Comprehensive Research Report & Full Ablation Study
This repository contains NLP models trained and evaluated by Wikilangs, specifically on Walloon 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:anêyes | anêyes | anêyes | anêyes | anêyes | | | | | | | | | Evenmints Personåli...
Vocab
Tokens
Count
8k
`▁anêyes ▁
▁anêyes ▁
16k
`▁anêyes ▁
▁anêyes ▁
32k
`▁anêyes ▁
▁anêyes ▁
64k
`▁anêyes ▁
▁anêyes ▁
Key Findings
Best Compression: 64k achieves 3.891x compression
Lowest UNK Rate: 8k with 0.2270% 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
13,161
13.68
50,899
15.9%
38.7%
2-gram
Subword
287 🏆
8.17
3,898
66.2%
98.9%
3-gram
Word
27,389
14.74
75,959
10.3%
28.8%
3-gram
Subword
2,270
11.15
30,979
28.7%
71.3%
4-gram
Word
43,477
15.41
111,769
10.0%
25.5%
4-gram
Subword
11,809
13.53
153,308
14.8%
41.1%
5-gram
Word
23,357
14.51
68,834
14.4%
33.1%
5-gram
Subword
40,103
15.29
373,182
8.5%
26.1%
Top 5 N-grams by Size
2-grams (Word):
Rank
N-gram
Count
1
c est
19,149
2
e walon
7,096
3
dins l
6,135
4
gn a
5,364
5
di l
5,071
3-grams (Word):
Rank
N-gram
Count
1
c est ene
4,038
2
c est on
3,474
3
c est l
2,865
4
i gn a
1,935
5
ciste anêye la
1,829
4-grams (Word):
Rank
N-gram
Count
1
ont vnou å monde
1,112
2
rilomés walons et waloneus
926
3
la rilomés walons et
919
4
ancyin ptit ban del
907
5
ptit ban del walonreye
881
5-grams (Word):
Rank
N-gram
Count
1
la rilomés walons et waloneus
919
2
ancyin ptit ban del walonreye
876
3
èn ancyin ptit ban del
862
4
est èn ancyin ptit ban
860
5
c est èn ancyin ptit
795
2-grams (Subword):
Rank
N-gram
Count
1
e _
346,952
2
s _
328,323
3
_ d
305,245
4
e s
247,267
5
_ l
199,665
3-grams (Subword):
Rank
N-gram
Count
1
e s _
158,630
2
_ d i
98,745
3
e _ d
74,828
4
_ d e
71,105
5
s _ d
59,946
4-grams (Subword):
Rank
N-gram
Count
1
_ l ' _
52,800
2
_ d i _
48,903
3
l e s _
45,518
4
_ d ' _
40,065
5
_ l e s
39,565
5-grams (Subword):
Rank
N-gram
Count
1
_ l e s _
39,370
2
_ d e s _
36,468
3
a e d j e
30,194
4
_ e s t _
28,789
5
w a l o n
26,821
Key Findings
Best Perplexity: 2-gram (subword) with 287
Entropy Trend: Decreases with larger n-grams (more predictable)
Coverage: Top-1000 patterns cover ~26% 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
0.8999
1.866
6.37
116,505
10.0%
1
Subword
0.9760
1.967
8.18
1,236
2.4%
2
Word
0.3359
1.262
1.91
738,986
66.4%
2
Subword
0.9487
1.930
6.02
10,102
5.1%
3
Word
0.1299
1.094
1.24
1,402,590
87.0%
3
Subword
0.8319
1.780
4.29
60,750
16.8%
4
Word
0.0508 🏆
1.036
1.08
1,734,432
94.9%
4
Subword
0.6556
1.575
2.86
260,171
34.4%
Generated Text Samples (Word-based)
Below are text samples generated from each word-based Markov chain model:
Context Size 1:
l pîce di a passé les diferincyî des floricontes thumb li holande c est ene cwårêye
a hesta li trope a sketé dji shijhele ou des viyaedjes so pås rfondeus do lussimbork
di scrire disk e l påye ki vént lére li 219inme po do calindrî grigoryin li
Context Size 2:
c est adon k ele s î ont dmoré dins les codjowaedjes et des sårts miertchamp rond
e walon eplaideye di jean collette rééditer c est vos k on lyi cåze dins l esplicant
dins l esplicant motî do tchestea rnåd mora l an 150 di filozofeye des loumires ou set
Context Size 3:
c est ene plaece sol fagne walone metans pol trouflaedje a stî foirt sibaré pal guere di la
c est on lingaedje do sud ess do nidjeria gn a eto des tchampions microscopikes les emacralêyès cawe...
c est l eshonna di totes les dujhances et des ovraedjes d ene metowe maladeye on djåzrè puvite
Context Size 4:
ont vnou å monde ciste anêye la ont morou ciste anêye la rilomés walons et waloneus arthur trigaux ô...
rilomés walons et waloneus ôtès djins fiesses nåcionåles ey eternåcionåles vey eto 27 di djanvî 28 d...
la rilomés walons et waloneus rené magritte ôtès djins fiesses nåcionåles ey eternåcionåles vey eto ...
