What do these changes do

Moves module-level imports of heavy external dependencies (transformers, torch, gensim, numpy, spacy) to class initialization methods, enabling module imports without requiring optional dependencies.

What was wrong

Module-level imports forced dependency resolution at import time, blocking:

• Documentation tools from importing modules without full dependency installation
• Module-level helper functions from working independently
• Clear error context (failures at import vs. instantiation)

Example failure case:

from pythainlp.word_vector import core  # Fails if gensim not installed
# Even if you only wanted to use a helper function, not WordVector class

How this fixes it

Defers dependency imports to where they're actually needed:

Pattern 1: Import in __init__

class KeyBERT:
    def __init__(self, model_name: str):
        from transformers import pipeline  # Loaded only when class instantiated
        self.ft_pipeline = pipeline(...)

Pattern 2: Import in method

def sentence_vectorizer(self, text: str) -> ndarray:
    from numpy import zeros  # Loaded only when method called
    vec = zeros((1, self.WV_DIM))

Pattern 3: Lazy module-level via getter

_tokenizer = None

def _get_tokenizer():
    global _tokenizer
    if _tokenizer is None:
        from transformers import CamembertTokenizer
        _tokenizer = CamembertTokenizer.from_pretrained(...)
    return _tokenizer

Used TYPE_CHECKING guards to preserve type hints without runtime imports. Updated type hints to use native Python 3.9+ lowercase types (list, dict) with Union from typing module for compatibility with Python 3.9 strict mode and type inspection tools.

Modified modules (9):

• word_vector/core.py - gensim, numpy
• summarize/keybert.py - numpy, transformers
• summarize/mt5.py - transformers
• wangchanberta/core.py - transformers (tokenizer)
• coref/_fastcoref.py - spacy
• coref/han_coref.py - spacy
• parse/transformers_ud.py - transformers, numpy, torch, ufal (uses Union[list[list[str]], str] syntax)
• augment/lm/wangchanberta.py - transformers
• augment/lm/fasttext.py - gensim

Your checklist for this pull request

• Passed code styles and structures
• Passed code linting checks and unit test

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.