Merge pull request #107 from meedan/CV2-4909-Identify-opportunities-t…

…o-improve-YAKE-keywords

adding auto langdetect and cleaning up keywords extracted by yake

lib/model/yake_keywords.py

@@ -7,8 +7,35 @@
 from lib import schemas
 
 import yake
+import cld3
 
 class Model(Model):
+
+    def keep_largest_overlapped_keywords(self, keywords):
+        cleaned_keywords = []
+        for i in range(len(keywords)):
+            keep_keyword = True
+            for j in range(len(keywords)):
+                current_keyword = keywords[i][0]
+                other_keyword = keywords[j][0]
+                if len(other_keyword) > len(current_keyword):
+                    if other_keyword.find(current_keyword + " ") >= 0 or other_keyword.find(" " + current_keyword) >= 0:
+                        keep_keyword = False
+                        break
+            if keep_keyword:
+                cleaned_keywords.append(keywords[i])
+        return cleaned_keywords
+
+    def normalize_special_characters(self, text):
+        replacement = {"`": "'",
+                       "‘": "'",
+                       "’": "'",
+                       "“": "\"",
+                       "”": "\""}
+        for k, v in replacement.items():
+            text = text.replace(k, v)
+        return text
+
     def run_yake(self, text: str,
                  language: str,
                  max_ngram_size: int,
@@ -26,15 +53,25 @@ def run_yake(self, text: str,
         :param num_of_keywords: int
         :returns: str
         """
+        ### if language is set to "auto", auto-detect it.
+        if language == 'auto':
+            language = cld3.get_language(text).language
+        ### normalize special characters
+        text = self.normalize_special_characters(text)
+        ### extract keywords
         custom_kw_extractor = yake.KeywordExtractor(lan=language, n=max_ngram_size, dedupLim=deduplication_threshold,
                                                     dedupFunc=deduplication_algo, windowsSize=window_size,
                                                     top=num_of_keywords, features=None)
-        return {"keywords": custom_kw_extractor.extract_keywords(text)}
+
+        ### Keep the longest keyword of if there is an overlap between two keywords.
+        keywords = custom_kw_extractor.extract_keywords(text)
+        keywords = self.keep_largest_overlapped_keywords(keywords)
+        return {"keywords": keywords}
 
     def get_params(self, message: schemas.Message) -> dict:
         params = {
             "text": message.body.text,
-            "language": message.body.parameters.get("language", "en"),
+            "language": message.body.parameters.get("language", "auto"),
             "max_ngram_size": message.body.parameters.get("max_ngram_size", 3),
             "deduplication_threshold": message.body.parameters.get("deduplication_threshold", 0.25),
             "deduplication_algo": message.body.parameters.get("deduplication_algo", 'seqm'),

requirements.txt

@@ -24,3 +24,4 @@ numpy==1.26.4
 protobuf==3.20.2
 openai==1.35.6
 anthropic==0.31.1
+pycld3==0.22

test/lib/model/test_yake_keywords.py

@@ -43,6 +43,16 @@ def test_run_yake_real(self):
         results = self.yake_model.run_yake(**self.yake_model.get_params(message))
         self.assertEqual(results, {"keywords": [('love Meedan', 0.0013670273525686505)]})
 
+    def test_keep_largest_overlapped_keywords(self):
+        keywords_test = [('Alegre', 0),('Alegre', 0),('Timpani', 0), ('Presto Timpani', 0), ('AlegreAlegre', 0), ('Alegre Alegre', 0), ("Presto", 0)]
+        expected = [('Presto Timpani', 0), ('AlegreAlegre', 0), ('Alegre Alegre', 0)]
+        self.assertEqual(self.yake_model.keep_largest_overlapped_keywords(keywords_test), expected)
+
+    def test_normalize_special_characters(self):
+        text = "`‘’“”"
+        expected = "'''\"\""
+        self.assertEqual(self.yake_model.normalize_special_characters(text), expected)
+
     def test_get_params_with_defaults(self):
         message = schemas.parse_message({
             "body": {
@@ -51,7 +61,7 @@ def test_get_params_with_defaults(self):
             },
             "model_name": "yake_keywords__Model"
         })
-        expected = {'text': 'Some Text', 'language': "en", 'max_ngram_size': 3, 'deduplication_threshold': 0.25, 'deduplication_algo': 'seqm', 'window_size': 0, 'num_of_keywords': 10}
+        expected = {'text': 'Some Text', 'language': "auto", 'max_ngram_size': 3, 'deduplication_threshold': 0.25, 'deduplication_algo': 'seqm', 'window_size': 0, 'num_of_keywords': 10}
         self.assertEqual(self.yake_model.get_params(message), expected)
 
     def test_get_params_with_specifics(self):