Generated Text Samples (Subword-based)
Below are text samples generated from each subword-based Markov chain model:
Context Size 1:
_doumwic'_ma_par
estoke,_lel'_da_
s_a_e_aericr,_ix
Context Size 2:
e_es_moxhamarou_d
s_ni_recis_re,_in
_di_li_zen_yu_cro
Context Size 3:
es_osse_bassé_pass
_di_shuvan_da_måvl
e_des_espal_ricnox
Context Size 4:
_l'_radio_pårteye_(
_di_fevrî-mont._met
les_tchaeffner:_min
Key Findings
Best Predictability: Context-4 (word) with 94.9% predictability
Branching Factor: Decreases with context size (more deterministic)
Memory Trade-off: Larger contexts require more storage (260,171 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
52,390
Total Tokens
2,080,878
Mean Frequency
39.72
Median Frequency
4
Frequency Std Dev
699.28
Most Common Words
Rank
Word
Frequency
1
l
60,706
2
a
49,899
3
di
49,556
4
d
44,591
5
li
41,896
6
les
40,671
7
des
36,781
8
on
34,019
9
e
30,971
10
est
29,225
Least Common Words (from vocabulary)
Rank
Word
Frequency
1
strivay
2
2
valkeneer
2
3
kotsifakos
2
4
cogolati
2
5
coprezide
2
6
lecocq
2
7
siclimboigne
2
8
pozzo
2
9
samourayes
2
10
diplomats
2
Zipf's Law Analysis
Metric
Value
Zipf Coefficient
1.1436
R² (Goodness of Fit)
0.997501
Adherence Quality
excellent
Coverage Analysis
Top N Words
Coverage
Top 100
48.1%
Top 1,000
72.7%
Top 5,000
86.8%
Top 10,000
91.6%
Key Findings
Zipf Compliance: R²=0.9975 indicates excellent adherence to Zipf's law
High Frequency Dominance: Top 100 words cover 48.1% of corpus
Long Tail: 42,390 words needed for remaining 8.4% 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
0.8697 🏆
0.3451
N/A
N/A
mono_64d
64
0.8678
0.2695
N/A
N/A
mono_128d
128
0.7751
0.1978
N/A
N/A
aligned_32d
32
0.8697
0.3408
0.0460
0.2260
aligned_64d
64
0.8678
0.2653
0.0800
0.3180
aligned_128d
128
0.7751
0.1961
0.1180
0.4000
Key Findings
Best Isotropy: mono_32d with 0.8697 (more uniform distribution)
Semantic Density: Average pairwise similarity of 0.2691. Lower values indicate better semantic separation.
Alignment Quality: Aligned models achieve up to 11.8% 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
0.268
High formulaic/idiomatic 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
sorpwès, soucant, schalon
-a
ashît, aschoûter, arivéve
-c
chanchesse, crac, coûte
-r
rékem, ritape, rahoucants
-d
dvuzlêyès, djilet, djoyes
-b
begnons, branmint, borins
-p
pattepårti, popes, preyale
-m
marker, manuels, montes
Productive Suffixes
Suffix
Examples
-e
frîmårtinisse, chanchesse, oyåve
-s
begnons, sorpwès, åmônes
-es
åmônes, popes, goidjes
-t
ashît, soucant, veyant
-ye
marveye, veskeveye, eveye
-je
kischoyaedje, laudje, redjårbaedje
-nt
soucant, veyant, branmint
-n
schalon, tramwegen, ploumtion
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
aedj
2.24x
178 contexts
aedje, saedje, taedje
tche
1.88x
250 contexts
tcheû, tchet, tcheu
êyes
2.10x
64 contexts
fêyes, idêyes, atêyes
sses
1.86x
93 contexts
asses, åsses, esses
edje
2.22x
38 contexts
nedje, aedje, wedje
djes
2.01x
38 contexts
ådjes, tidjes, vèdjes
ants
1.95x
35 contexts
wants, pzants, tnants
rijh
1.68x
55 contexts
prijhî, grijhe, prijhe
fran
1.95x
31 contexts
frane, franz, frank
ranc
1.69x
46 contexts
rance, franc, franci
scri
2.10x
18 contexts
scrit, scrip, scris
teut
2.28x
14 contexts
steut, eteut, asteut
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
-e
184 words
capitole, crustinnisse
-r
-e
183 words
riwaitaedje, rifômrece
-c
-s
170 words
curieus, crouwès
-s
-e
160 words
sicoreye, soucråde
-a
-e
146 words
ahèsse, ake
-p
-e
143 words
poelvoorde, poytreye
-d
-e
141 words
dialectologique, divizêye
-t
-e
131 words
turke, tontelange
-p
-s
130 words
purdans, potches
-s
-s
127 words
stitchîs, såvadjes
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
tchampionatn
tchampiona-t-n
7.5
t
cråxhoulet
cråxhoul-e-t
7.5
e
coirsulet
coirsul-e-t
7.5
e
bouxhreye
bouxhr-e-ye
7.5
e
pharmacien
pharmaci-e-n
7.5
e
forijhots
forijho-t-s
7.5
t
sloveneye
sloven-e-ye
7.5
e
fiziolodjeye
fiziolodj-e-ye
7.5
e
diswaibeye
diswaib-e-ye
7.5
e
omeyopateye
omeyopat-e-ye
7.5
e
påjhûlisté
påjhûli-s-té
7.5
s
tchimisse
tchimi-s-se
7.5
s
djouwreut
djouw-re-ut
7.5
re
ôrtografeye
ôrtograf-e-ye
7.5
e
plantisse
planti-s-se
7.5
s
6.6 Linguistic Interpretation
Automated Insight:
The language Walloon shows high morphological productivity. The subword models are significantly more efficient than word models, suggesting a rich system of affixation or compounding.
Note on Idiomaticity: The high Idiomaticity Gap suggests a large number of frequent multi-word expressions or formulaic sequences that are statistically distinct from their component parts.
7. Summary & Recommendations
Performance Dashboard
Production Recommendations
Component
Recommended
Rationale
Tokenizer
64k BPE
Best compression (3.89x)
N-gram
2-gram
Lowest perplexity (287)
Markov
Context-4
Highest predictability (94.9%)
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.
