Comprehensive Research Report & Full Ablation Study
This repository contains NLP models trained and evaluated by Wikilangs, specifically on Bangla 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
291,514
18.15
1,574,708
4.7%
13.8%
2-gram
Subword
2,633 🏆
11.36
151,712
33.5%
66.9%
3-gram
Word
772,868
19.56
2,366,241
2.2%
7.8%
3-gram
Subword
26,877
14.71
1,149,281
12.1%
33.1%
4-gram
Word
1,492,191
20.51
3,512,891
1.8%
5.9%
4-gram
Subword
176,159
17.43
5,668,680
6.6%
19.0%
5-gram
Word
1,031,104
19.98
2,302,686
2.2%
6.7%
5-gram
Subword
672,872
19.36
12,813,291
4.0%
12.3%
Top 5 N-grams by Size
2-grams (Word):
Rank
N-gram
Count
1
করা হয়
178,320
2
তথ্যসূত্র বহিঃসংযোগ
62,509
3
করা হয়েছিল
55,266
4
করা হয়েছে
52,752
5
হয় এবং
47,516
3-grams (Word):
Rank
N-gram
Count
1
থেকে সাল পর্যন্ত
15,509
2
করা হয় এবং
12,875
3
দায়িত্ব পালন করেন
11,918
4
উপর ভিত্তি করে
11,195
5
করা যেতে পারে
11,181
4-grams (Word):
Rank
N-gram
Count
1
তথ্যসূত্র বহিঃসংযোগ জন্ম ব্যক্তি
6,636
2
সংসদ সদস্য সংসদ সদস্য
6,370
3
হিসেবে দায়িত্ব পালন করেন
5,513
4
এপ্রিল জুন জুলাই সেপ্টেম্বর
5,102
5
জুলাই সেপ্টেম্বর অক্টোবর ডিসেম্বর
5,100
5-grams (Word):
Rank
N-gram
Count
1
জুন জুলাই সেপ্টেম্বর অক্টোবর ডিসেম্বর
5,049
2
এপ্রিল জুন জুলাই সেপ্টেম্বর অক্টোবর
5,048
3
মার্চ এপ্রিল জুন জুলাই সেপ্টেম্বর
5,040
4
জানুয়ারি মার্চ এপ্রিল জুন জুলাই
5,039
5
সদস্য সংসদ সদস্য সংসদ সদস্য
4,613
2-grams (Subword):
Rank
N-gram
Count
1
র _
10,460,613
2
_ এ
4,233,657
3
ন _
4,097,869
4
। _
3,608,688
5
_ ক
3,135,335
3-grams (Subword):
Rank
N-gram
Count
1
_ ক রে
1,380,255
2
এ বং _
1,266,527
3
_ এ বং
1,265,068
4
_ এ ক
991,360
5
ন । _
910,746
4-grams (Subword):
Rank
N-gram
Count
1
_ এ বং _
1,263,055
2
_ এ ক টি
584,296
3
এ ক টি _
578,361
4
_ তি নি _
473,133
5
_ ক রা _
429,980
5-grams (Subword):
Rank
N-gram
Count
1
_ এ ক টি _
571,779
2
_ হ য় । _
358,749
3
র _ জ ন্য _
344,350
4
_ ক রা _ হ
325,163
5
_ ক রে ন ।
253,567
Key Findings
Best Perplexity: 2-gram (subword) with 2,633
Entropy Trend: Decreases with larger n-grams (more predictable)
Coverage: Top-1000 patterns cover ~12% 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
0.8427
1.793
11.19
2,081,488
15.7%
1
Subword
0.9831
1.977
14.53
30,042
1.7%
2
Word
0.3490
1.274
2.13
23,273,031
65.1%
2
Subword
0.7496
1.681
6.59
436,358
25.0%
3
Word
0.1187
1.086
1.25
49,621,155
88.1%
3
Subword
0.5931
1.508
4.11
2,877,364
40.7%
4
Word
0.0412 🏆
1.029
1.07
61,780,303
95.9%
4
Subword
0.5053
1.419
2.78
11,819,297
49.5%
Generated Text Samples (Word-based)
Below are text samples generated from each word-based Markov chain model:
Context Size 1:
এবং পেনাল্টি শুট করা ঐতিহ্যগতভাবে মে তারিখে স্বাগতিক নিউজিল্যান্ড পুরুষ দীর্ঘ এবং ফোকসোনমি সালে u1 ১...
