batchalign 0.8.2.post10__tar.gz → 0.8.2.post12__tar.gz

{batchalign-0.8.2.post10/batchalign.egg-info → batchalign-0.8.2.post12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.8.2.post10
+Version: 0.8.2.post12
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

{batchalign-0.8.2.post10 → batchalign-0.8.2.post12}/batchalign/cli/dispatch.py

@@ -226,6 +226,7 @@ def _run_pipeline_for_file(command, pipeline, file, output, loader_info, writer_
 
         # Write annotated CHAT
         CHATFile(doc=result["doc"]).write(output,
+                                          write_mor=False,
                                           merge_abbrev=local_kwargs.get("merge_abbrev", False))
 
         # Write metrics as JSON sidecar for consolidated CSV later

{batchalign-0.8.2.post10 → batchalign-0.8.2.post12}/batchalign/formats/chat/file.py

@@ -93,7 +93,7 @@ class CHATFile(BaseFormat):
             self.__doc = doc
 
 
-    def write(self, path, write_wor=True, merge_abbrev=False):
+    def write(self, path, write_wor=True, write_mor=True, merge_abbrev=False):
         """Write the CHATFile to file.
 
         Parameters
@@ -102,13 +102,13 @@ class CHATFile(BaseFormat):
             Path of where the CHAT file should get str.
         """
 
-        str_doc = self.__generate(self.__doc, self.__special_mor, write_wor=write_wor, merge_abbrev=merge_abbrev)
+        str_doc = self.__generate(self.__doc, self.__special_mor, write_wor=write_wor, write_mor=write_mor, merge_abbrev=merge_abbrev)
 
         with open(path, 'w', encoding="utf-8") as df:
             df.write(str_doc)
 
     @staticmethod
-    def __generate(doc:Document, special=False, write_wor=True, merge_abbrev=False):
+    def __generate(doc:Document, special=False, write_wor=True, write_mor=True, merge_abbrev=False):
         utterances = doc.content
 
         def __get_birthdays(line):
@@ -129,6 +129,7 @@ class CHATFile(BaseFormat):
             else:
                 main.append(generate_chat_utterance(i, special and doc.langs[0] == "eng",
                                                     write_wor=write_wor,
+                                                    write_mor=write_mor,
                                                     merge_abbrev=merge_abbrev))
         main.append("@End\n")
 

{batchalign-0.8.2.post10 → batchalign-0.8.2.post12}/batchalign/formats/chat/generator.py

@@ -11,7 +11,7 @@ import warnings
 # document[3].text = None
 # document[3].model_dump()
 
-def generate_chat_utterance(utterance: Utterance, special_mor=False, write_wor=True, merge_abbrev=False):
+def generate_chat_utterance(utterance: Utterance, special_mor=False, write_wor=True, write_mor=True, merge_abbrev=False):
     """Converts at Utterance to a CHAT string.
 
     Parameters
@@ -89,7 +89,7 @@ def generate_chat_utterance(utterance: Utterance, special_mor=False, write_wor=T
             # I'm sorry. This is just mor line generation; there is actually not that much complexity here
             mor_elems.append("~".join(f"{m.pos}|{m.lemma}{'-' if any([m.feats.startswith(i) for i in UD__GENDERS]) else ('-' if m.feats else '')}{m.feats}"
                                     for m in mor))
-    if len(mor_elems) > 0:
+    if write_mor and len(mor_elems) > 0:
         # if the end is punct, drop the tag
         if mor_elems[-1].startswith("PUNCT"):
             mor_elems[-1] = mor_elems[-1].split("|")[1]
@@ -97,7 +97,7 @@ def generate_chat_utterance(utterance: Utterance, special_mor=False, write_wor=T
 
     #### GRA LINE GENERATION ####
     # gra list is not different for MWT tokens so we flatten it
-    if len(gras) > 0:
+    if write_mor and len(gras) > 0:
         flat_gras = [i for j in gras if j for i in j]
         if flat_gras:
             result.append(f"%{'u' if special_mor else ''}gra:\t"+" ".join([f"{i.id}|{i.dep_id}|{i.dep_type}" for i in flat_gras]))

{batchalign-0.8.2.post10 → batchalign-0.8.2.post12}/batchalign/pipelines/analysis/compare.py

@@ -159,8 +159,11 @@ def _find_best_segment(gold_tokens, main_tokens, mfn):
     multiset overlap with the gold utterance, ignoring order. To keep common
     words from swallowing later transcript material, it only considers windows
     near the gold utterance length. Among equally good windows it prefers the
-    latest one, not the earliest. The caller then runs the full Levenshtein
-    aligner inside that window to produce token annotations.
+    one with the fewest non-matching (wasted) tokens, then the most
+    Levenshtein alignment matches, then the latest one as a final tiebreaker
+    (so CHI repetitions / self-corrections beat earlier INV tokens). The caller then
+    runs the full Levenshtein aligner inside that window to produce token
+    annotations.
     """
     if not gold_tokens or not main_tokens:
         return 0, 0
@@ -174,7 +177,8 @@ def _find_best_segment(gold_tokens, main_tokens, mfn):
 
     best = (0, min(main_len, gold_len))
     best_score = -1.0
-    best_len_delta = None
+    best_waste = None
+    best_align_matches = -1
 
     for span in range(min_window, max_window + 1):
         window_counts = Counter(main_tokens[:span])
@@ -194,19 +198,30 @@ def _find_best_segment(gold_tokens, main_tokens, mfn):
                 overlap += min(window_counts[right], gold_counts[right])
 
             score = overlap / gold_len
-            len_delta = abs(span - gold_len)
+            waste = span - overlap  # non-matching tokens in the window
             end = start + span
 
+            # Alignment-based match count (Levenshtein) as tiebreaker
+            window = main_tokens[start:end]
+            alignment = align(window, gold_tokens, False, mfn)
+            align_matches = sum(1 for item in alignment if isinstance(item, Match))
+
             if score > best_score:
                 best = (start, end)
                 best_score = score
-                best_len_delta = len_delta
+                best_waste = waste
+                best_align_matches = align_matches
             elif score == best_score:
-                if best_len_delta is None or len_delta < best_len_delta:
-                    best = (start, end)
-                    best_len_delta = len_delta
-                elif len_delta == best_len_delta and end > best[1]:
+                if best_waste is None or waste < best_waste:
                     best = (start, end)
+                    best_waste = waste
+                    best_align_matches = align_matches
+                elif waste == best_waste:
+                    if align_matches > best_align_matches:
+                        best = (start, end)
+                        best_align_matches = align_matches
+                    elif align_matches == best_align_matches and end > best[1]:
+                        best = (start, end)
 
     # If no tokens overlap at all, return an empty window so the caller
     # doesn't consume main tokens that belong to a later gold utterance.

batchalign-0.8.2.post12/batchalign/version

@@ -0,0 +1,3 @@
+0.8.2-post.12
+April 15 2026
+compare beter

{batchalign-0.8.2.post10 → batchalign-0.8.2.post12/batchalign.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.8.2.post10
+Version: 0.8.2.post12
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

batchalign-0.8.2.post10/batchalign/version

@@ -1,3 +0,0 @@
-0.8.2-post.10
-April 02 2026
-segment