batchalign 0.8.2.post9__tar.gz → 0.8.2.post11__tar.gz

{batchalign-0.8.2.post9/batchalign.egg-info → batchalign-0.8.2.post11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.8.2.post9
+Version: 0.8.2.post11
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

{batchalign-0.8.2.post9 → batchalign-0.8.2.post11}/batchalign/pipelines/analysis/compare.py

@@ -159,8 +159,10 @@ 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 most positional (in-order) 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
@@ -175,6 +177,7 @@ 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_pos_matches = -1
 
     for span in range(min_window, max_window + 1):
         window_counts = Counter(main_tokens[:span])
@@ -197,16 +200,28 @@ def _find_best_segment(gold_tokens, main_tokens, mfn):
             len_delta = abs(span - gold_len)
             end = start + span
 
+            # Count how many tokens match the gold in the same position
+            pos_matches = sum(
+                1 for k in range(min(span, gold_len))
+                if mfn(main_tokens[start + k], gold_tokens[k])
+            )
+
             if score > best_score:
                 best = (start, end)
                 best_score = score
                 best_len_delta = len_delta
+                best_pos_matches = pos_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]:
-                    best = (start, end)
+                    best_pos_matches = pos_matches
+                elif len_delta == best_len_delta:
+                    if pos_matches > best_pos_matches:
+                        best = (start, end)
+                        best_pos_matches = pos_matches
+                    elif pos_matches == best_pos_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.
@@ -216,6 +231,30 @@ def _find_best_segment(gold_tokens, main_tokens, mfn):
     return best
 
 
+def _best_rotation(window_tokens, gold_tokens, mfn):
+    """Find the cyclic rotation of *window_tokens* that maximises matches.
+
+    Returns the rotation offset *r* such that
+    ``window_tokens[r:] + window_tokens[:r]`` best aligns to *gold_tokens*.
+    """
+    if len(window_tokens) <= 1:
+        return 0
+
+    best_r = 0
+    best_matches = -1
+    n = len(window_tokens)
+
+    for r in range(n):
+        rotated = window_tokens[r:] + window_tokens[:r]
+        alignment = align(rotated, gold_tokens, False, mfn)
+        matches = sum(1 for item in alignment if isinstance(item, Match))
+        if matches > best_matches:
+            best_matches = matches
+            best_r = r
+
+    return best_r
+
+
 def _get_pos(form):
     """Extract uppercased POS from a Form's morphology, or '?' if absent."""
     if form is not None and form.morphology:
@@ -362,8 +401,13 @@ class CompareEngine(BatchalignEngine):
                     utt_main_forms[utt_idx].append(m_form)
                     utt_main_speakers[utt_idx].append(m_utt_idx)
 
-            # Align the chosen window against this gold utterance
+            # Align the chosen window against this gold utterance,
+            # trying cyclic rotations to avoid spurious del/ins pairs.
             window_main = conformed_main[abs_start:abs_end]
+            window_len = len(window_main)
+            rotation = _best_rotation(window_main, g_tokens, match_fn)
+            if rotation > 0:
+                window_main = window_main[rotation:] + window_main[:rotation]
             utt_alignment = align(window_main, g_tokens, False, match_fn)
 
             local_main_cursor = 0
@@ -372,7 +416,7 @@ class CompareEngine(BatchalignEngine):
 
             for item in utt_alignment:
                 if isinstance(item, Match):
-                    global_main_idx = abs_start + local_main_cursor
+                    global_main_idx = abs_start + (local_main_cursor + rotation) % window_len
                     orig_main_idx = main_map[global_main_idx]
                     main_form = main_info[orig_main_idx][2]
                     orig_gold_idx = g_maps[local_gold_cursor]
@@ -403,7 +447,7 @@ class CompareEngine(BatchalignEngine):
                         local_gold_cursor += 1
 
                     else:
-                        global_main_idx = abs_start + local_main_cursor
+                        global_main_idx = abs_start + (local_main_cursor + rotation) % window_len
                         orig_main_idx = main_map[global_main_idx]
                         main_form = main_info[orig_main_idx][2]
 
@@ -464,6 +508,10 @@ class CompareEngine(BatchalignEngine):
                 utt.time = (timed_forms[0].time[0], timed_forms[-1].time[1])
                 utt.text = None
 
+        # Copy @Media header from the main doc if it has one
+        if doc.media is not None:
+            gold.media = doc.media
+
         return gold
 
 

batchalign-0.8.2.post11/batchalign/version

@@ -0,0 +1,3 @@
+0.8.2-post.11
+April 08 2026
+compare beter

{batchalign-0.8.2.post9 → batchalign-0.8.2.post11/batchalign.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: batchalign
-Version: 0.8.2.post9
+Version: 0.8.2.post11
 Summary: Python Speech Language Sample Analysis
 Author: Brian MacWhinney, Houjun Liu
 Author-email: macw@cmu.edu, houjun@cmu.edu

batchalign-0.8.2.post9/batchalign/version

@@ -1,3 +0,0 @@
-0.8.2-post.9
-April 02 2026
-segment