File size: 14,278 Bytes
f6f45d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fed70a
f6f45d5
 
 
 
 
 
 
 
 
1fed70a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6f45d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fed70a
f6f45d5
 
1fed70a
 
 
 
 
f6f45d5
 
 
1fed70a
f6f45d5
 
 
 
 
 
 
 
 
 
1fed70a
 
 
f6f45d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fed70a
f6f45d5
 
 
 
 
1fed70a
 
 
 
 
 
f6f45d5
 
 
 
1fed70a
 
 
 
 
 
 
 
 
f6f45d5
 
 
 
 
 
 
 
 
1fed70a
 
 
 
 
f6f45d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fed70a
f6f45d5
 
 
 
 
 
1fed70a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6f45d5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fed70a
f6f45d5
1fed70a
 
 
 
 
f6f45d5
1fed70a
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
"""
SinCode v3 β€” ByT5 Seq2Seq + XLM-RoBERTa MLM Reranker.

Pipeline (per word):
  Sinhala script  β†’ MLM scores in context (single candidate)
  English vocab   β†’ ByT5 generates Sinhala alternatives + English kept; MLM picks
  Everything else β†’ ByT5 generates top-5 candidates; MLM picks best
"""

import math
import re
import torch
import logging
from typing import List, Tuple, Optional

from transformers import AutoTokenizer, AutoModelForMaskedLM

from core.constants import (
    DEFAULT_MLM_MODEL, DEFAULT_BYT5_MODEL,
    MAX_CANDIDATES, MIN_ENGLISH_LEN, MIN_ENGLISH_PASSTHROUGH_LEN,
    PUNCT_PATTERN,
)
from core.english import ENGLISH_VOCAB
from seq2seq.infer import Transliterator

logger = logging.getLogger(__name__)

_SINHALA_RE = re.compile(r"[\u0D80-\u0DFF]")

# ── Numeric / special-token passthrough ──────────────────────────────────────
# These patterns detect tokens that must not go through ByT5 transliteration.
_RE_ORDINAL   = re.compile(r"^\d+(st|nd|rd|th)$", re.IGNORECASE)   # 1st, 2nd, 3rd
_RE_PURE_NUM  = re.compile(r"^\d+(?:[.,]\d+)*$")                    # 5, 10.30, 9.00
_RE_NUM_RANGE = re.compile(r"^\d+-\d+$")                            # 2-3, 10-20
_RE_CURRENCY  = re.compile(r"^\d+[/]-?$")                           # 500/-
_RE_AM_PM     = re.compile(r"^[ap]\.?m\.?$", re.IGNORECASE)        # a.m. p.m. am pm
_RE_NUM_PCT_K = re.compile(r"^(\d[\d,.]*%+)k$", re.IGNORECASE)     # 100%k β†’ 100%්࢚
_RE_NUM_K     = re.compile(r"^(\d[\d,.]*)k$", re.IGNORECASE)        # 5000k β†’ 5000්࢚
_RE_NUM_TA    = re.compile(r"^(\d[\d,.]*)ta$", re.IGNORECASE)       # 10.30ta β†’ 10.30ΰΆ§


def _numeric_passthrough(core: str) -> Optional[str]:
    """
    Return a (possibly lightly-transformed) value for numeric/special tokens.
    Returns None if the token should go through normal ByT5 + MLM processing.
    """
    # 100%k β†’ 100%්࢚  (check before plain Nk to avoid consuming the %)
    m = _RE_NUM_PCT_K.match(core)
    if m:
        return m.group(1) + "්࢚"
    # 5000k / 10k β†’ 5000්࢚ / 10්࢚
    m = _RE_NUM_K.match(core)
    if m:
        return m.group(1) + "්࢚"
    # 10.30ta β†’ 10.30ΰΆ§
    m = _RE_NUM_TA.match(core)
    if m:
        return m.group(1) + "ΰΆ§"
    # Ordinals: 1st, 2nd, 3rd … β†’ keep as-is
    if _RE_ORDINAL.match(core):
        return core
    # Pure numbers and decimals: 5, 100, 10.30, 9.00 β†’ keep
    if _RE_PURE_NUM.match(core):
        return core
    # Number ranges: 2-3 β†’ keep
    if _RE_NUM_RANGE.match(core):
        return core
    # Currency notation: 500/- β†’ keep
    if _RE_CURRENCY.match(core):
        return core
    # AM/PM markers: a.m. p.m. am pm β†’ keep
    if _RE_AM_PM.match(core):
        return core
    return None


class ScoredCandidate:
    __slots__ = ("text", "mlm_score")

    def __init__(self, text: str, mlm_score: float):
        self.text = text
        self.mlm_score = mlm_score


def _is_sinhala(text: str) -> bool:
    return bool(_SINHALA_RE.search(text))


class BeamSearchDecoder:
    """
    SinCode v3 contextual decoder.