ও বিদ্রোহী দুর্গগুলির ধ্বংসাবশেষ এবং মহিলা ফুটবল ক্লাবের দৃশ্যের মিল মালিক মুম্বাইয়ে গুজরাটি ভাষায়...
হয় যা মামলুকের পদক্ষেপকে ইসরায়েল বেইট শেমেশের কাছে উন্মুক্ত এবং বাবা মাকে ডেকে পিছনে চার্জার কেইস
Context Size 2:
করা হয় ১৫ নভেম্বর the day of francophonie ২০ মার্চ রাজ্য সরকার মাদুরাইয়ে দুটি আইটি ভিত্তিক সরঞ্জাম...
তথ্যসূত্র বহিঃসংযোগ উপজেলার ইউনিয়ন বিভাগের ইউনিয়ন জেলার ইউনিয়ন বিভাগের ইউনিয়ন জেলার ইউনিয়ন পরিষ...
করা হয়েছিল যে সামাজিক প্রভাবের প্রক্রিয়া যার মাধ্যমে গুগল টক ক্লায়েন্ট তৈরি করেন texier charles r...
Context Size 3:
থেকে সাল পর্যন্ত শাখাহার ইউনিয়নের পরপর পাঁচমেয়াদে নির্বাচিত চেয়ারম্যান ছিলেন তিনি থেকে সময়কালে অ...
করা হয় এবং এই ডকুমেন্ট সম্মেলনে আলোচনা হওয়ার পর সেখানেই একটি মাদরাসা প্রতিষ্ঠার ব্যাপারে অভিমত ব্য...
দায়িত্ব পালন করেন যেখানে তিনি দ্বিতীয় স্থান অধিকার করে সমালোচনামূলক প্রতিক্রিয়া রিভিউ অ্যাগ্রিগেট...
Context Size 4:
তথ্যসূত্র বহিঃসংযোগ জন্ম ব্যক্তি কোরীয় চলচ্চিত্র অভিনেত্রী কোরীয় নারী আইডল কোরীয় নারী মডেল কোরীয়...
সংসদ সদস্য সংসদ সদস্য সংসদ সদস্য সংসদ সদস্য সংসদ সদস্য সংসদ সদস্য সংসদ সদস্য সংসদ সদস্য সংসদ সদস্য স...
হিসেবে দায়িত্ব পালন করেন জাতীয় সংসদের স্পিকার উপরাষ্ট্রপতি নিয়োগের বিধান না থাকায় রাষ্ট্রপতির অব...
Generated Text Samples (Subword-based)
Below are text samples generated from each subword-based Markov chain model:
Context Size 1:
_প্রান্তে_সারকারীতামসদর_শের
র_আবহ_নের_থাকে_জেদেরজা
ন_সাইকে_স_পরিলাক্সিকাশ্মীর_
Context Size 2:
র_ভূপৃষ্ঠতলের_জনসংখ্যান_একা
_এবং_ডুবে_ঝুঁকিপূর্ণ_ছিলেন,_
ন_নাট্যধর্মীয়_লেজে_দৃশ্য_চলচ্চি
Context Size 3:
_করেন_এবং_সর্বোচ্চ_কানাডাব্যাপী_
এবং_এর_মাত্র_এবং_"প্রক্সিমালি"
_এবং_ওল্ফের_একটি_হল_৬৪-
Context Size 4:
_এবং_তাদের_তালিকা_১৬৩_±_০
_একটি_বিরোধ_দেখেন।_সাম্রাজ্যের_
একটি_কাজ_শুরু,_সালেই_প্রায়_৭
Key Findings
Best Predictability: Context-4 (word) with 95.9% predictability
Branching Factor: Decreases with context size (more deterministic)
Memory Trade-off: Larger contexts require more storage (11,819,297 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
838,913
Total Tokens
71,898,290
Mean Frequency
85.70
Median Frequency
4
Frequency Std Dev
2805.67
Most Common Words
Rank
Word
Frequency
1
এবং
1,267,871
2
ও
702,980
3
হয়
618,329
4
করে
616,816
5
একটি
586,525
6
তিনি
495,350
7
করা
454,721
8
থেকে
424,445
9
এই
402,971
10
তার
388,104
Least Common Words (from vocabulary)
Rank
Word
Frequency
1
সণ্ডিলা
2
2
শূকরক্ষেত
2
3
প্লীপেন
2
4
মস্ম্যান
2
5
শোরোশ
2
6
yohanna
2
7
katanacho
2
8
শোরোশের
2
9
ট্রাঞ্চবলের
2
10
হুলশফ
2
Zipf's Law Analysis
Metric
Value
Zipf Coefficient
1.0269
R² (Goodness of Fit)
0.987733
Adherence Quality
excellent
Coverage Analysis
Top N Words
Coverage
Top 100
23.9%
Top 1,000
50.1%
Top 5,000
