Comprehensive Research Report & Full Ablation Study
This repository contains NLP models trained and evaluated by Wikilangs, specifically on Occitan Wikipedia data.
We analyze tokenizers, n-gram models, Markov chains, vocabulary statistics, and word embeddings.
📋 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)
Trade-off: Larger vocabularies improve compression but increase model size
Recommendation: 32k vocabulary provides optimal balance for production use
2. N-gram Model Evaluation
N-gram Perplexity
N-gram Unique
N-gram Coverage
Results
N-gram
Variant
Perplexity
Entropy
Unique N-grams
Top-100 Coverage
Top-1000 Coverage
2-gram
Word
40,858
15.32
382,833
16.0%
32.3%
2-gram
Subword
256 🏆
8.00
9,724
69.1%
99.1%
3-gram
Word
99,251
16.60
691,705
13.3%
25.5%
3-gram
Subword
2,095
11.03
74,490
29.1%
73.4%
4-gram
Word
144,878
17.14
1,152,073
14.3%
26.4%
4-gram
Subword
11,826
13.53
411,965
14.7%
42.1%
5-gram
Word
78,202
16.25
807,722
17.0%
32.1%
5-gram
Subword
46,870
15.52
1,292,254
8.9%
27.4%
Top 5 N-grams by Size
2-grams (Word):
Rank
N-gram
Count
1
de la
213,501
2
de l
104,259
3
es una
53,804
4
e la
52,175
5
dins lo
51,918
3-grams (Word):
Rank
N-gram
Count
1
es una comuna
35,541
2
e monuments personalitats
35,308
3
monuments personalitats ligadas
31,330
4
e la region
31,001
5
ligams extèrnes nòtas
30,871
4-grams (Word):
Rank
N-gram
Count
1
e monuments personalitats ligadas
31,329
2
luòcs e monuments personalitats
29,121
3
ligadas amb la comuna
28,401
4
personalitats ligadas amb la
28,400
5
monuments personalitats ligadas amb
27,977
5-grams (Word):
Rank
N-gram
Count
1
personalitats ligadas amb la comuna
28,398
2
e monuments personalitats ligadas amb
27,976
3
monuments personalitats ligadas amb la
27,973
4
luòcs e monuments personalitats ligadas
27,614
5
demografia luòcs e monuments personalitats
27,263
2-grams (Subword):
Rank
N-gram
Count
1
a _
3,234,853
2
e _
3,172,126
3
s _
3,063,418
4
_ d
3,054,602
5
_ l
2,264,276
3-grams (Subword):
Rank
N-gram
Count
1
_ d e
2,037,812
2
d e _
1,436,433
3
_ l a
884,600
4
l a _
867,260
5
a s _
793,636
4-grams (Subword):
Rank
N-gram
Count
1
_ d e _
1,401,348
2
_ l a _
689,470
3
d e _ l
434,454
4
i o n _
370,419
5
a _ d e
370,367
5-grams (Subword):
Rank
N-gram
Count
1
_ d e _ l
432,360
2
e _ l a _
285,044
3
d e _ l a
266,977
4
s _ d e _
254,104
5
a _ d e _
253,709
Key Findings
Best Perplexity: 2-gram (subword) with 256
Entropy Trend: Decreases with larger n-grams (more predictable)
Coverage: Top-1000 patterns cover ~27% of corpus
Recommendation: 4-gram or 5-gram for best predictive performance
3. Markov Chain Evaluation
Markov Entropy
Markov Contexts
Markov Branching
Results
Context
Variant
Avg Entropy
Perplexity
Branching Factor
Unique Contexts
Predictability
1
Word
0.9719
1.961
8.13
622,832
2.8%
1
Subword
0.8487
1.801
5.90
5,768
15.1%
2
Word
0.3692
1.292
2.14
5,057,510
63.1%
2
Subword
0.7604
1.694
4.98
33,998
24.0%
3
Word
0.1556
1.114
1.32
10,822,711
84.4%
3
Subword
0.7496
1.681
4.18
169,232
25.0%
4
Word
0.0609 🏆
1.043
1.10
14,277,493
93.9%
4
Subword
0.6863
1.609
3.33
707,794
31.4%
Generated Text Samples (Word-based)
Below are text samples generated from each word-based Markov chain model:
Context Size 1:
de comunas vesinas e solidaritat s auçant quitament d una comuna veire tanben ligams extèrnes nòtas
la corona mas es l intoxicacion son concentradas de govèrn francés livre premier estudi meninosa d
e posicion relativa istòria l entorn istòria revòlta del grand glise est attestée semble que depend
Context Size 2:
de la municipalitat qu es connectat e diferents ph es segon la definicion d un rai de
de l arnm pòrta l anèl latin digitus annularis det de l industria unica de l union
es una proprietat sus la luna esquèrra vinheta moïses trencant las taules de la nauta marna e
Context Size 3:
es una comuna francesa del departament de tarn e garona ligams extèrnes nòtas de gironda de la regio...
