Comprehensive Research Report & Full Ablation Study
This repository contains NLP models trained and evaluated by Wikilangs, specifically on Tsonga 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)
Sample 3:+Jamhuri ya Kenya 125px 125px (Flag) (Coat of Arms) <small></big> 300px Kenya i ...
Vocab
Tokens
Count
8k
▁+ ja m huri ▁ya ▁kenya ▁ 1 2 5 ... (+30 more)
40
16k
▁+ jamhuri ▁ya ▁kenya ▁ 1 2 5 px ▁ ... (+28 more)
38
32k
▁+ jamhuri ▁ya ▁kenya ▁ 1 2 5 px ▁ ... (+28 more)
38
Key Findings
Best Compression: 32k achieves 4.757x compression
Lowest UNK Rate: 8k with 0.1543% 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
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
3,674
11.84
6,696
17.5%
52.9%
2-gram
Subword
203 🏆
7.66
1,426
73.8%
99.8%
3-gram
Word
5,300
12.37
7,866
12.8%
41.5%
3-gram
Subword
1,457
10.51
10,147
33.7%
80.2%
4-gram
Word
9,675
13.24
12,992
9.7%
28.5%
4-gram
Subword
6,717
12.71
42,330
16.4%
49.9%
5-gram
Word
6,251
12.61
8,596
13.3%
34.1%
5-gram
Subword
18,231
14.15
82,731
8.9%
32.4%
Top 5 N-grams by Size
2-grams (Word):
Rank
N-gram
Count
1
hi ku
816
2
na ku
518
3
tani hi
433
4
lembe ra
416
5
hi lembe
403
3-grams (Word):
Rank
N-gram
Count
1
hi lembe ra
342
2
hi siku leri
165
3
a ku ri
136
4
member of the
119
5
ku sukela hi
105
4-grams (Word):
Rank
N-gram
Count
1
add text add text
103
2
text add text add
81
3
flag coat of arms
70
4
coat of arms small
67
5
hi ku ya hi
66
5-grams (Word):
Rank
N-gram
Count
1
add text add text add
81
2
text add text add text
81
3
flag coat of arms small
67
4
life of a south african
65
5
of a south african tribe
65
2-grams (Subword):
Rank
N-gram
Count
1
a _
92,010
2
i _
42,197
3
a n
25,791
4
_ n
22,549
5
u _
22,416
3-grams (Subword):
Rank
N-gram
Count
1
n a _
15,155
2
k a _
14,105
3
_ k u
11,845
4
w a _
11,809
5
y a _
10,410
4-grams (Subword):
Rank
N-gram
Count
1
_ k u _
7,542
2
_ y a _
7,064
3
_ h i _
6,736
4
_ n a _
6,398
5
_ w a _
5,407
5-grams (Subword):
Rank
N-gram
Count
1
_ e k a _
3,747
2
a _ k u _
3,382
3
a _ s w i
2,884
4
a _ h i _
2,861
5
a _ n a _
2,635
Key Findings
Best Perplexity: 2-gram (subword) with 203
Entropy Trend: Decreases with larger n-grams (more predictable)
Coverage: Top-1000 patterns cover ~32% 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
0.7323
1.661
4.40
27,675
26.8%
1
Subword
1.1293
2.187
8.40
380
0.0%
2
Word
0.2742
1.209
1.61
121,281
72.6%
2
Subword
1.0411
2.058
5.99
3,190
0.0%
3
Word
0.0957
1.069
1.15
195,065
90.4%
3
Subword
0.8570
1.811
3.83
19,088
14.3%
4
Word
0.0308 🏆
1.022
1.04
223,778
96.9%
4
Subword
0.5850
1.500
2.39
72,986
41.5%
Generated Text Samples (Word-based)
Below are text samples generated from each word-based Markov chain model:
Context Size 1:
ku tiyisisa leswaku mati eka buccaneers loko ku betsa timbalelo u tirhile tanihi ntiro wavutshila mi...
ya siku relero e nkaveni kenya i xirho xa mozambhiki ni moya wun we people amp
hi maribye ya 70 sangiovese na vasuvuki va a xihlanganisi xa vatsari va minhlangano leyi humesiweke
Context Size 2:
hi ku angarhela va vuriwa vatatana ntirho wa vukorhokeri mati na chukele leswi bakiweke hi mahiselo ...
na ku thyakisiwa vito hi nandzu wa ku tihlanganisa na swilo kumbe swiendlakalo swin wana leswi a
tani hi psitjemba kambe a va swi dyaka swinene mapa lawa ya nyiketeriweke eka hosi kheto uve
Context Size 3:
hi lembe ra huvo leyi ku hlanganisa na swihlawulekisi swa yona xivutiso lexi saleke hi ta mihleketo ...
hi siku leri lava tswariweke hi siku leri nelson mandela khale ka phresidenti ya afrika dzonga hinkw...
a ku ri xihaha mpfhuka lexi na xona kutani va famba va cela makhele ehenhla ka xona lomu
Context Size 4:
add text add text add text add text add text add text jkl add text add text add text
text add text add text add text add text m add text add text add text xyz add text
flag coat of arms small big 300px mauritius i tiko ra afrika leri kumekaka exikhari ka afrika dzonga...
