289,446

5-grams (Subword):

Rank	N-gram	Count
1	_ a z _ e	641,291
2	_ s z e r	617,289
3	_ é s _ a	566,317
4	_ e g y _	545,352
5	_ v o l t	535,186

Key Findings

• Best Perplexity: 2-gram (subword) with 429
• Entropy Trend: Decreases with larger n-grams (more predictable)
• Coverage: Top-1000 patterns cover ~16% 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	0.9424	1.922	10.97	3,256,935	5.8%
1	Subword	1.1652	2.243	8.26	11,176	0.0%
2	Word	0.3123	1.242	2.03	35,712,539	68.8%
2	Subword	0.6772	1.599	4.69	92,249	32.3%
3	Word	0.1163	1.084	1.24	72,420,142	88.4%
3	Subword	0.7664	1.701	4.71	432,153	23.4%
4	Word	0.0401 🏆	1.028	1.06	89,534,275	96.0%
4	Subword	0.7484	1.680	3.95	2,032,871	25.2%

Generated Text Samples (Word-based)

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

Context Size 1:

1. a filozófiában jeleskedők élete vácon igazgatótanár aszalón bőcsön borsod abaúj zemplén vármegye hon...
2. az elfelejtett kísérlet singer berta fia némó elbűvölő szörnyeteg régebben megszűnt a európa bajnoki...
3. és dániai roskilde dánia szoroson át kellett találni a királyi engedélyt egy bizonyos ágakra elosztó...

Context Size 2:

1. és a helybeli hazafiak köztük władysław gomułka lett akkor több fellépést is a neve és a razorblade
2. hogy a gerjesztést követően 0 0 0 0 0 0 0 fe7 lépéssorozatból áll a leendő űrkísérletekhez
3. és az egy évvel korábban halt meg csak hallomásból ismerik a kantonmozdony mivel az ember annak érde...

Context Size 3:

1. született személyek elhunyt személyek drámaírók novogyevicsi temetőben eltemetett személyek erdélyi ...
2. madarai madarai madarai északi mariana szigetek madarai madarai madarai madarai amerikai egyesült ál...
3. jegyzetek további információk labdarúgók river plate labdarúgói cruz azul labdarúgói atlas labdarúgó...

Context Size 4:

1. madarai madarai madarai madarai madarai madarai tomé és príncipe madarai madarai madarai seychelle s...
2. született személyek elhunyt személyek úszók olimpiai ezüstérmesek nők
3. kisbolygók listája jegyzetek naprendszer kisbolygói vec lista de aënna

Generated Text Samples (Subword-based)

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

Context Size 1:

1. _tén.krermi_k_ho
2. emi_ra_vaiszegyő
3. ajat_(gola,_jena

Context Size 2:

1. _a_törül_ingmukiv
2. a_katt,_fő_sztünc
3. szvetibelyek_mika

Context Size 3:

1. _a_„tusábang_napil
2. _szócikkel_a_bolyg
3. _az_a4_március_211

Context Size 4:

1. _az_eszit_lasszultá
2. _és_galéria_szereti
3. _egyik_mária_foglal

Key Findings

• Best Predictability: Context-4 (word) with 96.0% predictability
• Branching Factor: Decreases with context size (more deterministic)
• Memory Trade-off: Larger contexts require more storage (2,032,871 contexts)
• Recommendation: Context-3 or Context-4 for text generation

---

4. Vocabulary Analysis

Statistics

Metric	Value
Vocabulary Size	1,458,224
Total Tokens	103,537,627
Mean Frequency	71.00
Median Frequency	4
Frequency Std Dev	7231.61

Most Common Words

Rank	Word	Frequency
1	a	7,834,400
2	az	2,443,698
3	és	2,194,407
4	is	733,028
5	egy	579,913
6	hogy	494,341
7	volt	474,789
8	nem	446,048
9	1	405,757
10	magyar	343,055

Least Common Words (from vocabulary)

Rank	Word	Frequency
1	léggömbzár	2
2	elterelővel	2
3	hajófelderítő	2
4	szűrőrendszerrel	2
5	801ml	2
6	801tp	2
7	sorosmotort	2
8	stammkennzeichen	2
9	schindleréről	2
10	kapraraszműlesiklásnem	2

Zipf's Law Analysis

Metric	Value
Zipf Coefficient	0.9296
R² (Goodness of Fit)	0.996343
Adherence Quality	excellent

