Comprehensive Research Report & Full Ablation Study
This repository contains NLP models trained and evaluated by Wikilangs, specifically on Central Dusun 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
7,224
12.82
18,432
17.6%
40.2%
2-gram
Subword
227 🏆
7.82
2,665
72.6%
99.5%
3-gram
Word
10,598
13.37
17,860
12.0%
30.6%
3-gram
Subword
1,902
10.89
18,913
28.7%
75.5%
4-gram
Word
17,687
14.11
21,653
5.2%
18.7%
4-gram
Subword
10,332
13.33
90,801
14.5%
42.9%
5-gram
Word
9,233
13.17
10,312
5.0%
23.1%
5-gram
Subword
32,680
15.00
217,159
9.9%
28.5%
Top 5 N-grams by Size
2-grams (Word):
Rank
N-gram
Count
1
nopo nga
11,657
2
id suang
2,821
3
toi ko
1,861
4
ontok toun
1,828
5
nga iso
1,049
3-grams (Word):
Rank
N-gram
Count
1
nopo nga iso
951
2
diti nopo nga
935
3
id suang do
660
4
nopo nga songulun
600
5
nopo diti nga
439
4-grams (Word):
Rank
N-gram
Count
1
nopo nga iso mantad
117
2
nopo nga iso kawo
79
3
nopo nga songulun mimingkono
75
4
nopo nga kohompit no
71
5
nopo nga iso pogun
70
5-grams (Word):
Rank
N-gram
Count
1
archived from the original on
42
2
toi ko lobi ointutunan sabaagi
34
3
koposion pogulu om pondidikan nosusu
25
4
toun uhu kono saluran tv
24
5
mw parser output reflist lower
24
2-grams (Subword):
Rank
N-gram
Count
1
a n
132,420
2
n _
100,917
3
o _
92,031
4
i _
88,621
5
o n
79,747
3-grams (Subword):
Rank
N-gram
Count
1
a n _
56,169
2
d o _
34,236
3
_ n o
33,345
4
_ d o
32,858
5
_ k o
28,766
4-grams (Subword):
Rank
N-gram
Count
1
_ d o _
30,800
2
_ i d _
22,452
3
_ o m _
19,951
4
_ n g a
17,310
5
n o p o
15,354
5-grams (Subword):
Rank
N-gram
Count
1
_ n g a _
14,567
2
_ n o p o
14,303
3
n o p o _
14,096
4
o n t o k
12,540
5
n t o k _
12,488
Key Findings
Best Perplexity: 2-gram (subword) with 227
Entropy Trend: Decreases with larger n-grams (more predictable)
Coverage: Top-1000 patterns cover ~29% 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
0.8540
1.808
5.51
70,711
14.6%
1
Subword
0.8991
1.865
5.16
1,986
10.1%
2
Word
0.2712
1.207
1.62
388,589
72.9%
2
Subword
0.6820
1.604
4.13
10,241
31.8%
3
Word
0.0811
1.058
1.13
628,158
91.9%
3
Subword
0.7746
1.711
3.85
42,293
22.5%
4
Word
0.0237 🏆
1.017
1.03
709,279
97.6%
4
Subword
0.6516
1.571
2.76
162,763
34.8%
Generated Text Samples (Word-based)
Below are text samples generated from each word-based Markov chain model:
Context Size 1:
do tasu piipiro posis nopo nga bagas menteri malaysia toi ko 7 3w 7 808 güzelbahçe
id boros sweden maamaso timpu pogulu nosusu i nopo nga okito nogi i rajaa do amu
om papaharo sikul takawas id keningau diti nga kohompit om gisom pinoposiliu do dudumagang maritim m...
Context Size 2:
nopo nga okito id posorili do kuil kuil bongunan bongunan winonsoi o kinoyonon diti galeri sukuon pa...
id suang pambalajalan loolobi id gana do sains sosial om ekonomi mogigion do pulau bali winonsoi o
toi ko bandar raya santiago gurun atacama ii gersang id utara chile nopo nga kosoruan ointutunan sab...
Context Size 3:
nopo nga iso kakadayan komponen kalas ko 5 id kointayadan do 50 tondu yahudi di bobos boroson id
diti nopo nga kiwaa totos okuri nopo nga kirati do tudan udan talasu om i bobos poinwagu nopo
id suang do watas tenom om id siriba kotoinaan do upis watas keningau di laid abaabayan dii nopo
Context Size 4:
nopo nga iso mantad tolu puruan tinimungan slav kosilahon ii kakal po do pharo ii suai nopo nga monu...
nopo nga iso kawo boros dayak i popohompit do duo dialek daro om matu dialek mantad boros austronesi...
nopo nga songulun mimingkono di abantung kopio maya piipiro film miagal ko x men apocalypse om nogi ...
