FangSen9000 commited on
Commit
a16e5a2
·
1 Parent(s): 8af7be8

Add semantic text-to-gloss option

Browse files
Files changed (3) hide show
  1. index.js +117 -6
  2. pipeline01_text2gloss.py +136 -1
  3. viser_backend.py +11 -4
index.js CHANGED
@@ -95,6 +95,30 @@ function ensureStyles() {
95
  font-size: 16px;
96
  color: #0f172a;
97
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98
  .stable-video-frame {
99
  position: relative;
100
  width: 100%;
@@ -190,6 +214,30 @@ function ensureStyles() {
190
  .stable-control-actions button {
191
  flex: 1;
192
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
193
  .stable-status-chip {
194
  font-size: 13px;
195
  padding: 6px 12px;
@@ -628,6 +676,7 @@ export default class StableSignerPlugin extends ControlWavePlugin {
628
  <div class="stable-control-actions">
629
  <button id="stable-prewarm" class="btn secondary">Prewarm</button>
630
  <button id="stable-translate" class="btn secondary">Text → Available Gloss</button>
 
631
  </div>
632
  <span class="stable-status-chip" id="stable-status">Idle</span>
633
  </div>
@@ -635,6 +684,10 @@ export default class StableSignerPlugin extends ControlWavePlugin {
635
  <label>Prompt / Text</label>
636
  <textarea id="stable-prompt-text" placeholder="Describe the action or context; the system will translate it to gloss."></textarea>
637
  <div class="stable-hint">Input text auto-translates and fills the uppercase gloss field below (still editable).</div>
 
 
 
 
638
  <div class="form-group" style="margin-top:8px;font-size:13px;color:#475569;">
639
  <div>Prompt samples (picked from the latest prompt2gloss eval logs with strongest gloss matches):</div>
640
  <div>1. Can you provide instruction on the sign language for &ldquo;You&rsquo;re going to hit it just a little more behind.&rdquo;?</div>
@@ -752,7 +805,11 @@ export default class StableSignerPlugin extends ControlWavePlugin {
752
  <div class="stable-text-playground stable-signer-panel">
753
  <div class="stable-text-playground-title">
754
  <h3>Selection Translation Playground</h3>
755
- <span class="stable-status-chip">Select text to translate</span>
 
 
 
 
756
  </div>
757
  <div id="stable-selection-template" class="stable-reading-area stable-dialogue-area" tabindex="0" contenteditable="true">
758
  <div class="stable-dialogue-turn question">
@@ -820,6 +877,10 @@ export default class StableSignerPlugin extends ControlWavePlugin {
820
  this.refineLogsPre = container.querySelector('#stable-refine-logs');
821
  this.prewarmBtn = container.querySelector('#stable-prewarm');
822
  this.translateBtn = container.querySelector('#stable-translate');
 
 
 
 
823
  this.selectionTemplate = container.querySelector('#stable-selection-template');
824
  this.selectionPopover = container.querySelector('#stable-selection-popover');
825
  this.selectionTranslateBtn = container.querySelector('#stable-selection-translate');
@@ -873,6 +934,9 @@ export default class StableSignerPlugin extends ControlWavePlugin {
873
  attachEvents() {
874
  this.prewarmBtn.addEventListener('click', () => this.handlePrewarm());
875
  this.translateBtn.addEventListener('click', () => this.handleTranslate());
 
 
 
876
  this.generateBtn.addEventListener('click', () => this.handleGenerate());
877
  this.videoRefreshBtn.addEventListener('click', () => this.refreshRecentVideos());
878
  this.videoSelect.addEventListener('change', () => this.handleRecentSelection());
@@ -885,6 +949,9 @@ export default class StableSignerPlugin extends ControlWavePlugin {
885
  this.selectionTemplate.addEventListener('mouseup', () => this.handleTemplateSelection());
886
  this.selectionTemplate.addEventListener('keyup', () => this.handleTemplateSelection());
887
  }
 
 
 
888
  if (this.selectionTranslateBtn) {
889
  this.selectionTranslateBtn.addEventListener('click', () => this.handleSelectedTextGenerate(true));
890
  }
@@ -929,7 +996,8 @@ export default class StableSignerPlugin extends ControlWavePlugin {
929
  refImagePath: this.refImageInput.value,
930
  normalizePose: this.normalizeCheckbox.checked,
931
  hideTorsoLines: this.hideTorsoCheckbox.checked,
932
- translationMode: 'text2gloss'
 