e monuments personalitats ligadas amb la comuna véser tanben ligams extèrnes nòtas e referéncias de ...
monuments personalitats ligadas amb la comuna véser tanben ligams extèrnes nòtas de la nauta garona ...
Context Size 4:
e monuments personalitats ligadas amb la comuna véser tanben ligams extèrnes nòtas dels vòges
luòcs e monuments personalitats ligadas amb la comuna véser tanben ligams extèrnes nòtas de normandi...
ligadas amb la comuna véser tanben ligams extèrnes nòtas de normandia de la marga
Generated Text Samples (Subword-based)
Below are text samples generated from each subword-based Markov chain model:
Context Size 1:
_daurenèrd'anaio
art_deabesime,_p
e_uzarive_se_ge_
Context Size 2:
a_doppsi_morlà_10
e_menregièrnasist
s_panar_mil_de_pl
Context Size 3:
_desfistòria_cap_a
de_jacque_dismeniv
_la_(∗)_régions_en
Context Size 4:
_de_la_grat_de_la_c
_la_fibrairie_e_avi
de_lieux_forcèt_l'a
Key Findings
Best Predictability: Context-4 (word) with 93.9% predictability
Branching Factor: Decreases with context size (more deterministic)
Memory Trade-off: Larger contexts require more storage (707,794 contexts)
Recommendation: Context-3 or Context-4 for text generation
4. Vocabulary Analysis
Zipf's Law
Top Words
Coverage Curve
Statistics
Metric
Value
Vocabulary Size
298,767
Total Tokens
19,561,503
Mean Frequency
65.47
Median Frequency
4
Frequency Std Dev
3567.99
Most Common Words
Rank
Word
Frequency
1
de
1,412,762
2
la
704,581
3
e
516,061
4
d
382,843
5
en
367,851
6
lo
364,128
7
l
357,372
8
a
301,072
9
es
226,360
10
un
196,170
Least Common Words (from vocabulary)
Rank
Word
Frequency
1
shonkinita
2
2
piròp
2
3
lherzolita
2
4
miéj
2
5
mangiato
2
6
ignaure
2
7
langfors
2
8
accouplés
2
9
theodiscus
2
10
nyamuragira
2
Zipf's Law Analysis
Metric
Value
Zipf Coefficient
1.0382
R² (Goodness of Fit)
0.998226
Adherence Quality
excellent
Coverage Analysis
Top N Words
Coverage
Top 100
45.3%
Top 1,000
64.4%
Top 5,000
78.5%
Top 10,000
84.0%
Key Findings
Zipf Compliance: R²=0.9982 indicates excellent adherence to Zipf's law
High Frequency Dominance: Top 100 words cover 45.3% of corpus
Long Tail: 288,767 words needed for remaining 16.0% coverage
5. Word Embeddings Evaluation
Embedding Isotropy
Similarity Matrix
t-SNE Words
t-SNE Sentences
5.1 Cross-Lingual Alignment
Alignment Quality
Multilingual t-SNE
5.2 Model Comparison
Model
Dimension
Isotropy
Semantic Density
Alignment R@1
Alignment R@10
mono_32d
32
0.7759
0.3605
N/A
N/A
mono_64d
64
0.7311
0.2808
N/A
N/A
mono_128d
128
0.7021
0.2184
N/A
N/A
aligned_32d
32
0.7759 🏆
0.3741
0.2480
0.6180
aligned_64d
64
0.7311
0.2733
0.3600
0.7300
aligned_128d
128
0.7021
0.2172
0.5080
0.8180
Key Findings
Best Isotropy: aligned_32d with 0.7759 (more uniform distribution)
Semantic Density: Average pairwise similarity of 0.2874. Lower values indicate better semantic separation.