71.3%
Top 10,000
78.8%
Key Findings
Zipf Compliance: R²=0.9877 indicates excellent adherence to Zipf's law
High Frequency Dominance: Top 100 words cover 23.9% of corpus
Long Tail: 828,913 words needed for remaining 21.2% 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
0.8095 🏆
0.3709
N/A
N/A
mono_64d
64
0.8011
0.2937
N/A
N/A
mono_128d
128
0.7560
0.2281
N/A
N/A
aligned_32d
32
0.8095
0.3802
0.0980
0.4600
aligned_64d
64
0.8011
0.2992
0.2280
0.6000
aligned_128d
128
0.7560
0.2319
0.3880
0.7640
Key Findings
Best Isotropy: mono_32d with 0.8095 (more uniform distribution)
Semantic Density: Average pairwise similarity of 0.3007. Lower values indicate better semantic separation.
Alignment Quality: Aligned models achieve up to 38.8% 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
-0.452
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
Productive Suffixes
Suffix
Examples
-র
মোহানপুর, গুরবানীর, শান্তির
-ের
সেরাদের, সাইট্রেটের, লিওঁনের
-ার
ভোজভোদিনার, স্পেকটার, মনোরোগবিদ্যার
-কে
দুঃস্বপ্নকে, ক্লাইনকে, হাংচৌকে
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
ress
3.30x
93 contexts
press, dress, cress
nter
3.28x
88 contexts
enter, unter, anter
atio
3.33x
77 contexts
ratio, ation, natio
ctio
3.38x
50 contexts
action, lectio, suction
stor
2.96x
87 contexts
astor, stora, stori
mber
3.07x
60 contexts
umber, ember, amber
ence
3.40x
37 contexts
pence, fence, bence
ersi
3.11x
43 contexts
ersin, persia, persie
nati
3.22x
34 contexts
natio, nativa, nation
ical
3.23x
33 contexts
epical, apical, micali
ieve
3.35x
25 contexts
sieve, lieve, pieve
embe
3.34x
20 contexts
ember, rember, embers
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.
No significant affix co-occurrences detected.
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
স্যাপারের
স্যাপ-ার-ের
6.0
স্যাপ
ক্রুসেডারের
ক্রুসেড-ার-ের
6.0
ক্রুসেড
পরিষদসমূহের
পরিষদসমূহ-ের
4.5
পরিষদসমূহ
তন্তুগুলিকে
তন্তুগুলি-কে
4.5
তন্তুগুলি
ইতালিয়াসের
ইতালিয়াস-ের
4.5
ইতালিয়াস
দ্বিতীয়কে
দ্বিতীয়-কে
4.5
দ্বিতীয়
অ্যাসপার্টের
অ্যাসপার্ট-ের
4.5
অ্যাসপার্ট
পেটারসেনের
পেটারসেন-ের
4.5
পেটারসেন
হার্জেগোভিনাকে
হার্জেগোভিনা-কে
4.5
হার্জেগোভিনা
অ্যাক্টিনের
অ্যাক্টিন-ের
4.5
অ্যাক্টিন
মাইগ্রেশনের
মাইগ্রেশন-ের
4.5
মাইগ্রেশন
এরদোয়ানকে
এরদোয়ান-কে
4.5
এরদোয়ান
ক্রীড়াঙ্গণের
ক্রীড়াঙ্গণ-ের
4.5
ক্রীড়াঙ্গণ
অ্যাপোপটোসিসের
অ্যাপোপটোসিস-ের
4.5
অ্যাপোপটোসিস
জার্নালকে
জার্নাল-কে
4.5
জার্নাল
6.6 Linguistic Interpretation
Automated Insight:
The language Bangla 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 (5.04x)
N-gram
2-gram
Lowest perplexity (2,633)
Markov
Context-4
Highest predictability (95.9%)
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.