933
  };
934
  }
935
 
@@ -1040,11 +1108,18 @@ export default class StableSignerPlugin extends ControlWavePlugin {
1040
  this.setTranslateLoading(true);
1041
  }
1042
  try {
1043
- const gloss = await this.translatePromptToGloss(normalizedPrompt, maxTokens, promptTextRaw);
 
1044
  this.glossInput.value = gloss;
1045
  if (!isAuto) {
1046
  this.updateStatus('Gloss translation completed');
1047
- this.logsPre.textContent = `Translation result: ${gloss}`;
 
 
 
 
 
 
1048
  }
1049
  return gloss;
1050
  } finally {
@@ -1059,12 +1134,15 @@ export default class StableSignerPlugin extends ControlWavePlugin {
1059
  throw new Error('Prompt/Text is empty');
1060
  }
1061
  const payloadPrompt = originalPrompt || promptText;
 
1062
  const response = await window.controlWaveWebSocket.sendRequest(this.id, {
1063
  action: 'text_to_gloss',
1064
  data: {
1065
  text: promptText,
1066
  prompt: payloadPrompt,
1067
- maxNewTokens
 
 
1068
  }
1069
  });
1070
  if (!response || response.status !== 'success' || !response.gloss) {
@@ -1073,7 +1151,37 @@ export default class StableSignerPlugin extends ControlWavePlugin {
1073
  if (response.dropped && response.dropped.length > 0) {
1074
  console.info('[StableSigner] dropped non-renderable text tokens', response.dropped);
1075
  }
1076
- return response.gloss;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1077
  }
1078
 
1079
  handleTemplateSelection() {
@@ -1154,6 +1262,9 @@ export default class StableSignerPlugin extends ControlWavePlugin {
1154
  throw new Error('WebSocket not ready');
1155
  }
1156
  const payload = this.collectPayload();
 
 
 
1157
  payload.maxNewTokens = Math.min(toInt(this.maxNewTokensInput.value, 64), 18);
1158
  payload.maxCandidates = Math.min(toInt(this.maxCandidatesInput.value, 10), 3);
1159
  const response = await window.controlWaveWebSocket.sendRequest(this.id, {
 
95
  font-size: 16px;
96
  color: #0f172a;
97
  }
98
+ .stable-playground-actions {
99
+ display: flex;
100
+ align-items: center;
101
+ justify-content: flex-end;
102
+ gap: 10px;
103
+ flex-wrap: wrap;
104
+ }
105
+ .stable-playground-actions label {
106
+ display: inline-flex;
107
+ align-items: center;
108
+ gap: 6px;
109
+ margin: 0;
110
+ font-size: 13px;
111
+ font-weight: 600;
112
+ color: #334155;
113
+ }
114
+ .stable-playground-actions input {
115
+ width: auto;
116
+ margin: 0;
117
+ }
118
+ .stable-playground-actions button {
119
+ padding: 6px 10px;
120
+ white-space: nowrap;
121
+ }
122
  .stable-video-frame {
123
  position: relative;
124
  width: 100%;
 
214
  .stable-control-actions button {
215
  flex: 1;
216
  }
217
+ .stable-translation-options {
218
+ display: flex;
219
+ align-items: center;
220
+ justify-content: space-between;
221
+ gap: 10px;
222
+ margin-top: 8px;
223
+ font-size: 13px;
224
+ color: #334155;
225
+ }
226
+ .stable-translation-options label {
227
+ display: inline-flex;
228
+ align-items: center;
229
+ gap: 6px;
230
+ margin: 0;
231
+ font-weight: 600;
232
+ }
233
+ .stable-translation-options input {
234
+ width: auto;
235
+ margin: 0;
236
+ }
237
+ .stable-translation-options button {
238
+ white-space: nowrap;
239
+ padding: 6px 10px;
240
+ }
241
  .stable-status-chip {
242
  font-size: 13px;
243
  padding: 6px 12px;
 
676
  <div class="stable-control-actions">
677
  <button id="stable-prewarm" class="btn secondary">Prewarm</button>
678
  <button id="stable-translate" class="btn secondary">Text → Available Gloss</button>
679
+ <button id="stable-semantic-demo" class="btn secondary">Semantic Demo</button>
680
  </div>
681
  <span class="stable-status-chip" id="stable-status">Idle</span>
682
  </div>
 