    Replaces the rule engine + dictionary + hardcoded maps with a single
    ByT5-small seq2seq model fine-tuned on 1,000,000 Singlish→Sinhala pairs.
    XLM-RoBERTa reranks the top-5 beam candidates by masked-LM probability.
    """

    def __init__(
        self,
        mlm_model_name: str = DEFAULT_MLM_MODEL,
        byt5_model_path: str = DEFAULT_BYT5_MODEL,
        device: Optional[str] = None,
    ):
        self.device = device or ("cuda" if torch.cuda.is_available() else "cpu")

        logger.info("Loading MLM reranker: %s", mlm_model_name)
        self.tokenizer = AutoTokenizer.from_pretrained(mlm_model_name)
        self.model = AutoModelForMaskedLM.from_pretrained(mlm_model_name)
        self.model.to(self.device)
        self.model.eval()

        logger.info("Loading ByT5 transliterator: %s", byt5_model_path)
        self.transliterator = Transliterator(model_path=byt5_model_path, device=self.device)

    # ── Normalization ─────────────────────────────────────────────────────────

    @staticmethod
    def _softmax_normalize(raw_scores: List[float]) -> List[float]:
        if not raw_scores:
            return []
        if len(raw_scores) == 1:
            return [1.0]
        max_s = max(raw_scores)
        exps = [math.exp(s - max_s) for s in raw_scores]
        total = sum(exps)
        return [e / total for e in exps]

    # ── MLM batch scoring ─────────────────────────────────────────────────────

    def _batch_mlm_score(
        self,
        left_contexts: List[str],
        right_contexts: List[str],
        candidates: List[str],
    ) -> List[float]:
        """Score each candidate with XLM-RoBERTa multi-mask log-probability."""
        if not candidates:
            return []

        mask = self.tokenizer.mask_token
        mask_token_id = self.tokenizer.mask_token_id

        cand_token_ids: List[List[int]] = []
        for c in candidates:
            ids = self.tokenizer.encode(c, add_special_tokens=False)
            cand_token_ids.append(ids if ids else [self.tokenizer.unk_token_id])

        batch_texts: List[str] = []
        for i in range(len(candidates)):
            n_masks = len(cand_token_ids[i])
            mask_str = " ".join([mask] * n_masks)
            parts = [p for p in [left_contexts[i], mask_str, right_contexts[i]] if p]
            batch_texts.append(" ".join(parts))

        inputs = self.tokenizer(
            batch_texts,
            return_tensors="pt",
            padding=True,
            truncation=True,
        ).to(self.device)

        with torch.no_grad():
            logits = self.model(**inputs).logits

        scores: List[float] = []
        for i, target_ids in enumerate(cand_token_ids):
            token_ids = inputs.input_ids[i]
            mask_positions = (token_ids == mask_token_id).nonzero(as_tuple=True)[0]

            if mask_positions.numel() == 0 or not target_ids:
                scores.append(-100.0)
                continue

            n = min(len(target_ids), mask_positions.numel())
            total = 0.0
            for j in range(n):
                pos = mask_positions[j].item()
                log_probs = torch.log_softmax(logits[i, pos, :], dim=0)
                total += log_probs[target_ids[j]].item()

            scores.append(total / n)

        return scores

    # ── Public decode ─────────────────────────────────────────────────────────

    def decode(self, sentence: str) -> Tuple[str, List[str], List[Tuple[str, List[str]]]]:
        """
        Decode a Singlish sentence word-by-word using ByT5 + XLM-RoBERTa MLM.
        Returns (transliterated_sentence, trace_logs, word_candidates).

        word_candidates is a list of (selected_word, [all_candidates]) per word,
        in input word order. Single-candidate words (passthrough/sinhala) have an
        empty alternatives list.
        """
        words = sentence.split()
        if not words:
            return "", [], []

        # ── Phase 1: batch ByT5 candidate generation ──────────────────────────
        # Collect only the words that need ByT5 (non-Sinhala), run in one pass
        cores: List[str] = []
        core_meta: List[tuple] = []  # (index_into_words, prefix, core, suffix, core_lower)

        for i, raw in enumerate(words):
            match = PUNCT_PATTERN.match(raw)
            prefix, core, suffix = match.groups() if match else ("", raw, "")
            if not _is_sinhala(core):
                # Skip numeric/special tokens β€” they don't need ByT5
                if _numeric_passthrough(core) is not None:
                    continue
                cores.append(core)
                core_meta.append((i, prefix, core, suffix, core.lower()))

        # Single ByT5 forward pass for all non-Sinhala words
        byt5_results: List[List[str]] = (
            self.transliterator.batch_candidates(cores, k=MAX_CANDIDATES)
            if cores else []
        )

        byt5_map: dict = {}  # word index β†’ list of raw ByT5 strings
        for (i, prefix, core, suffix, core_lower), cands in zip(core_meta, byt5_results):
            byt5_map[i] = (prefix, suffix, core_lower, cands or [core])

        word_infos: List[dict] = []
        for i, raw in enumerate(words):
            match = PUNCT_PATTERN.match(raw)
            raw_prefix, core, raw_suffix = match.groups() if match else ("", raw, "")

            if _is_sinhala(core):
                word_infos.append({"kind": "sinhala", "candidates": [raw]})
                continue

