Problem in bunsetu spanning

When I parsed a sentence “彼によって行われる” with GiNZA v4.0.1, I got the following result:

# text = 彼によって行われる
1	彼	彼	PRON	代名詞	_	5	obl	_	SpaceAfter=No|BunsetuBILabel=B|BunsetuPositionType=SEM_HEAD|NP_B|Reading=カレ
2	に	に	ADP	助詞-格助詞	_	1	case	_	SpaceAfter=No|BunsetuBILabel=I|BunsetuPositionType=SYN_HEAD|Reading=ニ
3	よっ	よる	VERB	動詞-一般	_	2	fixed	_	SpaceAfter=No|BunsetuBILabel=B|BunsetuPositionType=CONT|Inf=五段-ラ行,連用形-促音便|Reading=ヨッ
4	て	て	SCONJ	助詞-接続助詞	_	2	fixed	_	SpaceAfter=No|BunsetuBILabel=I|BunsetuPositionType=SYN_HEAD|Reading=テ
5	行わ	行う	VERB	動詞-一般	_	0	root	_	SpaceAfter=No|BunsetuBILabel=I|BunsetuPositionType=ROOT|Inf=五段-ワア行,未然形-一般|Reading=オコナワ
6	れる	れる	AUX	助動詞	_	5	aux	_	SpaceAfter=No|BunsetuBILabel=I|BunsetuPositionType=SYN_HEAD|Inf=助動詞-レル,終止形-一般|Reading=レル

This result indicates that there are two bunsetu spans in the sentence, “彼に” and “よって行われる”, which should be “彼によって” and “行われる”.

The problem seems to occur because the bunsetu spanning process in BunsetuRecognizer.__call__() checks only right children for each bunsetu head token, not all right descendants iteratively.

Here I attach a patch I believe to fix it.

--- ginza/bunsetu_recognizer.py~	2020-08-31 22:02:47.000000000 +0900
+++ ginza/bunsetu_recognizer.py	2020-09-03 11:08:32.000000000 +0900
@@ -193,13 +193,20 @@
             t = doc[head_i]
             if next_start < len(bunsetu_bi):
                 bunsetu_bi[next_start] = "B"
-            right = t
-            for sub in t.rights:
-                if heads[sub.i]:
-                    right = right.right_edge
-                    break
-                right = sub
-            next_start = right.i + 1
+            rights = [t]
+            while len(rights):
+                right = rights.pop(0)
+                for sub in right.rights:
+                    if heads[sub.i]:
+                        next_start = right.right_edge.i + 1
+                        break
+                    else:
+                        rights.append(sub)
+                else:
+                    continue
+                break
+            else:
+                next_start = t.right_edge.i + 1
 
         doc.user_data["bunsetu_heads"] = bunsetu_heads
         doc.user_data["bunsetu_bi_labels"] = bunsetu_bi

[tadashikumano]

As far as examining your attached diff, the revised code seems to work much appropriate than the previous revision. Thanks a lot. I’ll report you if I find any other problem.

[hiroshi-matsuda-rit]

@tadashikumano Thank you for reporting. I’ve ever implemented above recursive algorithm before releasing v4.0.0, but found some side-effects. I’d like to reevaluate this approach and try to improve total accuracy.