684
  <label>Prompt / Text</label>
685
  <textarea id="stable-prompt-text" placeholder="Describe the action or context; the system will translate it to gloss."></textarea>
686
  <div class="stable-hint">Input text auto-translates and fills the uppercase gloss field below (still editable).</div>
687
+ <div class="stable-translation-options">
688
+ <label><input type="checkbox" id="stable-semantic-gloss"> Semantic constrained gloss</label>
689
+ <span>Uses conservative synonym fallback; output remains limited to available poses.</span>
690
+ </div>
691
  <div class="form-group" style="margin-top:8px;font-size:13px;color:#475569;">
692
  <div>Prompt samples (picked from the latest prompt2gloss eval logs with strongest gloss matches):</div>
693
  <div>1. Can you provide instruction on the sign language for &ldquo;You&rsquo;re going to hit it just a little more behind.&rdquo;?</div>
 
805
  <div class="stable-text-playground stable-signer-panel">
806
  <div class="stable-text-playground-title">
807
  <h3>Selection Translation Playground</h3>
808
+ <div class="stable-playground-actions">
809
+ <label><input type="checkbox" id="stable-selection-semantic-gloss" checked> Semantic gloss</label>
810
+ <button id="stable-selection-semantic-demo" type="button" class="btn secondary">Semantic Demo</button>
811
+ <span class="stable-status-chip">Select text to translate</span>
812
+ </div>
813
  </div>
814
  <div id="stable-selection-template" class="stable-reading-area stable-dialogue-area" tabindex="0" contenteditable="true">
815
  <div class="stable-dialogue-turn question">
 
877
  this.refineLogsPre = container.querySelector('#stable-refine-logs');
878
  this.prewarmBtn = container.querySelector('#stable-prewarm');
879
  this.translateBtn = container.querySelector('#stable-translate');
880
+ this.semanticDemoBtn = container.querySelector('#stable-semantic-demo');
881
+ this.semanticGlossCheckbox = container.querySelector('#stable-semantic-gloss');
882
+ this.selectionSemanticCheckbox = container.querySelector('#stable-selection-semantic-gloss');
883
+ this.selectionSemanticDemoBtn = container.querySelector('#stable-selection-semantic-demo');
884
  this.selectionTemplate = container.querySelector('#stable-selection-template');
885
  this.selectionPopover = container.querySelector('#stable-selection-popover');
886
  this.selectionTranslateBtn = container.querySelector('#stable-selection-translate');
 
934
  attachEvents() {
935
  this.prewarmBtn.addEventListener('click', () => this.handlePrewarm());
936
  this.translateBtn.addEventListener('click', () => this.handleTranslate());
937
+ if (this.semanticDemoBtn) {
938
+ this.semanticDemoBtn.addEventListener('click', () => this.handleSemanticDemo());
939
+ }
940
  this.generateBtn.addEventListener('click', () => this.handleGenerate());
941
  this.videoRefreshBtn.addEventListener('click', () => this.refreshRecentVideos());
942
  this.videoSelect.addEventListener('change', () => this.handleRecentSelection());
 
949
  this.selectionTemplate.addEventListener('mouseup', () => this.handleTemplateSelection());
950
  this.selectionTemplate.addEventListener('keyup', () => this.handleTemplateSelection());
951
  }
952
+ if (this.selectionSemanticDemoBtn) {
953
+ this.selectionSemanticDemoBtn.addEventListener('click', () => this.handleSelectionSemanticDemo());
954
+ }
955
  if (this.selectionTranslateBtn) {
956
  this.selectionTranslateBtn.addEventListener('click', () => this.handleSelectedTextGenerate(true));
957
  }
 
996
  refImagePath: this.refImageInput.value,
997
  normalizePose: this.normalizeCheckbox.checked,
998
  hideTorsoLines: this.hideTorsoCheckbox.checked,
999
+ translationMode: this.semanticGlossCheckbox && this.semanticGlossCheckbox.checked ? 'semantic' : 'text2gloss',
1000
+ useSemantic: !!(this.semanticGlossCheckbox && this.semanticGlossCheckbox.checked)
1001
  };
1002
  }
1003
 