Generated Text Samples (Subword-based)
Below are text samples generated from each subword-based Markov chain model:
Context Size 1:
_2_,_suhyl_palal
aheacasomomoid_p
ombaaiayosiesili
Context Size 2:
an_gan_ka_kopoko_
n_mek_koudions_gr
o_dukul_bihaguluh
Context Size 3:
an_abaagu_di_aut"_
do_sukuon_debutang
_nokobol_kopo_ngam
Context Size 4:
_do_ponuan_chillage
_id_sabaagi_gisom_n
_om_institud_5.11-3
Key Findings
Best Predictability: Context-4 (word) with 97.6% predictability
Branching Factor: Decreases with context size (more deterministic)
Memory Trade-off: Larger contexts require more storage (162,763 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
30,571
Total Tokens
714,971
Mean Frequency
23.39
Median Frequency
4
Frequency Std Dev
322.81
Most Common Words
Rank
Word
Frequency
1
do
30,939
2
id
22,604
3
om
20,001
4
nga
15,882
5
nopo
14,210
6
di
13,677
7
i
9,637
8
mantad
7,460
9
ontok
6,784
10
sabaagi
5,793
Least Common Words (from vocabulary)
Rank
Word
Frequency
1
nın
2
2
tarihçesi
2
3
paü
2
4
eğitim
2
5
dergisi
2
6
sayı
2
7
mongumang
2
8
mikattiwang
2
9
sisimbarpulou
2
10
koz
2
Zipf's Law Analysis
Metric
Value
Zipf Coefficient
1.0496
R² (Goodness of Fit)
0.994075
Adherence Quality
excellent
Coverage Analysis
Top N Words
Coverage
Top 100
41.6%
Top 1,000
66.1%
Top 5,000
84.5%
Top 10,000
91.2%
Key Findings
Zipf Compliance: R²=0.9941 indicates excellent adherence to Zipf's law
High Frequency Dominance: Top 100 words cover 41.6% of corpus
Long Tail: 20,571 words needed for remaining 8.8% 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
0.8679 🏆
0.3272
N/A
N/A
mono_64d
64
0.7620
0.2632
N/A
N/A
mono_128d
128
0.3462
0.2417
N/A
N/A
aligned_32d
32
0.8679
0.3226
0.0560
0.2820
aligned_64d
64
0.7620
0.2720
0.1060
0.3860
aligned_128d
128
0.3462
0.2427
0.2020
0.5260
Key Findings
Best Isotropy: mono_32d with 0.8679 (more uniform distribution)
Semantic Density: Average pairwise similarity of 0.2782. Lower values indicate better semantic separation.
Alignment Quality: Aligned models achieve up to 20.2% 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
-0.189
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
-po
poinkilong, pointounda, poninong
-ko
kopogonuan, kontinjen, kokomoi
-mo
monongkuyaan, mongingit, mohd
-mi
mind, millennium, minsingumbal
-ma
maru, many, matter
Productive Suffixes
Suffix
Examples
-n
louson, sukun, monongkuyaan
-an
monongkuyaan, kopogonuan, keahlian
-on
louson, southampton, unsubon
-ng
poinkilong, skateboarding, dropping
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
anga
1.64x
146 contexts
ganga, tanga, manga
ngan
1.88x
34 contexts
songan, jangan, dengan
oros
2.02x
26 contexts
boros, oroso, doros
anta
1.48x
88 contexts
banta, manta, antad
boro
2.19x
19 contexts
boros, oboros, borough
ongu
1.63x
50 contexts
tongue, tongus, mongua
impu
1.96x
24 contexts
limpu, timpu, limput
mont
1.81x
26 contexts
monto, montk, monte
ampa
1.48x
47 contexts
campa, gampa, rampa
uang
1.59x
33 contexts
huang, duang, ruang
ogun
1.79x
21 contexts
oguno, pogun, koguno
mpai
1.95x
13 contexts
ampai, rumpai, mimpai
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
-ko
-n
164 words
kolintuhunan, koyomutan
-po
-n
148 words
poimpohon, porundangan
-ko
-an
121 words
kolintuhunan, koyomutan
-po
-an
109 words
porundangan, pomutulan
-po
-on
39 words
poimpohon, potingkodon
-ko
-on
37 words
kohinoon, kosogubon
-mi
-ng
29 words
minanamong, minongisonong
-mi
-n
23 words
million, miimpohon
-mo
-ng
22 words
momoguring, moyang
-po
-ng
16 words
poring, poning
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
kopomolobusan
ko-po-mo-lobus-an
9.0
lobus
popokobong
po-po-ko-bong
7.5
bong
pomokritik
po-mo-kritik
6.0
kritik
popobibas
po-po-bibas
6.0
bibas
momooboros
mo-mo-oboros
6.0
oboros
mamagakom
ma-ma-gakom
6.0
gakom
pomodolinan
po-mo-dolin-an
4.5
dolin
koingkuri
ko-ingkuri
4.5
ingkuri
tungkusan
tungkus-an
4.5
tungkus
pengurusan
pengurus-an
4.5
pengurus
kopogisuusuayan
ko-po-gisuusuay-an
4.5
gisuusuay
pesisiran
pesisir-an
4.5
pesisir
kopomoogian
ko-po-mo-ogian
4.5
ogian
pomudagangan
po-mudaga-ng-an
4.5
mudaga
pomobodilan
po-mo-bodil-an
4.5
bodil
6.6 Linguistic Interpretation
Automated Insight:
The language Central Dusun 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 (4.96x)
N-gram
2-gram
Lowest perplexity (227)
Markov
Context-4
Highest predictability (97.6%)
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
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.