Generated Text Samples (Subword-based)
Below are text samples generated from each subword-based Markov chain model:
Context Size 1:
_keka_yowi._nya_
awambo_lo_yana_h
i_ncavu_mananatl
Context Size 2:
a_mpito_yi_ntso_h
i_tioatimisi._mu_
anitsof_tiwa_hezi
Context Size 3:
na_le_ealt=blackso
ka_mbana_e_tala_nh
_ku_andla_kufiketo
Context Size 4:
_ku_tirho_leyintirh
_ya_le_makarta)_abu
_hi_fambia_nelsprud
Key Findings
Best Predictability: Context-4 (word) with 96.9% predictability
Branching Factor: Decreases with context size (more deterministic)
Memory Trade-off: Larger contexts require more storage (72,986 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
11,975
Total Tokens
236,184
Mean Frequency
19.72
Median Frequency
4
Frequency Std Dev
174.31
Most Common Words
Rank
Word
Frequency
1
ku
7,660
2
ya
7,078
3
hi
6,805
4
na
6,438
5
wa
5,468
6
a
4,111
7
eka
3,768
8
va
3,734
9
ka
3,663
10
ra
2,819
Least Common Words (from vocabulary)
Rank
Word
Frequency
1
übersetzerinnen
2
2
übersetzern
2
3
allgemeinen
2
4
digitalisierung
2
5
linguistische
2
6
übersetzungsdienstleistungen
2
7
elektronischer
2
8
literarischen
2
9
übersetzungssektor
2
10
erfolgt
2
Zipf's Law Analysis
Metric
Value
Zipf Coefficient
1.1100
R² (Goodness of Fit)
0.990922
Adherence Quality
excellent
Coverage Analysis
Top N Words
Coverage
Top 100
49.2%
Top 1,000
75.0%
Top 5,000
92.4%
Top 10,000
98.3%
Key Findings
Zipf Compliance: R²=0.9909 indicates excellent adherence to Zipf's law
High Frequency Dominance: Top 100 words cover 49.2% of corpus
Long Tail: 1,975 words needed for remaining 1.7% 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
0.4521
0.4099
N/A
N/A
mono_64d
64
0.0928
0.3968
N/A
N/A
mono_128d
128
0.0116
0.4051
N/A
N/A
aligned_32d
32
0.4521 🏆
0.3983
0.0120
0.0800
aligned_64d
64
0.0928
0.4034
0.0220
0.1360
aligned_128d
128
0.0116
0.4050
0.0220
0.1020
Key Findings
Best Isotropy: aligned_32d with 0.4521 (more uniform distribution)
Semantic Density: Average pairwise similarity of 0.4031. Lower values indicate better semantic separation.
Alignment Quality: Aligned models achieve up to 2.2% 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
0.051
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
-m
mavulavulelo, mankweng, mindzhuti
-ma
mavulavulelo, mankweng, makumekaka
-xi
xitanga, xiendla, xiximiwa
-s
swisaka, switereka, siku
-t
thoveriwa, tiviwaka, tise
-ti
tiviwaka, tise, tikonkulu
-n
ntlambi, nwaka, ngoni
-e
endliwa, evuhlongeni, endyangwini
Productive Suffixes
Suffix
Examples
-a
thoveriwa, rhandziwa, vitana
-i
ntlambi, wansati, imini
-e
gewünschte, have, compare
-o
hikwalaho, mavulavulelo, ko
-ni
imini, koroni, evuhlongeni
-le
fuwile, uyile, chukele
-wa
thoveriwa, rhandziwa, endliwa
-ka
nwaka, swisaka, switereka
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
ungu
1.62x
54 contexts
hungu, kungu, lungu
andz
1.52x
54 contexts
andza, pandza, kandza
hamb
1.74x
27 contexts
hamba, hambi, rhambu
isiw
1.56x
36 contexts
yisiwa, hisiwa, nwisiwa
tirh
1.54x
35 contexts
tirha, tirhe, tirhi
karh
1.70x
24 contexts
karhi, nkarhi, mikarhi
ngan
1.52x
34 contexts
ngana, ngani, angana
lela
1.62x
26 contexts
hlela, fulela, leland
riwa
1.61x
25 contexts
siriwa, mariwa, soriwa
tson
1.68x
21 contexts
watson, tsongo, tsonga
arhi
1.66x
16 contexts
harhi, karhi, marhi
ngul
1.66x
15 contexts
angula, nguluve, sungule
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
-t
-a
253 words
titimela, tivekaka
-m
-a
207 words
mfukuzana, mintlwa
-v
-i
187 words
vulavisisi, vukarhi
-m
-i
177 words
mthombheni, mimiti
-v
-a
141 words
vona, vatshila
-e
-i
138 words
ematini, enyameni
-m
-o
133 words
mikomiso, minxaxamelo
-t
-i
129 words
tshuri, tihanyi
-s
-a
128 words
swekeriwa, swokoma
-t
-e
125 words
tumbuluxe, tshahiwile
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
tumbuluxiweke
tumbuluxiw-e-ke
7.5
e
chavisaka
chavis-a-ka
7.5
a
mudyandhzaka
mudyandhz-a-ka
7.5
a
swivulavulelo
swivulavu-le-lo
7.5
le
xirimbyati
xirimby-a-ti
7.5
a
wusunguleke
wusungu-le-ke
7.5
le
okmalumkoolkat
okmalumkoolk-a-t
7.5
a
nyangweni
nyangw-e-ni
7.5
e
fikeleleke
fikele-le-ke
7.5
le
xikalanga
xi-ka-langa
7.5
langa
nhlengani
nhleng-a-ni
7.5
a
leswivulaka
leswivu-la-ka
7.5
la
robertson
robert-s-on
7.5
s
exihlaleni
exihla-le-ni
7.5
le
hlanganani
hlangan-a-ni
7.5
a
6.6 Linguistic Interpretation
Automated Insight:
The language Tsonga 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
32k BPE
Best compression (4.76x)
N-gram
2-gram
Lowest perplexity (203)
Markov
Context-4
Highest predictability (96.9%)
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.