 
1108
  this.setTranslateLoading(true);
1109
  }
1110
  try {
1111
+ const result = await this.translatePromptToGloss(normalizedPrompt, maxTokens, promptTextRaw);
1112
+ const gloss = typeof result === 'string' ? result : result.gloss;
1113
  this.glossInput.value = gloss;
1114
  if (!isAuto) {
1115
  this.updateStatus('Gloss translation completed');
1116
+ const semantic = result.semantic && result.semantic.length > 0
1117
+ ? `\nSemantic fallback: ${result.semantic.map(([src, dst]) => `${src}→${dst}`).join(', ')}`
1118
+ : '';
1119
+ const dropped = result.dropped && result.dropped.length > 0
1120
+ ? `\nDropped: ${result.dropped.join(', ')}`
1121
+ : '';
1122
+ this.logsPre.textContent = `Translation result: ${gloss}\nMode: ${result.mode || 'rule'}${semantic}${dropped}`;
1123
  }
1124
  return gloss;
1125
  } finally {
 
1134
  throw new Error('Prompt/Text is empty');
1135
  }
1136
  const payloadPrompt = originalPrompt || promptText;
1137
+ const useSemantic = !!(this.semanticGlossCheckbox && this.semanticGlossCheckbox.checked);
1138
  const response = await window.controlWaveWebSocket.sendRequest(this.id, {
1139
  action: 'text_to_gloss',
1140
  data: {
1141
  text: promptText,
1142
  prompt: payloadPrompt,
1143
+ maxNewTokens,
1144
+ mode: useSemantic ? 'semantic' : 'rule',
1145
+ useSemantic
1146
  }
1147
  });
1148
  if (!response || response.status !== 'success' || !response.gloss) {
 
1151
  if (response.dropped && response.dropped.length > 0) {
1152
  console.info('[StableSigner] dropped non-renderable text tokens', response.dropped);
1153
  }
1154
+ return response;
1155
+ }
1156
+
1157
+ handleSemanticDemo() {
1158
+ if (this.semanticGlossCheckbox) {
1159
+ this.semanticGlossCheckbox.checked = true;
1160
+ }
1161
+ this.promptTextInput.value = 'Could you aid me? I bought a mobile and go home later.';
1162
+ this.handleTranslate(false).catch((error) => {
1163
+ this.updateStatus('Semantic demo failed');
1164
+ this.logsPre.textContent = (error && error.message) ? error.message : 'Semantic demo failed';
1165
+ });
1166
+ }
1167
+
1168
+ handleSelectionSemanticDemo() {
1169
+ if (this.selectionSemanticCheckbox) {
1170
+ this.selectionSemanticCheckbox.checked = true;
1171
+ }
1172
+ const demoText = 'Could you aid me? I bought a mobile and go home later.';
1173
+ if (this.selectionTemplate) {
1174
+ this.selectionTemplate.innerHTML = `<div class="stable-dialogue-turn question"><span class="stable-dialogue-speaker">User</span><p>${demoText}</p></div>`;
1175
+ }
1176
+ this.selectedTemplateText = demoText;
1177
+ this.lastGeneratedSelection = '';
1178
+ if (this.selectionTextEl) {
1179
+ this.selectionTextEl.textContent = demoText;
1180
+ }
1181
+ if (this.selectionPopover) {
1182
+ this.selectionPopover.style.display = 'block';
1183
+ }
1184
+ this.handleSelectedTextGenerate(true);
1185
  }
1186
 
1187
  handleTemplateSelection() {
 
1262
  throw new Error('WebSocket not ready');
1263
  }
1264
  const payload = this.collectPayload();
1265
+ const selectionUseSemantic = !!(this.selectionSemanticCheckbox && this.selectionSemanticCheckbox.checked);
1266
+ payload.translationMode = selectionUseSemantic ? 'semantic' : 'text2gloss';
1267
+ payload.useSemantic = selectionUseSemantic;
1268
  payload.maxNewTokens = Math.min(toInt(this.maxNewTokensInput.value, 64), 18);
1269
  payload.maxCandidates = Math.min(toInt(this.maxCandidatesInput.value, 10), 3);
1270
  const response = await window.controlWaveWebSocket.sendRequest(this.id, {
pipeline01_text2gloss.py CHANGED
@@ -108,12 +108,100 @@ PHRASE_MAP = {
108
  }
109
 
