NLP Research Tokenization

Why LLMs Perform Better on Jumbled English than Bengali

Author: Rakesh Saha ⏱ 5 min read Tech Philosophy Series

Large Language Models (LLMs) display a remarkable capability to reconstruct coherent English sentences from jumbled inputs. However, the same capability is significantly weaker in low-resource languages like Bengali. Let's analyze the underlying statistical parameters causing this disparity.

How LLMs Predict Words

Transformers do not comprehend language logically; they evaluate statistical confidence using probability chains over sequences of discrete tokens.

Mathematical probability chain:
P(w₁, w₂, ..., wₙ) = Π P(wᵢ | w₁, ..., wᵢ₋₁)

The prediction pipeline consists of four major steps:

Case Study: English Reconstruction

Reconstructing jumbled English sentences yields high-accuracy results due to rigid syntactic structures and high-resource training corpora.

Input Payload:

percent nine genius and perspiration ninety percent is one inspiration

Output Result:

"Genius is one percent inspiration and ninety-nine percent perspiration."

The model resolves this perfectly because the quote exists repeatedly in its historical pre-training index.

Why English Reconstruction Excels:

Case Study: Bengali Reconstruction

reconstructing jumbled Bengali sentences introduces significant contextual ambiguity, leading to poor or disjointed outputs.

Input Payload:

বেলা সুরে সুর সাঁঝবেলাতে তোমার যে মেলাতে আমার যায় সুরে

Expected Output:

"আমার সাঁঝবেলাতে যে তোমার সুরে সুরে মেলাতে যায়।"

Why Bengali Reconstruction Struggles:

Comparative Language Evaluation

A comparison of the linguistic and training parameters that dictate LLM reconstruction capabilities in English and Bengali.

Linguistic Aspect English Pipeline Bengali Pipeline
Training Corpus Volume Extremely High (Trillions of tokens) Low to Moderate
Phrase Familiarity Confidence High (Strong attention baselines) Low (Sparse association chains)
Word Order Flexibility Rigid Subject-Verb-Object (SVO) Highly Flexible (SOV baseline)
Tokenization Quality Clean, whole-word mappings Frequent sub-word fragmentations
Reconstruction Accuracy High (>90% accuracy on common prompts) Lower, frequent semantic drift

Processing Logic

The sequence of operations when an LLM evaluates and reconstructs jumbled inputs:

  1. Token Identification: Converting raw inputs into integer index vectors.
  2. Pattern Identification: Attention heads search for historical quote fragments or associations.
  3. Permutation Evaluation: Simulating possible syntax connections.
  4. Probability Selection: Outputting the phrase configuration matching the highest softmax confidence.

Data Scarcity vs. Linguistic Limitation

The lower accuracy in Bengali is **not** an inherent flaw of the language itself. Rather, it is a reflection of statistical confidence. Because LLMs operate on mathematical probability, a language with lower training resources yields lower confidence outputs.

Paths to Optimization

To bridge the language gap and build high-fidelity multilingual models, developers should focus on:

Conclusion

LLMs reconstruct sentences by evaluating statistical probabilities over learned patterns. English excels due to massive historical pre-training resources and rigid syntax constraints. Bengali's morphological depth and flexible word order introduce natural ambiguities, making it an excellent case study in low-resource language optimization.