go back to sacrebleu for in-lines computation

logmetric.py

@@ -19,6 +19,7 @@ import re
 import dateutil.parser
 import numpy as np
 from difflib import SequenceMatcher
+import sacrebleu
 
 import time
 
@@ -70,7 +71,8 @@ class LogMetric(evaluate.Metric):
     # Constant regex to get timestrings
     timestamp_regex = r'^\s*\[?\s*(\d{4}[-/.]\d{2}[-/.]\d{2}(?:[ T]\d{2}[:]\d{2}(?:[:]\d{2}(?:[.,]\d+)?)?(?:Z|[+-]\d{2}[:]\d{2})?)?)\s*\]?\s*'
     timestamp_pattern = re.compile(timestamp_regex, re.MULTILINE)
-    
+    sacrebleu_metric = evaluate.load("evaluate-metric/sacrebleu")
+
 
     def _info(self):
         # TODO: Specifies the evaluate.EvaluationModuleInfo object
@@ -113,13 +115,15 @@ class LogMetric(evaluate.Metric):
 
     # Use minimum edit distance between two sentences
     def get_overall_similarity(self, sentence1, sentence2):
-        s1split = sentence1.split()
-        s2split = sentence2.split()
+        return self.sacrebleu_metric.compute(predictions=sentence1, references=sentence2)
+        
+        # s1split = sentence1.split()
+        # s2split = sentence2.split()
 
-        jaccard_score = self.get_jaccard_similarity(set(s1split), set(s2split))
-        length_score = self.get_length_score(s1split, s2split)
+        # jaccard_score = self.get_jaccard_similarity(set(s1split), set(s2split))
+        # length_score = self.get_length_score(s1split, s2split)
 
-        return (jaccard_score * 0.7 + length_score * 0.3) * 100.0
+        # return (jaccard_score * 0.7 + length_score * 0.3) * 100.0
 
     def getLogMetric(self, pred : str, ref : str):
         ref = ref.strip(' \t\n\r')
@@ -194,11 +198,10 @@ class LogMetric(evaluate.Metric):
                 monotonicallyIncreasingScore = 0.0
     
         # apply jaccard-similarity to every pred-ref pair and then take mean score * 100
-        local_score = np.mean([self.get_overall_similarity(p, r) for p,r in 
-                                       zip(
+        local_score = self.get_overall_similarity(
                                         list(map(lambda t: t[1], pred_logentries))[:min_logentries], 
                                         list(map(lambda t: t[1], ref_logentries))[:min_logentries]
-                                        )])
+                                        )
 
 
         # we aggregate the bleu scores where we weight the difference in logentries with a score of 0