110
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
111
  @dataclass
112
  class TextToGlossResult:
113
  gloss: str
114
  tokens: List[str]
115
  matched: List[Tuple[str, str]]
116
  dropped: List[str]
 
117
 
118
 
119
  def load_available_glosses(path: Path | str = DEFAULT_POSE_DICT) -> Set[str]:
@@ -168,6 +256,41 @@ def _best_vocab_match(token: str, vocab: Set[str]) -> Optional[str]:
168
  return None
169
 
170
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
171
  def _apply_phrase_map(tokens: Sequence[str], vocab: Set[str]) -> Tuple[List[str], List[Tuple[str, str]]]:
172
  output: List[str] = []
173
  matched: List[Tuple[str, str]] = []
@@ -195,11 +318,16 @@ def translate_text_to_gloss(
195
  text: str,
196
  vocab: Optional[Iterable[str]] = None,
197
  max_tokens: int = 64,
 
198
  ) -> TextToGlossResult:
199
  """Translate English text into a renderable uppercase gloss sequence."""
200
  vocab_set = {item.upper() for item in vocab} if vocab is not None else load_available_glosses()
201
  raw_tokens = tokenize_english(text)
202
  phrase_tokens, phrase_matches = _apply_phrase_map(raw_tokens, vocab_set)
 
 
 
 
203
 
204
  gloss_tokens: List[str] = []
205
  matched: List[Tuple[str, str]] = list(phrase_matches)
@@ -212,6 +340,10 @@ def translate_text_to_gloss(
212
  dropped.append(token)
213
  continue
214
  gloss = _best_vocab_match(token, vocab_set)
 
 
 
 
215
  if gloss:
216
  gloss_tokens.append(gloss)
217
  if token != gloss:
@@ -224,6 +356,7 @@ def translate_text_to_gloss(
224
  tokens=gloss_tokens,
225
  matched=matched,
226
  dropped=dropped,
 
227
  )
228
 
229
 
@@ -232,11 +365,12 @@ def main() -> None:
232
  parser.add_argument("text", help="English text to translate.")
233
  parser.add_argument("--pose-dict", default=str(DEFAULT_POSE_DICT), help="WLASL-style pose dictionary JSON.")
234
  parser.add_argument("--max-tokens", type=int, default=64)
 
235
  parser.add_argument("--json", action="store_true", help="Print structured JSON instead of only gloss.")
236
  args = parser.parse_args()
237
 
238
  vocab = load_available_glosses(args.pose_dict)
239
- result = translate_text_to_gloss(args.text, vocab=vocab, max_tokens=args.max_tokens)
240
  if args.json:
241
  print(
242
  json.dumps(
@@ -245,6 +379,7 @@ def main() -> None:
245
  "tokens": result.tokens,
246
  "matched": result.matched,
247
  "dropped": result.dropped,
 