Alignment Quality: Aligned models achieve up to 50.8% 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
0.170
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
-a
arrecebèva, auroish, apròhe
-s
sfrf, suris, saëns
-ma
manqueront, mahlkirch, maçacans
-c
colomberiis, chaohusaurus, campanhard
-b
bièle, brixey, bartl
-m
mcgowan, manqueront, mahlkirch
-p
pennante, pousser, pisuerga
-ca
campanhard, casalabriva, castelpers
Productive Suffixes
Suffix
Examples
-s
suris, kohs, colomberiis
-a
goja, fonologica, arrecebèva
-e
pennante, bièle, podiosalicone
-t
manqueront, projèct, convertissent
-n
mcgowan, esteron, réligion
-as
termonuclearas, taças, refractàrias
-es
ecoulettes, vongnes, neuffontaines
-on
esteron, réligion, diferencièron
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
itat
2.00x
167 contexts
pitat, gitat, itata
acio
2.06x
128 contexts
acion, bacio, racion
ogra
1.83x
133 contexts
dogra, logran, lograr
raci
1.80x
136 contexts
racim, oraci, braci
tats
2.06x
67 contexts
stats, états, etats
ntre
1.86x
105 contexts
antre, entre, intre
énci
2.13x
49 contexts
éncia, réncia, siéncia
icio
1.84x
83 contexts
licio, vicios, bricio
stra
1.35x
282 contexts
stray, strat, strad
lita
1.67x
94 contexts
litas, elita, clita
anbe
2.53x
19 contexts
anben, tanbe, tanben
tanb
2.49x
19 contexts
tanbn, tanbe, tanban
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
-s
210 words
conreats, cippus
-a
-s
171 words
annexis, annuentes
-p
-s
168 words
palays, prébois
-s
-s
126 words
senˈtises, sevas
-c
-a
119 words
casalta, conoguda
-a
-a
101 words
abjura, abominabla
-b
-s
94 words
brindas, barangays
-c
-e
94 words
colloverge, coroe
-p
-a
91 words
partidària, plantada
-m
-s
80 words
meus, majusculas
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
velhiment
velhi-me-nt
7.5
me
hermaphroditism
hermaphroditi-s-m
7.5
s
tuscaloosa
tuscaloo-s-a
7.5
s
drepanocitòsi
drepanocitò-s-i
7.5
s
sarrasiet
sarrasi-e-t
7.5
e
acomplisca
acompli-s-ca
7.5
s
daissarem
daissar-e-m
7.5
e
condusent
condus-e-nt
7.5
e
étroussat
étrous-s-at
7.5
s
garrwanas
garrw-an-as
7.5
an
prehistoria
p-re-historia
7.5
historia
cerevisiae
cerevisi-a-e
7.5
a
billinghurst
billinghur-s-t
7.5
s
europeans
europe-an-s
7.5
an
cherquesses
cherques-s-es
7.5
s
6.6 Linguistic Interpretation
Automated Insight:
The language Occitan 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
Performance Dashboard
Production Recommendations
Component
Recommended
Rationale
Tokenizer
64k BPE
Best compression (4.44x)
N-gram
2-gram
Lowest perplexity (256)
Markov
Context-4
Highest predictability (93.9%)
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
Compare within model families: Metrics are most meaningful when comparing models of the same type (e.g., 8k vs 64k tokenizer).
Consider trade-offs: Better performance on one metric often comes at the cost of another (e.g., compression vs. OOV rate).
Context matters: Optimal values depend on downstream tasks. Text generation may prioritize different metrics than classification.
Corpus influence: All metrics are influenced by corpus characteristics. Wikipedia text differs from social media or literature.
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
About This Project
Data Source
Models trained on wikipedia-monthly - a monthly snapshot of Wikipedia articles across 300+ languages.
Project
A project by Wikilangs - Open-source NLP models for every Wikipedia language.
