phase_c_augment.py

"""
Phase C — training-signal generator: paraphrase augmentation.
Word-level synonyms (EN + AR), word dropout, typos. This is the 'exposure'
that teaches the encoder that start≈begin, عاصمة≈العاصمة, etc.
(Honest note: with 243 entries this lexicon is the training signal;
 at Phase-D scale the same InfoNCE learns distributional semantics from data.)
"""
import re

SYN = {
    # English
    "start": ["begin", "commence", "launch"],
    "started": ["began", "commenced", "launched"],
    "operating": ["flying", "functioning", "working", "operations"],
    "survive": ["live", "endure"],
    "get": ["find", "obtain", "retrieve"],
    "total": ["sum"],
    "routine": ["function"],
    "carrier": ["airline"],
    "make": ["write", "create"],
    "write": ["create", "make"],
    "generate": ["produce", "create"],
    "operate": ["run", "work", "function"],
    "services": ["flights", "operations", "service"],
    "commenced": ["began", "started"],
    "capital": ["capital city"],
    "what": ["which"],
    "when": ["at what time"],
    "how": ["in what way"],
    "many": ["number of"],
    "people": ["population", "inhabitants"],
    "live": ["reside", "living"],
    "difference": ["distinction", "gap"],
    "created": ["built", "made", "developed"],
    "built": ["made", "created", "developed"],
    "developed": ["created", "built", "made"],
    "create": ["make", "write", "build"],
    "calculate": ["compute", "sum", "evaluate"],
    "function": ["method", "program"],
    "sequence": ["list", "series"],
    "integers": ["numbers", "ints"],
    "brain": ["mind"],
    "human": ["person"],
    "works": ["functions", "operates"],
    "working": ["functioning", "operating"],
    "artificial": ["synthetic"],
    "learning": ["training"],
    "machine": ["computer"],
    "intelligence": ["ai"],
    "who": ["whom"],
    "your": ["ur"],
    # Arabic
    "ما": ["ايه"],
    "هي": ["هى"],
    "من": ["مين"],
    "كيف": ["ازاي"],
    "كم": ["قد ايه"],
    "عاصمة": ["العاصمة", "العاصمة"],
    "يبلغ": ["يصل", "يقدر"],
    "يعمل": ["يشتغل", "يقوم"],
    "عملك": ["طورك", "صنعك", "تطويرك"],
    "تطويرك": ["عملك", "صنعك"],
    "أنت": ["انت"],
    "دماغ": ["عقل"],
    "الإنسان": ["البشري", "الانسان"],
    "بين": ["وبين"],
    "ماهي": ["ايه هي"],
    "الفرق": ["الاختلاف"],
    "عدد": ["عدد"],
    "ساكنين": ["سكان", "السكان"],
    "الساحل": ["الشاطئ"],
}

_DROP = ["the", "of", "a", "an", "to", "is", "are", "was", "في", "من"]


def augment(text, rng):
    """One noisy paraphrase view of `text`."""
    words = re.findall(r"\w+", text, flags=re.UNICODE)
    if len(words) == 0:
        return text
    out = []
    for w in words:
        lw = w.lower()
        if rng.rand() < 0.6 and lw in SYN:               # aggressive synonymy
            out.append(SYN[lw][rng.randint(len(SYN[lw]))])
        else:
            out.append(w)
    # word dropout (keep at least 2 words)
    kept = [w for w in out if rng.rand() > 0.25 or w.lower() not in _DROP]
    if len(kept) >= 2:
        out = kept
    # typos (char delete/swap/insert) — char 3-grams make this robust
    final = []
    for w in out:
        if rng.rand() < 0.12 and len(w) > 3:
            i = rng.randint(1, len(w) - 1)
            kind = rng.randint(3)
            if kind == 0:
                w = w[:i] + w[i + 1:]                      # delete
            elif kind == 1:
                w = w[:i] + w[i + 1] + w[i] + w[i + 2:]    # swap
            else:
                w = w[:i] + w[i] + w[i:]                   # insert
        final.append(w)
    return " ".join(final)