248
  },
249
  ensure_ascii=True,
250
  indent=2,
 
108
  }
109
 
110
 
111
+ SEMANTIC_PHRASE_MAP = {
112
+ "HELP ME": ["HELP", "I"],
113
+ "NEED HELP": ["NEED", "HELP"],
114
+ "I NEED HELP": ["I", "NEED", "HELP"],
115
+ "PLEASE HELP": ["PLEASE", "HELP"],
116
+ "GO TO SCHOOL": ["GO", "SCHOOL"],
117
+ "GO HOME": ["GO", "HOME"],
118
+ "COME HERE": ["COME", "HERE"],
119
+ "SEE YOU": ["SEE", "YOU"],
120
+ "NICE TO MEET YOU": ["NICE", "MEET", "YOU"],
121
+ "HOW ARE YOU": ["HOW", "YOU"],
122
+ }
123
+
124
+
125
+ SEMANTIC_TOKEN_MAP = {
126
+ "ASSIST": "HELP",
127
+ "AID": "HELP",
128
+ "SUPPORT": "HELP",
129
+ "SUPPORTING": "HELP",
130
+ "REQUIRE": "NEED",
131
+ "REQUIRES": "NEED",
132
+ "REQUIRED": "NEED",
133
+ "MUST": "NEED",
134
+ "SHOULD": "NEED",
135
+ "PURCHASE": "BUY",
136
+ "PURCHASED": "BUY",
137
+ "PURCHASING": "BUY",
138
+ "GET": "BUY",
139
+ "BOUGHT": "BUY",
140
+ "OBTAIN": "BUY",
141
+ "SPEAK": "TALK",
142
+ "SPEAKING": "TALK",
143
+ "TELL": "SAY",
144
+ "TOLD": "SAY",
145
+ "VIEW": "SEE",
146
+ "WATCH": "SEE",
147
+ "OBSERVE": "SEE",
148
+ "UNDERSTOOD": "UNDERSTAND",
149
+ "UNDERSTANDING": "UNDERSTAND",
150
+ "CHILDREN": "CHILD",
151
+ "KIDS": "CHILD",
152
+ "KID": "CHILD",
153
+ "ADULTS": "ADULT",
154
+ "PEOPLE": "PERSON",
155
+ "HUMAN": "PERSON",
156
+ "PERSONS": "PERSON",
157
+ "PHYSICIAN": "DOCTOR",
158
+ "MEDICAL": "DOCTOR",
159
+ "HOSPITAL": "DOCTOR",
160
+ "INSTRUCTOR": "TEACHER",
161
+ "EDUCATOR": "TEACHER",
162
+ "PUPIL": "STUDENT",
163
+ "LEARNER": "STUDENT",
164
+ "AUTOMOBILE": "CAR",
165
+ "VEHICLE": "CAR",
166
+ "CELL": "PHONE",
167
+ "MOBILE": "PHONE",
168
+ "TELEPHONE": "PHONE",
169
+ "RESIDENCE": "HOME",
170
+ "HOUSE": "HOME",
171
+ "WEATHER": "WEATHER",
172
+ "RAINING": "RAIN",
173
+ "SUNNY": "SUN",
174
+ "COLD": "COLD",
175
+ "HOT": "HOT",
176
+ "BEAUTIFUL": "NICE",
177
+ "GREAT": "GOOD",
178
+ "EXCELLENT": "GOOD",
179
+ "FINE": "GOOD",
180
+ "HAPPY": "HAPPY",
181
+ "SAD": "SAD",
182
+ "AFRAID": "FEAR",
183
+ "SCARED": "FEAR",
184
+ "SOON": "LATER",
185
+ "AFTER": "LATER",
186
+ "CURRENTLY": "NOW",
187
+ "IMMEDIATELY": "NOW",
188
+ "YESTERDAY": "YESTERDAY",
189
+ "TOMORROW": "TOMORROW",
190
+ }
191
+
192
+
193
+ AMBIGUOUS_SEMANTIC_BLOCKLIST = {
194
+ "CAN", "MAY", "MIGHT", "LEFT", "RIGHT", "LIGHT", "WELL", "KIND", "MEAN", "PART", "SECOND"
195
+ }
196
+
197
+
198
  @dataclass
199
  class TextToGlossResult:
200
  gloss: str
201
  tokens: List[str]
202
  matched: List[Tuple[str, str]]
203
  dropped: List[str]
204
+ semantic: List[Tuple[str, str]]
205
 
206
 
207
  def load_available_glosses(path: Path | str = DEFAULT_POSE_DICT) -> Set[str]:
 
256
  return None
257
 
258
 
259
+ def _best_semantic_match(token: str, vocab: Set[str]) -> Optional[str]:
260
+ if token in AMBIGUOUS_SEMANTIC_BLOCKLIST:
261
+ return None
262
+ candidates = _simple_lemma(token)
263
+ for candidate in candidates:
264
+ mapped = SEMANTIC_TOKEN_MAP.get(candidate)
265
+ if mapped and mapped in vocab:
266
+ return mapped
267
+ return None
268
+
269
+
270
+ def _apply_semantic_phrase_map(tokens: Sequence[str], vocab: Set[str]) -> Tuple[List[str], List[Tuple[str, str]]]:
271
+ output: List[str] = []
272
+ matched: List[Tuple[str, str]] = []
273
+ i = 0
274
+ phrases = sorted(SEMANTIC_PHRASE_MAP.items(), key=lambda item: len(item[0].split()), reverse=True)
275
+ while i < len(tokens):
276
+ consumed = False
277
+ for phrase, gloss_tokens in phrases:
278
+ phrase_tokens = phrase.split()
279
+ if list(tokens[i : i + len(phrase_tokens)]) != phrase_tokens:
280
+ continue
281
+ available = [gloss for gloss in gloss_tokens if gloss in vocab]
282
+ if available:
283
+ output.extend(available)
284
+ matched.append((phrase, " ".join(available)))
285
+ i += len(phrase_tokens)
286
+ consumed = True
287
+ break
288
+ if not consumed:
289
+ output.append(tokens[i])
290
+ i += 1
291
+ return output, matched
292
+
293
+
294
  def _apply_phrase_map(tokens: Sequence[str], vocab: Set[str]) -> Tuple[List[str], List[Tuple[str, str]]]:
295
  output: List[str] = []
296
  matched: List[Tuple[str, str]] = []
 
