Translation¶

Translate between Bahasa Melayu, English, and Manglish - with entity preservation and formal output.

Why translation?¶

Cross-language content processing, bilingual chatbot support, document translation for Malaysian businesses, and normalizing Manglish for formal contexts. Google Translate struggles with Manglish because it's not a standard language - "xpe la bro" isn't in any training corpus.

malaysian-manglish-nlp uses a rule-based approach that handles the actual patterns Malaysians use.

Load module¶

import malaysian_manglish_nlp as mnlp

# BM to English
result = mnlp.to_english("Saya nak pergi makan nasi lemak")
print(result)
# "I want to go eat nasi lemak"

# English to BM
result = mnlp.to_malay("I want to eat fried rice")
print(result)
# "Saya mahu makan nasi goreng"

Basic usage¶

BM → English¶

mnlp.to_english("Aku nak pergi kedai jap")
# "I want to go to the shop for a moment"

mnlp.to_english("Diaorang semua dah sampai rumah")
# "They all have arrived at the house"

mnlp.to_english("Jangan lupa bawa payung, hujan nanti")
# "Don't forget to bring an umbrella, it will rain later"

English → BM¶

mnlp.to_malay("I am going to the market")
# "Saya sedang pergi ke pasar"

mnlp.to_malay("She already ate lunch")
# "Dia sudah makan tengah hari"

General translate function¶

# Auto-detect and translate
mnlp.translate("Saya suka makan nasi lemak")
# Auto-detects BM → translates to English

mnlp.translate("I like eating coconut rice")
# Auto-detects English → translates to BM

Formal translation¶

Convert informal Manglish to proper formal Bahasa Melayu:

mnlp.to_formal("Aku nk g mkn jap")
# "Saya hendak pergi makan sebentar"

mnlp.to_formal("xpe la bro, aku tunggu je kat sini")
# "Tidak mengapalah, saya tunggu sahaja di sini"

mnlp.to_formal("ko dah makan ke belum?")
# "Awak sudah makan atau belum?"

Formalize vs to_formal

mnlp.formalize() normalizes spelling and grammar. mnlp.to_formal() translates the register to formal BM. Use formalize when you just want clean text; use to_formal when you need proper formal register.

Word-level translation¶

Translate individual words:

mnlp.word_translate("makan")
# "eat"

mnlp.word_translate("pergi")
# "go"

mnlp.word_translate("beautiful")
# "cantik"

Real Manglish examples¶

The rule-based approach handles common Manglish patterns:

# Shortforms expanded during translation
mnlp.to_english("sy nk tny brp harga")
# "I want to ask how much is the price"

# Particles preserved
mnlp.to_english("Sedap lah makanan ni")
# "This food is delicious"

# Code-switched input
mnlp.to_english("I nak order satu teh tarik")
# "I want to order one teh tarik"

# Informal contractions
mnlp.to_english("Diaorg pg kedai mamak")
# "They went to the mamak shop"

Entity preservation¶

Proper nouns and entities are preserved across translation:

mnlp.to_english("Ahmad pergi Petronas Tower KL")
# "Ahmad went to Petronas Tower KL"

mnlp.to_english("Harga Myvi baru RM50,000")
# "The price of new Myvi is RM50,000"

Translation limitations

malaysian-manglish-nlp translation is rule-based, not neural. It excels at common patterns but may struggle with: - Very long or complex sentences - Domain-specific jargon - Idiomatic expressions not in the dictionary - Creative/slang wordplay

For production-grade translation quality, pair with a neural MT system and use malaysian-manglish-nlp for preprocessing.

Detect and translate¶

Automatically detect the input language and translate to the other:

mnlp.detect_and_translate("Saya lapar gila")
# "I am very hungry" (detected BM → EN)

mnlp.detect_and_translate("I am very hungry")
# "Saya sangat lapar" (detected EN → BM)

CLI usage¶

# BM to English
$ mnlp translate "Saya nak makan nasi lemak" --to en
I want to eat nasi lemak

# English to BM
$ mnlp translate "I want to eat" --to bm
Saya mahu makan

# To formal BM
$ mnlp translate "aku nk g mkn" --to formal
Saya hendak pergi makan

# Pipe input
$ echo "Best gila makanan tu" | mnlp translate --to en
That food is incredibly delicious

# Auto-detect
$ mnlp translate "Saya suka makan"
I like to eat

How it works¶

malaysian-manglish-nlp uses a rule-based translation pipeline:

Normalization - expand shortforms (nk→nak, sy→saya)
Tokenization - split into tokens with particle awareness
Dictionary lookup - 15,000+ BM↔EN word pairs
Phrase matching - multi-word expressions ("nasi lemak" stays intact)
Grammar rules - word order transformation (BM SVO → EN SVO with modifier placement)
Entity preservation - proper nouns pass through unchanged

This approach is fast (no model loading), deterministic (same input = same output), and handles Manglish patterns that neural models miss.

Performance¶

Metric	Score
BLEU (BM→EN, news)	42.3
BLEU (BM→EN, social)	35.8
BLEU (EN→BM, news)	38.7
Formal accuracy	87.1%
Throughput	15,000 texts/sec
Latency (single)	< 1ms

BLEU scores context

Rule-based BLEU scores are lower than neural MT on standard benchmarks. However, on Manglish text (which neural MT handles poorly), malaysian-manglish-nlp outperforms Google Translate by a wide margin.