fix precision bug

chinesebleu.py

@@ -56,12 +56,12 @@ Examples:
     >>> my_new_module = evaluate.load("chinesebleu")
     >>> results = my_new_module.compute(references=["這裡就是香港都會大學"], predictions=["這裡是香港都會大學"])
     >>> print(results)
-    {'score': 71.89393375176813, 'counts': [9, 7, 5, 4], 'totals': [9, 8, 7, 6], 'bp': 1.0, tokenizer: 'chinese'}
+    {'score': 71.89393375176813, 'counts': [9, 7, 5, 4], 'totals': [9, 8, 7, 6], 'bp': 1.0, 'sys_len': 9, 'ref_len': 10, tokenizer: 'chinese'}
 """
 
 @evaluate.utils.file_utils.add_start_docstrings(_DESCRIPTION, _KWARGS_DESCRIPTION)
 class ChineseBLEU(evaluate.Metric):
-    """TODO: Chinese BLEU - a BLEU-based metric for Chinese sentences"""
+    """Chinese BLEU - a BLEU-based metric for Chinese sentences"""
 
     def _info(self):
         return evaluate.MetricInfo(
@@ -134,19 +134,22 @@ class ChineseBLEU(evaluate.Metric):
             raise ValueError("Predictions and references must have the same length.")
 
         max_n = 4 # Default n-gram = 4
+
+        counts = [0] * max_n
+        totals = [0] * max_n
+        
         tokenizer = 'chinese'
         
         pred_tokens = [self._tokenize_chinese(p, tokenizer) for p in predictions]
         ref_tokens = [self._tokenize_chinese(r, tokenizer) for r in references]
 
-        # For total number of tokens < 4, fallback to character-level tokenizations
+        # For total number of tokens < 4, fallback to SacreBLEU
         if len(pred_tokens[0]) < 4 or len(ref_tokens[0]) < 4:
             tokenizer = 'char'
-            pred_tokens = [self._tokenize_chinese(p, tokenizer) for p in predictions]
-            ref_tokens = [self._tokenize_chinese(r, tokenizer) for r in references]
-    
-        counts = [0] * max_n
-        totals = [0] * max_n
+            sacrebleu = evaluate.load('sacrebleu')
+            bleu_result = sacrebleu.compute(predictions=predictions, references=references, tokenize="zh")
+            bleu_result['tokenizer'] = tokenizer
+            return bleu_result
         
         for n in range(1, max_n + 1):
             clipped_counts_n = 0
@@ -168,15 +171,11 @@ class ChineseBLEU(evaluate.Metric):
         
         # Compute precisions
         precisions = []
-        for i, (c, t) in enumerate(zip(counts, totals)):
+        for c, t in zip(counts, totals):
             if t == 0:
                 precisions.append(0.0)
             else:
-                # SacreBLEU floor smoothing for n=4 on short sentences (no matches, total=1)
-                p = float(c) / t
-                if i == 3 and c == 0 and t == 1: # i=3 is 4-gram (0-indexed)
-                    p = 1.0 / 2.0  # 0.5, matching SacreBLEU's heuristic
-                precisions.append(p)
+                precisions.append(float(c) / t)
         
         # Geometric mean of precisions
         if any(p == 0 for p in precisions):