318
  text: str,
319
  vocab: Optional[Iterable[str]] = None,
320
  max_tokens: int = 64,
321
+ use_semantic: bool = False,
322
  ) -> TextToGlossResult:
323
  """Translate English text into a renderable uppercase gloss sequence."""
324
  vocab_set = {item.upper() for item in vocab} if vocab is not None else load_available_glosses()
325
  raw_tokens = tokenize_english(text)
326
  phrase_tokens, phrase_matches = _apply_phrase_map(raw_tokens, vocab_set)
327
+ semantic_matches: List[Tuple[str, str]] = []
328
+ if use_semantic:
329
+ phrase_tokens, semantic_phrase_matches = _apply_semantic_phrase_map(phrase_tokens, vocab_set)
330
+ semantic_matches.extend(semantic_phrase_matches)
331
 
332
  gloss_tokens: List[str] = []
333
  matched: List[Tuple[str, str]] = list(phrase_matches)
 
340
  dropped.append(token)
341
  continue
342
  gloss = _best_vocab_match(token, vocab_set)
343
+ if not gloss and use_semantic:
344
+ gloss = _best_semantic_match(token, vocab_set)
345
+ if gloss:
346
+ semantic_matches.append((token, gloss))
347
  if gloss:
348
  gloss_tokens.append(gloss)
349
  if token != gloss:
 
356
  tokens=gloss_tokens,
357
  matched=matched,
358
  dropped=dropped,
359
+ semantic=semantic_matches,
360
  )
361
 
362
 
 
365
  parser.add_argument("text", help="English text to translate.")
366
  parser.add_argument("--pose-dict", default=str(DEFAULT_POSE_DICT), help="WLASL-style pose dictionary JSON.")
367
  parser.add_argument("--max-tokens", type=int, default=64)
368
+ parser.add_argument("--semantic", action="store_true", help="Enable conservative semantic synonym fallback.")
369
  parser.add_argument("--json", action="store_true", help="Print structured JSON instead of only gloss.")
370
  args = parser.parse_args()
371
 
372
  vocab = load_available_glosses(args.pose_dict)
373
+ result = translate_text_to_gloss(args.text, vocab=vocab, max_tokens=args.max_tokens, use_semantic=args.semantic)
374
  if args.json:
375
  print(
376
  json.dumps(
 
379
  "tokens": result.tokens,
380
  "matched": result.matched,
381
  "dropped": result.dropped,
382
+ "semantic": result.semantic,
383
  },
384
  ensure_ascii=True,
385
  indent=2,
viser_backend.py CHANGED
@@ -213,17 +213,20 @@ class StableSignerBackend(PluginBackendBase):
213
  decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0].strip().upper()
214
  return decoded
215
 
216
- def _generate_constrained_text_gloss(self, text: str, max_tokens: int) -> Dict[str, Any]:
217
  result = translate_text_to_gloss(
218
  text,
219
  vocab=self._get_available_glosses(),
220
  max_tokens=max_tokens,
 
221
  )
222
  return {
223
  "gloss": result.gloss,
224
  "tokens": result.tokens,
225
  "matched": result.matched,
226
  "dropped": result.dropped,
 
 
227
  }
228
 
229
  def _build_output_path(self, gloss_tokens: List[str], draw_style: str) -> Path:
@@ -345,8 +348,10 @@ class StableSignerBackend(PluginBackendBase):
345
  if not text:
346
  return {"status": "error", "message": "Text is empty."}
347
  max_tokens = _safe_int(data.get("maxNewTokens"), 64)
 
