Context-Aware & Grammar Validation

Documentation Index

Fetch the complete documentation index at: https://docs.myspellchecker.com/llms.txt

Use this file to discover all available pages before exploring further.

A word can be spelled correctly yet still be wrong. Myanmar homophones like စား (eat) and စာ (letter) pass dictionary checks but produce nonsense when swapped. Catching these errors requires looking beyond the word itself. mySpellChecker uses two complementary strategies: Syntactic Grammar Checking (Layer 2.5) applies rigid POS-based rules, while N-gram Probability (Layer 3) scores how likely a word sequence is given the surrounding context. Example (Myanmar homophone confusion):

“သူ စား ချင်တယ်။” (He wants to eat.) Correct. “သူ စာ ချင်တယ်။” (He letter wants-to, which is ungrammatical; ချင် requires a preceding verb.) Spelling correct, context incorrect.

Another example:

“ကျောင်း သား တစ်ယောက်” (A student) Correct. “ကျောင်း သာ တစ်ယောက်” (School only one person) Grammatically awkward.

​

Syntactic Grammar Checking

This layer uses Part-of-Speech (POS) tagging and deterministic rules to catch grammatical errors that statistical models might miss due to data sparsity.

​

How it Works

1. POS Tagging: Every word in the dictionary can optionally have a POS tag (e.g., N for Noun, V for Verb, P for Particle).
2. Rule Engine: A set of linguistic rules defines valid and invalid sequences.

​

Example Rules

• Verb + Particle Agreement:
  • Invalid: သွား (Go/Verb) + ကျောင်း (School/Noun) → “Go School” (Grammatically awkward, missing particle)
  • Correction: ကျောင်းသွား or ကျောင်းကို သွား → “Go to school”
• Nominalizer Particle ကြောင်း vs Noun ကျောင်း:
  • သွားကြောင်း ပြောတယ် (Said that [he] went). Correct: ကြောင်း nominalizes the verb.
  • သွားကျောင်း ပြောတယ် is invalid because ကျောင်း (school) cannot follow a verb directly
• Particle Selection (မှာ vs မှ):
  • ရုံးမှာ ရှိတယ် (Is at the office). Correct: မှာ indicates location.
  • ရုံးမှ လာတယ် (Came from the office). Correct: မှ indicates origin.
  • Rule: After Noun, မှာ typically means “at/in”. မှ typically means “from” or marks conditional.
• Subject Marker Agreement (က vs ကို):
  • သူက စာအုပ်ကို ဖတ်တယ် (He reads the book). Correct.
  • သူကို စာအုပ်က ဖတ်တယ် (The book reads him). Semantically incorrect.
  • Rule: Animate subjects typically take က; objects take ကို
• Question Particle Matching:
  • ဘယ်သူလဲ (Who is it?). Correct: ဘယ် question word + လဲ particle.
  • ဘယ်သူလား is also valid but carries a different nuance (softer question)
  • ဘာလဲ vs ဘာပဲ have different meanings: “What?” vs “Whatever”

​

N-gram Probability

mySpellChecker uses N-gram models (Bigrams and Trigrams) to calculate the probability of word sequences.

1. Bigram: Probability of Word B following Word A ($P(B|A)$).
2. Trigram: Probability of Word C following A and B ($P(C|A,B)$).

​

The Algorithm

1. Detection: When the checker encounters a sequence of words, it queries the database for the frequency of that sequence. If $P(Word_i | Word_{i-1})$ is below bigram_threshold, the word is flagged as suspicious.
2. Correction: The system generates candidates for the suspicious word (using SymSpell or Phonetic matching). It then re-calculates probabilities for each candidate in the sentence. Example: Input sentence “သူ စာ ချင်တယ်” (suspicious word: စာ)
   • Candidate “စား” (eat): $P(\text{စား} | \text{သူ}) = 0.08$ (high, because it is a common verb after a pronoun)
   • Candidate “စာ” (letter): $P(\text{စာ} | \text{သူ}) = 0.002$ (low, because it is less common as a standalone)
   The system suggests “စား” because it fits the context better with the verb-wanting pattern “ချင်တယ်”.

​

Advanced Strategies

The N-gram checker employs several heuristics to handle unseen data and improve accuracy:

​

1. Backoff Smoothing (Unigram Check)

If a bigram probability is zero (unseen sequence), the checker looks at the unigram frequency of the word.

• If the word is very common globally (high unigram frequency), we assume it is likely correct but used in a novel context. It is not flagged as an error.
• If the word is rare, it is more likely to be a typo.

Example:

• Input: မြန်မာ ဂီတ (Myanmar music), a bigram unseen in the corpus
• ဂီတ has high unigram frequency (common word for “music”)
• Result: Not flagged as error, assumed to be valid novel combination

​

2. Typo Heuristic

For unseen rare words, the checker searches for “neighbors” (words with Edit Distance = 1) that fit the context with high probability.

• If a neighbor has a high bigram probability ($P > \text{threshold} \times 10$), we assume the current word is a typo of that neighbor and flag it.

Example:

• Input: စာအုပ် ဖတ်တတ် (rare/unseen word ဖတ်တတ်)
• Neighbor found: ဖတ်တယ် (reads) with Edit Distance = 1
• $P(\text{ဖတ်တယ်} | \text{စာအုပ်}) = 0.15$ (high bigram probability)
• Result: Flag ဖတ်တတ် as likely typo of ဖတ်တယ်

​

Tone Disambiguation

In Myanmar language, tone marks ( ့, း) drastically change the meaning of a word. Many spelling errors involve missing or incorrect tone marks (e.g., ငါ vs ငါး). The ToneDisambiguator module uses a specialized context window to resolve these ambiguities.

​

How it Works

It maintains a list of Ambiguous Groups (e.g., the “Three Tones of Ka”). When it encounters a word from such a group, it checks the surrounding +/- 3 words against a set of context patterns. Example 1: သံ (Sound/Iron) vs သုံး (Three)

• Input: သံ ယောက် (Iron person?)
• Context: ယောက် (classifier for people) follows the word.
• Pattern Match: The pattern ("ယောက်", "ခု", "လုံး") is associated with the number သုံး (Three).
• Correction: သုံး ယောက် (Three people).

Example 2: ငါ (I/me) vs ငါး (Fish/Five)

• Input: ငါ ကောင် (I animal?)
• Context: ကောင် (classifier for animals) follows the word.
• Pattern Match: Classifiers for counting animals/fish follow numbers.
• Correction: ငါး ကောင် (Five animals/fish).

Example 3: အစ (Beginning) vs အစာ (Food)

• Input: အစ စား (Eat beginning?)
• Context: စား (eat) follows, and eating requires food (အစာ)
• Pattern Match: The verb စား collocates with အစာ (food), not အစ (beginning).
• Correction: အစာ စား (Eat food)

This system operates alongside the N-gram checker but provides higher confidence for specific, well-known ambiguity patterns.

​

Common Myanmar Grammar Errors Detected