llm-ie 0.3.2__py3-none-any.whl → 0.3.4__py3-none-any.whl

llm_ie/extractors.py

@@ -143,7 +143,7 @@ class FrameExtractor(Extractor):
         self.tokenizer = RegexpTokenizer(r'\w+|[^\w\s]')
 
 
-    def _jaccard_score(self, s1:set, s2:set) -> float:
+    def _jaccard_score(self, s1:Set[str], s2:Set[str]) -> float:
         """
         This method calculates the Jaccard score between two sets of word tokens.
         """
@@ -166,6 +166,8 @@ class FrameExtractor(Extractor):
     def _get_closest_substring(self, text:str, pattern:str, buffer_size:float=0.2) -> Tuple[Tuple[int, int], float]:
         """
         This method finds the closest (highest Jaccard score) substring in text that matches the pattern.
+        the substring must start with the same word token as the pattern. This is due to the observation that 
+        LLM often generate the first few words consistently. 
 
         Parameters
         ----------
@@ -179,25 +181,29 @@ class FrameExtractor(Extractor):
         Returns : Tuple[Tuple[int, int], float]
             a tuple of 2-tuple span and Jaccard score.
         """
+        if not text or not pattern:
+            return None, 0
+
         text_tokens, text_spans = self._get_word_tokens(text)
         pattern_tokens, _ = self._get_word_tokens(pattern)
-        pattern_tokens = set(pattern_tokens)
+        pattern_tokens_set = set(pattern_tokens)
         window_size = len(pattern_tokens)
-        window_size_min = int(window_size * (1 - buffer_size))
-        window_size_max = int(window_size * (1 + buffer_size))
-        closest_substring_spans = None
+        window_size_min = max(1, int(window_size * (1 - buffer_size)))
+        window_size_max = int(window_size * (1 + buffer_size)) + 1
+        closest_substring_span = None
         best_score = 0
         
         for i in range(len(text_tokens) - window_size_max):
             for w in range(window_size_min, window_size_max): 
-                sub_str_tokens = set(text_tokens[i:i + w])
-                score = self._jaccard_score(sub_str_tokens, pattern_tokens)
-                if score > best_score:
-                    best_score = score
-                    sub_string_word_spans = text_spans[i:i + w]
-                    closest_substring_spans = (sub_string_word_spans[0][0], sub_string_word_spans[-1][-1])
+                sub_str_tokens = text_tokens[i:i + w]
+                if len(sub_str_tokens) > 0 and sub_str_tokens[0] == pattern_tokens[0]:
+                    score = self._jaccard_score(set(sub_str_tokens), pattern_tokens_set)
+                    if score > best_score:
+                        best_score = score
+                        sub_string_word_spans = text_spans[i:i + w]
+                        closest_substring_span = (sub_string_word_spans[0][0], sub_string_word_spans[-1][-1])
 
-        return closest_substring_spans, best_score
+        return closest_substring_span, best_score
 
 
     def _find_entity_spans(self, text: str, entities: List[str], case_sensitive:bool=False, 
@@ -235,7 +241,7 @@ class FrameExtractor(Extractor):
 
             # Exact match  
             match = re.search(re.escape(entity), text)
-            if match:
+            if match and entity:
                 start, end = match.span()
                 entity_spans.append((start, end))
                 # Replace the found entity with spaces to avoid finding the same instance again
@@ -243,7 +249,7 @@ class FrameExtractor(Extractor):
             # Fuzzy match
             elif fuzzy_match:
                 closest_substring_span, best_score = self._get_closest_substring(text, entity, buffer_size=fuzzy_buffer_size)
-                if best_score >= fuzzy_score_cutoff:
+                if closest_substring_span and best_score >= fuzzy_score_cutoff:
                     entity_spans.append(closest_substring_span)
                     # Replace the found entity with spaces to avoid finding the same instance again
                     text = text[:closest_substring_span[0]] + ' ' * (closest_substring_span[1] - closest_substring_span[0]) + text[closest_substring_span[1]:]
@@ -432,7 +438,7 @@ class BasicFrameExtractor(FrameExtractor):
                 frame = LLMInformationExtractionFrame(frame_id=f"{i}", 
                             start=start,
                             end=end,
-                            entity_text=ent[entity_key],
+                            entity_text=text[start:end],
                             attr={k: v for k, v in ent.items() if k != entity_key and v != ""})
                 frame_list.append(frame)
         return frame_list
@@ -731,12 +737,13 @@ class SentenceFrameExtractor(FrameExtractor):
             for ent, span in zip(entity_json, spans):
                 if span is not None:
                     start, end = span
+                    entity_text = sent['sentence_text'][start:end]
                     start += sent['sentence_start']
                     end += sent['sentence_start']
                     frame = LLMInformationExtractionFrame(frame_id=f"{len(frame_list)}", 
                                 start=start,
                                 end=end,
-                                entity_text=ent[entity_key],
+                                entity_text=entity_text,
                                 attr={k: v for k, v in ent.items() if k != entity_key and v != ""})
                     frame_list.append(frame)
         return frame_list

{llm_ie-0.3.2.dist-info → llm_ie-0.3.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: llm-ie
-Version: 0.3.2
+Version: 0.3.4
 Summary: An LLM-powered tool that transforms everyday language into robust information extraction pipelines.
 License: MIT
 Author: Enshuo (David) Hsu

{llm_ie-0.3.2.dist-info → llm_ie-0.3.4.dist-info}/RECORD

@@ -16,8 +16,8 @@ llm_ie/asset/prompt_guide/SentenceFrameExtractor_prompt_guide.txt,sha256=oKH_QeD
 llm_ie/asset/prompt_guide/SentenceReviewFrameExtractor_prompt_guide.txt,sha256=oKH_QeDgpw771ZdHk3L7DYz2Jvfm7OolUoTiJyMJI30,9541
 llm_ie/data_types.py,sha256=hPz3WOeAzfn2QKmb0CxHmRdQWZQ4G9zq8U-RJBVFdYk,14329
 llm_ie/engines.py,sha256=PTYs7s_iCPmI-yFUCVCPY_cMGS77ma2VGoz4rdNkODI,9308
-llm_ie/extractors.py,sha256=bZQh_qZP1yIwNLXIx1ZzVGN702q3qzrlwiTcHuMsJt0,64051
+llm_ie/extractors.py,sha256=j9L9USybJBmYZM4RAjGO6DR4StYBgzhqCN6nEFZZQVQ,64523
 llm_ie/prompt_editor.py,sha256=Xc5ZHsEnM8-YYITokIsM6BVsf2Ec_8ajJDaldPf-P8U,8577
-llm_ie-0.3.2.dist-info/METADATA,sha256=Lpzaq0n-kpfiIpzic3qUQncVDoCpfw6CGmxc5NftXro,41266
-llm_ie-0.3.2.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
-llm_ie-0.3.2.dist-info/RECORD,,
+llm_ie-0.3.4.dist-info/METADATA,sha256=-5Tf9TCWczCVOsdMavkBZ-KnYPGnbNrV1rsU-pMHfPA,41266
+llm_ie-0.3.4.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
+llm_ie-0.3.4.dist-info/RECORD,,