Bapynshngain/Khasi-SpellChecker-v1

Khasi Spell Checker v1

Khasi Spell Checker v1 is a statistical spell-checking system for the Khasi language built using a large monolingual corpus (~700k sentences). The system detects and corrects spelling errors using edit-distance candidate generation and contextual ranking using a probabilistic language model.

The goal of this project is to provide basic NLP infrastructure for Khasi, a low-resource language, enabling improved writing assistance, text preprocessing, and downstream NLP applications.

🚀 Live Demo: https://huggingface.co/spaces/Bapynshngain/Khasi-Spell-Checker

---

Overview

This spell checker follows a classical architecture historically used in early search engines and spelling correction systems.

   Input sentence 
         ↓ 
    Tokenization 
         ↓ 
Suspicious word detection 
         ↓ 

Candidate generation (edit distance)

         ↓ 
Language model scoring 
         
         ↓ 

Best correction selection

         ↓ 
 Corrected sentence

The system combines:

• Edit-distance candidate generation
• Word frequency model
• Bigram language model
• Bidirectional context scoring

---

Training Data

The model is derived from a Khasi monolingual corpus containing ~700,000 sentences.

From this corpus we extracted:

Resource	Description
Vocabulary	Unique Khasi tokens
Word frequencies	Frequency counts for each token
Bigram frequencies	Context probabilities between word pairs

After cleaning, the vocabulary contains ~58,000 unique Khasi words.

---

Detection of Misspelled Words

The system first determines whether a word is likely to be incorrect.

A word is trusted if it is:

1. Present in the vocabulary
2. Sufficiently frequent in the corpus

Formally:

$$\text{If }w\in V\text{ and }freq(w)>\tau \backslash text\{If\; \}\; w\; \backslash in\; V\; \backslash text\{\; and\; \}\; freq(w)\; >\; \backslash tau$$

then the word is accepted.

Otherwise the system attempts correction.

Example:

Word	Frequency	Action
nga	high	keep
ka	very high	keep
shnogn	low	correct

---

Candidate Generation

Candidate corrections are generated using edit distance operations.

Allowed operations:

Operation	Description
Deletion	remove a character
Insertion	add a character
Replacement	replace a character
Transposition	swap adjacent characters

Example:

sngewhuh → sngewthuh

shnogn → shnong

Candidates are generated using: edits1(word) edits2(word)

Where:

$$edit{s}_2(w)=edit{s}_1(edit{s}_1(w))edits\_2(w)\; =\; edits\_1(edits\_1(w))$$

After generation:

$$Candidates=edits(w)\cap VocabularyCandidates\; =\; edits(w)\; \backslash cap\; Vocabulary$$

Only candidates present in the vocabulary are retained.

---

Probabilistic Ranking

Once candidate corrections are generated, the system ranks them probabilistically.

The classical spelling correction objective is:

$$\widehat{c}=\arg \underset{c}{\max }P(c\mathrm{\mid }w)\backslash hat\{c\}\; =\; \backslash arg\backslash max\_c\; P(c\; |\; w)$$

Using Bayes' theorem:

$$P(c\mathrm{\mid }w)=\frac{P(w\mathrm{\mid }c)P(c)}{P(w)}P(c\; |\; w)\; =\; \backslash frac\{P(w|c)P(c)\}\{P(w)\}$$

Since (P(w)) is constant:

$$\widehat{c}=\arg \underset{c}{\max }P(w\mathrm{\mid }c)P(c)\backslash hat\{c\}\; =\; \backslash arg\backslash max\_c\; P(w|c)P(c)$$

In this implementation:

• (P(c)) is approximated using word frequency
• contextual probabilities are modeled using bigram statistics

---

Language Model

A bigram language model is used to model contextual probability.

$$P(w_i\mathrm{\mid }{w}_{i-1})=\frac{count(w_{i-1},w_i)}{count(w_{i-1})}P(w\_i\; |\; w\_\{i-1\})\; =\; \backslash frac\{count(w\_\{i-1\},\; w\_i)\}\{count(w\_\{i-1\})\}$$

Example:

ban sngewthuh → common ban sngewleh → rare

Thus:

$$P(sngewthuh\mathrm{\mid }ban)>P(sngewleh\mathrm{\mid }ban)P(sngewthuh\; |\; ban)\; >\; P(sngewleh\; |\; ban)$$

---

Bidirectional Context Scoring

To improve correction accuracy, both left and right context are used.

The final candidate score is:

$$Score(c)=\log P(c)+\log P(c\mathrm{\mid }{w}_{i-1})+\log P(w_{i+1}\mathrm{\mid }c)Score(c)\; =\; \backslash log\; P(c)\; +\; \backslash log\; P(c\; |\; w\_\{i-1\})\; +\; \backslash log\; P(w\_\{i+1\}\; |\; c)$$

Where:

Term	Meaning
(P(c))	candidate word frequency
(P(c	w_{i-1}))
(P(w_{i+1}	c))

This allows the system to evaluate phrases such as: me khlem leit

instead of only evaluating: me khlem

---

Implementation Details

Language: Python

Framework: Gradio (via Hugging Face Spaces)

---

Limitations

Current limitations include:

• No explicit typo probability model (P(w|c))
• Candidate explosion for short words
• No phonetic error modeling
• No neural context understanding

Example challenging case:

khlm → kum vs khlem

Because:

frequency(kum) >> frequency(khlem)

---

---

Future Improvements

Character Error Model

Learn probabilities for common typing errors.

Trigram Language Model

Replace bigram model with:

$$P(w_i\mathrm{\mid }{w}_{i-1},w_{i-2})P(w\_i\; |\; w\_\{i-1\},\; w\_\{i-2\})$$

using tools such as KenLM.

---

Neural Spell Correction

Future versions may incorporate neural models such as:

• BERT
• T5
• sequence-to-sequence transformers

for improved contextual understanding.

---

Intended Use

This spell checker is designed for:

• Khasi writing assistance
• educational tools
• preprocessing Khasi text
• improving downstream NLP pipelines

---

Citation

If you use this work, please cite:

@software{nongkynrih2026khasi_spellchecker_v1,
  author       = {Nongkynrih, Bapynshngainlang},
  title        = {Khasi Spell Checker v1},
  version      = {1.0},
  year         = {2026},
  month        = mar,
  day          = 13
  publisher    = {Hugging Face},
  doi          = {10.57967/hf/7999},
  url          = {https://huggingface.co/Bapynshngain/Khasi-SpellChecker-v1}
}

APA Citation

Nongkynrih, B. (2026, March 13). Khasi Spell Checker v1 (Version 1.0) [Software]. Hugging Face. https://doi.org/10.57967/hf/7999