            # Numeric/special passthrough β€” keep or lightly transform, skip MLM
            passthrough_val = _numeric_passthrough(core)
            if passthrough_val is not None:
                word_infos.append({"kind": "passthrough", "candidates": [raw_prefix + passthrough_val + raw_suffix]})
                continue

            prefix, suffix, core_lower, byt5_cands = byt5_map[i]
            sinhala_cands = [prefix + c + suffix for c in byt5_cands]

            if core_lower in ENGLISH_VOCAB and len(core_lower) >= MIN_ENGLISH_LEN:
                if len(core_lower) >= MIN_ENGLISH_PASSTHROUGH_LEN:
                    # Long English loanwords (β‰₯6 chars): unambiguously English in
                    # code-mixed Singlish β€” bypass MLM to avoid Sinhala-bias override.
                    word_infos.append({"kind": "english", "candidates": [raw]})
                else:
                    # Short English vocab words (3–5 chars) may also be Singlish
                    # homophones (mage/mama/game/call) β€” let MLM disambiguate.
                    candidates = [raw] + [c for c in sinhala_cands if c != raw]
                    word_infos.append({"kind": "singlish", "candidates": candidates[:MAX_CANDIDATES + 1]})
            else:
                word_infos.append({"kind": "singlish", "candidates": sinhala_cands})

        # ── Phase 2: greedy left-to-right pass (builds dynamic left context) ──
        # Right context is seeded from first ByT5 candidate (pre-decode estimate)
        stable_right = [info["candidates"][0] for info in word_infos]
        selected_words: List[str] = []

        for t, info in enumerate(word_infos):
            # English-detected words: always keep raw form β€” skip MLM (MLM Sinhala bias
            # would otherwise score a Sinhala transliteration higher than the English token)
            if info["kind"] in ("english", "passthrough", "sinhala"):
                selected_words.append(info["candidates"][0])
                continue
            candidates = info["candidates"]
            left_ctx = " ".join(selected_words)
            right_ctx = " ".join(stable_right[t + 1:])
            raw_mlm = self._batch_mlm_score(
                [left_ctx] * len(candidates),
                [right_ctx] * len(candidates),
                candidates,
            )
            norm_mlm = self._softmax_normalize(raw_mlm)
            best = max(zip(candidates, norm_mlm), key=lambda x: x[1])
            selected_words.append(best[0])

        # ── Phase 3: re-score with full decoded sentence as context ───────────
        # Right context is now the actual decoded output, not the pre-decode estimate
        trace_logs: List[str] = []
        final_words: List[str] = []
        word_candidates: List[Tuple[str, List[str]]] = []  # (selected, [all_cands])

        for t, info in enumerate(word_infos):
            raw_word = words[t]
            kind = info["kind"]
            candidates = info["candidates"]

            # English-detected, Sinhala, and passthrough words bypass MLM scoring.
            # For English: the MLM model is Sinhala-biased and would otherwise prefer
            # a Sinhala transliteration over the correct English token.
            if kind == "sinhala":
                final_words.append(candidates[0])
                word_candidates.append((candidates[0], []))
                trace_logs.append(
                    f"**Step {t+1}: `{raw_word}`** β†’ `{candidates[0]}` (Sinhala passthrough)\n"
                )
                continue
            if kind == "passthrough":
                final_words.append(candidates[0])
                word_candidates.append((candidates[0], []))
                trace_logs.append(
                    f"**Step {t+1}: `{raw_word}`** β†’ `{candidates[0]}` (numeric/passthrough)\n"
                )
                continue
            if kind == "english":
                final_words.append(candidates[0])
                word_candidates.append((candidates[0], []))
                trace_logs.append(
                    f"**Step {t+1}: `{raw_word}`** β†’ `{candidates[0]}` (English vocab passthrough)\n"
                )
                continue

            left_ctx = " ".join(final_words)
            right_ctx = " ".join(selected_words[t + 1:])

            raw_mlm = self._batch_mlm_score(
                [left_ctx] * len(candidates),
                [right_ctx] * len(candidates),
                candidates,
            )
            norm_mlm = self._softmax_normalize(raw_mlm)

            scored = sorted(
                [ScoredCandidate(text=c, mlm_score=norm_mlm[i]) for i, c in enumerate(candidates)],
                key=lambda x: x.mlm_score,
                reverse=True,
            )
            best = scored[0]
            final_words.append(best.text)
            word_candidates.append((best.text, [s.text for s in scored]))

            trace_logs.append(
                f"**Step {t+1}: `{raw_word}`** β†’ `{best.text}` "
                f"(MLM={best.mlm_score:.3f})\n"
                + "\n".join(f"  - `{s.text}` {s.mlm_score:.3f}" for s in scored)
            )

        return " ".join(final_words), trace_logs, word_candidates