 
348
  try:
349
- result = self._generate_constrained_text_gloss(text, max_tokens)
350
  if not result["gloss"]:
351
  return {
352
  "status": "error",
@@ -508,10 +513,11 @@ class StableSignerBackend(PluginBackendBase):
508
  self._generation_lock.release()
509
  return {"status": "error", "message": "Please enter Prompt/Text or a gloss sequence."}
510
  translation_mode = (data.get("translationMode") or "text2gloss").strip().lower()
 
511
  if translation_mode == "prompt2gloss":
512
  gloss_input = self._generate_gloss(text_input, prompt_input, max_new_tokens)
513
  else:
514
- gloss_input = self._generate_constrained_text_gloss(text_input, max_new_tokens)["gloss"]
515
 
516
  gloss_tokens = [token for token in gloss_input.upper().split() if token]
517
  if not gloss_tokens:
@@ -603,10 +609,11 @@ class StableSignerBackend(PluginBackendBase):
603
  if not text_input:
604
  return {"status": "error", "message": "Please enter Prompt/Text or a gloss sequence."}
605
  translation_mode = (data.get("translationMode") or "text2gloss").strip().lower()
 
606
  if translation_mode == "prompt2gloss":
607
  gloss_input = self._generate_gloss(text_input, prompt_input, max_new_tokens)
608
  else:
609
- gloss_result = self._generate_constrained_text_gloss(text_input, max_new_tokens)
610
  gloss_input = gloss_result["gloss"]
611
 
612
  gloss_tokens = [token for token in gloss_input.upper().split() if token]
 
213
  decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0].strip().upper()
214
  return decoded
215
 
216
+ def _generate_constrained_text_gloss(self, text: str, max_tokens: int, use_semantic: bool = False) -> Dict[str, Any]:
217
  result = translate_text_to_gloss(
218
  text,
219
  vocab=self._get_available_glosses(),
220
  max_tokens=max_tokens,
221
+ use_semantic=use_semantic,
222
  )
223
  return {
224
  "gloss": result.gloss,
225
  "tokens": result.tokens,
226
  "matched": result.matched,
227
  "dropped": result.dropped,
228
+ "semantic": result.semantic,
229
+ "mode": "semantic" if use_semantic else "rule",
230
  }
231
 
232
  def _build_output_path(self, gloss_tokens: List[str], draw_style: str) -> Path:
 
348
  if not text:
349
  return {"status": "error", "message": "Text is empty."}
350
  max_tokens = _safe_int(data.get("maxNewTokens"), 64)
351
+ mode = str(data.get("mode") or data.get("translationMode") or "rule").lower()
352
+ use_semantic = bool(data.get("useSemantic")) or mode in {"semantic", "semantic_text2gloss", "semantic-text2gloss"}
353
  try:
354
+ result = self._generate_constrained_text_gloss(text, max_tokens, use_semantic=use_semantic)
355
  if not result["gloss"]:
356
  return {
357
  "status": "error",
 
513
  self._generation_lock.release()
514
  return {"status": "error", "message": "Please enter Prompt/Text or a gloss sequence."}
515
  translation_mode = (data.get("translationMode") or "text2gloss").strip().lower()
516
+ use_semantic = bool(data.get("useSemantic")) or translation_mode in {"semantic", "semantic_text2gloss", "semantic-text2gloss"}
517
  if translation_mode == "prompt2gloss":
518
  gloss_input = self._generate_gloss(text_input, prompt_input, max_new_tokens)
519
  else:
520
+ gloss_input = self._generate_constrained_text_gloss(text_input, max_new_tokens, use_semantic=use_semantic)["gloss"]
521
 
522
  gloss_tokens = [token for token in gloss_input.upper().split() if token]
523
  if not gloss_tokens:
 
609
  if not text_input:
610
  return {"status": "error", "message": "Please enter Prompt/Text or a gloss sequence."}
611
  translation_mode = (data.get("translationMode") or "text2gloss").strip().lower()
612
+ use_semantic = bool(data.get("useSemantic")) or translation_mode in {"semantic", "semantic_text2gloss", "semantic-text2gloss"}
613
  if translation_mode == "prompt2gloss":
614
  gloss_input = self._generate_gloss(text_input, prompt_input, max_new_tokens)
615
  else:
616
+ gloss_result = self._generate_constrained_text_gloss(text_input, max_new_tokens, use_semantic=use_semantic)
617
  gloss_input = gloss_result["gloss"]
618
 
619
  gloss_tokens = [token for token in gloss_input.upper().split() if token]