Thai Text Processing

Thai Text Processing (Segmentation, Normalization, Collation, Romanization)

Overview

Thai script has no spaces between words within a sentence — "ฉันรักการเขียนโค้ด".split(" ") returns the entire string as one token, not a word list, because there is nothing to split on. A model without this skill reaches for whitespace-based tokenization out of habit (it works for English) and silently produces garbage for Thai. This skill flags that failure mode and points to the correct tools — it does not reimplement a Thai tokenizer or transliterator, since both need a maintained dictionary/model to be accurate, not hand-written rules.

When to use

• ตัดคำภาษาไทยเป็นคำๆ (word segmentation/tokenization) สำหรับ search, NLP, หรือ text analytics
• Normalizing Thai text before comparison, search, deduplication, or storage (Unicode NFC)
• Sorting Thai strings correctly (Thai collation order differs from raw Unicode codepoint order)
• Romanizing Thai text to Latin script (e.g. for a passport-style name, a street sign, or an international form) using the RTGS (Royal Thai General System of Transcription) convention

When NOT to use

• Plain reading/writing/translating Thai text with no segmentation, sorting, or comparison step involved
• The romanization only needs to be "readable," not standards-compliant (e.g. casual transliteration in a chat) — RTGS precision isn't necessary there

Core knowledge

Word segmentation: Thai does not use spaces between words (spaces mark clause/sentence boundaries, not word boundaries). Segmentation requires a dictionary- or model-based tokenizer — PyThaiNLP (pip install pythainlp, pythainlp.tokenize.word_tokenize) is the standard open-source option. This skill does not vendor or reimplement a tokenizer; a hand-rolled rule-based splitter would silently produce wrong word boundaries on real text.

Unicode normalization (NFC): Thai vowels and tone marks are combining Unicode characters. The same visible text can be encoded as a single precomposed codepoint or as a base character plus a combining mark — two strings that look identical can compare unequal, fail a lookup, or break deduplication if they aren't normalized to the same form first. Normalize to NFC (unicodedata.normalize("NFC", text), pure Python stdlib, no dependency needed) before comparing, searching, hashing, or storing user-submitted Thai text.

Collation (sorting): Thai alphabetical order is not the same as sorting by raw Unicode codepoint value — vowels that visually appear before their consonant (เ, แ, โ, ใ, ไ) are still ordered by the consonant they attach to, and tone marks/other combining characters sort after the base character rather than by codepoint. A naive sorted() on raw strings gives an order a Thai reader won't recognize as alphabetical. Correct collation needs locale-aware sorting (e.g. Python's locale module with a Thai locale, or ICU-based collation via a library like PyICU) — this skill flags the need, it does not implement a full Thai collation table by hand.

Romanization (RTGS): the Royal Institute's RTGS (Royal Thai General System of Transcription) is Thailand's official Thai-to-Latin transliteration standard, used on road signs and in official documents. It is not a simple 1:1 letter-substitution — the same Thai consonant can romanize differently depending on its position in a syllable (initial vs. final), and several distinct Thai sounds collapse to the same Latin letter (RTGS deliberately drops tone marks). Use a maintained implementation (e.g. PyThaiNLP's romanize() function) rather than hand-writing substitution rules — the exception list is large enough that an ad hoc mapping will be wrong on real names/addresses often enough to matter.

Common mistakes

1. Using .split(" ") (or any whitespace-based split) as if it segments Thai words — it doesn't, because Thai doesn't use spaces between words.
2. Comparing or deduplicating Thai strings without normalizing to NFC first, causing visually-identical text to be treated as different.
3. Sorting Thai text with a plain codepoint-based sorted() and presenting it as "alphabetical order" — a Thai reader will notice it's wrong, particularly around the leading vowels (เ/แ/โ/ใ/ไ).
4. Hand-writing a Thai-to-Latin transliteration table instead of using a maintained RTGS implementation — the exceptions are numerous enough that ad hoc rules produce visibly wrong romanizations on real names.
5. Assuming a segmented word list can be rejoined with spaces to reconstruct "readable" Thai — spaces between every word are not natural Thai writing; only use segmentation output for machine processing (search indexes, NLP pipelines), not for display back to a Thai reader.

Code

examples/normalize.py — pure stdlib NFC normalization, no dependency. Run python3 normalize.py for the self-test.

examples/segmentation.py — demonstrates the naive-split failure (dependency-free) and wraps PyThaiNLP's tokenizer if installed (pip install pythainlp). Run python3 segmentation.py; the tokenizer half of the self-test runs only if PyThaiNLP is present, and reports a clear skip message if not.

Known limitations

• This skill deliberately does not vendor PyThaiNLP or any transliteration table — those are maintained third-party projects with their own update cadence (dictionary coverage, model accuracy); pin and install them as a real dependency in the target project rather than copying code out of this skill.
• Collation is described here, not implemented — there's no dependency-free stdlib equivalent to a full Thai ICU collation table worth hand-rolling.
• RTGS romanization is lossy by design (drops tone information) — it is the official transliteration standard, not necessarily the most phonetically precise one; if the task needs tone-preserving phonetic transcription, RTGS is the wrong tool.