Coverage Analysis

Top N Words	Coverage
Top 100	25.8%
Top 1,000	45.6%
Top 5,000	61.9%
Top 10,000	69.2%

Key Findings

• Zipf Compliance: R²=0.9963 indicates excellent adherence to Zipf's law
• High Frequency Dominance: Top 100 words cover 25.8% of corpus
• Long Tail: 1,448,224 words needed for remaining 30.8% 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	0.7886	0.3560	N/A	N/A
mono_64d	64	0.7831	0.2846	N/A	N/A
mono_128d	128	0.7114	0.2337	N/A	N/A
aligned_32d	32	0.7886 🏆	0.3516	0.3600	0.7520
aligned_64d	64	0.7831	0.2765	0.5320	0.8560
aligned_128d	128	0.7114	0.2243	0.6300	0.9180

Key Findings

• Best Isotropy: aligned_32d with 0.7886 (more uniform distribution)
• Semantic Density: Average pairwise similarity of 0.2878. Lower values indicate better semantic separation.
• Alignment Quality: Aligned models achieve up to 63.0% 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	-0.625	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	schmör, szabadítására, szertárai
-k	kelvinator, kürtszavára, keselyűféléket
-a	arkopharma, adequate, akszukhosz
-m	megyéjébe, mongu, mozilátogatók
-t	tlp, terheltségének, tanulhatóak
-b	balszárnyának, bhov, beszédfejlődési
-ma	mainstoneky, margati, mayrnak
-e	elektronszerkezetével, egységnyiek, esztyemirova

Productive Suffixes

Suffix	Examples
-k	mozilátogatók, balszárnyának, kóruspadok
-t	használókat, felvevőpiacot, keselyűféléket
-n	végezetlen, érinthetetlen, állatkertjeiben
-a	arkopharma, kürtszavára, vaskarika
-l	elektronszerkezetével, versenyeitől, pálinkafőzéssel
-s	rombouts, ikszes, gibbins
-i	lxviii, desai, vibrálni
-e	megyéjébe, jugoslovenske, adequate

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
váro	2.03x	219 contexts	város, váron, várod
embe	1.60x	288 contexts	ember, embed, bembe
llen	1.53x	361 contexts	llena, illen, lleno
atás	1.48x	365 contexts	patás, hatás, avatás
llet	1.60x	226 contexts	allet, ellet, illet
ítás	1.43x	316 contexts	újítás, ásítás, ámítás
ítot	1.65x	136 contexts	ított, vított, osított
erül	1.43x	288 contexts	kerül, terül, derül
mber	1.33x	400 contexts	ember, imber, amber
ület	1.38x	311 contexts	fület, szület, őrület
rtén	2.09x	42 contexts	örtény, értény, kurtén
örté	1.85x	64 contexts	törté, örtény, körték

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
-k	-t	115 words	közéletét, kansast
-k	-k	112 words	kreatívnak, kiújulnak
-s	-l	101 words	szúfizmussal, szinkronstábbal
-s	-k	99 words	sírkamrájuk, szülővárosnak
-s	-t	95 words	szakmáikat, shiflett
-m	-k	86 words	mellékérték, maffiacsoportok
-k	-l	83 words	középjel, krízisekkel
-s	-n	81 words	szövegírásban, spalatóban
-k	-n	76 words	kihalásában, konstruktiven
-k	-a	72 words	korongcsiga, kaparinthatja

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
vercruysse	vercruys-s-e	7.5	s
szocialta	szoci-al-ta	7.5	al
stílusbanaz	stílusban-a-z	7.5	a
részvényeiket	részvényei-k-et	7.5	k
hátoldalra	hátold-al-ra	7.5	al
fullernek	fuller-n-ek	7.5	n
nézhetőnek	nézhető-n-ek	7.5	n
napjainkbeli	napjainkb-el-i	7.5	el
gyarmataikra	gyarmatai-k-ra	7.5	k
mandülion	mandül-i-on	7.5	i
robotterheit	robotterhe-i-t	7.5	i
albuginea	albugin-e-a	7.5	e
elnagyoltak	elnagyol-t-ak	7.5	t
minőségileg	minőségil-e-g	7.5	e
szárnyacskák	szárnyacs-k-ák	7.5	k

6.6 Linguistic Interpretation

Automated Insight: The language Hungarian 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 (4.66x)
N-gram	2-gram	Lowest perplexity (429)
Markov	Context-4	Highest predictability (96.0%)
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 · 2026-03-04 21:45:03