Xapian devel - Jan 2024 - Possible bug using FLAG_WORD_BREAKS with fullwidth Unicode codepoints

I think I found a bug in Xapian 1.5 when using FLAG_WORD_BREAKS for input that
contains characters in Unicode Halfwidth and Fullwidth Forms
(https://unicode.org/charts/PDF/UFF00.pdf).

Since I am undecided yet if and how to fix this in Xapian I haven't come up
with a pull request. Because trac currently is offline, I could not file a bug.
I hope it's OK to post my analysis here first, I'll be happy to follow
up reporting that bug proper later (should we conclude that it actually is a
bug).

Imagine indexing the following Japanese text "??????????????" which in
English denotes the "Mitsubishi UFJ Factors Limited" bank.

Using word segmentation in Xapian 1.5 this causes the following terms to get
indexed:
?????
??
????
???

Note that last term, which starts with FULLWIDTH LATIN CAPITAL LETTER U'
(U+FF35). Xapian's Unicode library correctly assigns this the
UPPERCASE_LETTER category and indexes this verbatim.

However, querying for ??? produces the query Query(???@1). That is, it queries
for the lowercase form which seems to be the result of unconditional
lower-casing at
https://github.com/xapian/xapian/blob/master/xapian-core/queryparser/queryparser.lemony#L1459.
As a result, the query returns no result.

I have written code that demonstrates this at
https://gist.github.com/rsto/168a61536793e10a0a07c3920977e5eb

Now, I think that much of this issue can be prevented by normalizing both
indexed text and queries before passing them over the Xapian, but this requires
to rewrite indexes so isn't necessarily a quick fix. As a workaround, I
chose to detect such queries and query for both the lower-cased and original
uppercase forms in our systems.

Still, I do think it is a bug for Xapian not to return a result when querying
for a term that's verbatim in the original input and the database. Should
you agree I will be happy to discuss how to fix this and might come up with a
pull request once we agreed on a solution.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.xapian.org/pipermail/xapian-devel/attachments/20240104/19872ad8/attachment.htm>

On Thu, Jan 04, 2024 at 05:50:22PM +0100, Robert Stepanek
wrote:
> Since I am undecided yet if and how to fix this in Xapian I haven't
> come up with a pull request. Because trac currently is offline, I
> could not file a bug. I hope it's OK to post my analysis here first,
> I'll be happy to follow up reporting that bug proper later (should we
> conclude that it actually is a bug).
I've restarted trac.
> Still, I do think it is a bug for Xapian not to return a result when
> querying for a term that's verbatim in the original input and the
> database. Should you agree I will be happy to discuss how to fix this
> and might come up with a pull request once we agreed on a solution.
I think the cause is that the list of ranges of characters which need
word breaks findings includes this block of full- and half-width latin
forms, coupled with an assumption that there's no lowercase vs uppercase
forms in these alphabets.

Assuming the latter is valid, just removing this block (or removing the
parts of it which are Lu or Ll) should fix the problem as then
tokenisation will switch mode - I tried this and it fixes your case at
least:

diff --git a/xapian-core/queryparser/word-breaker.cc
b/xapian-core/queryparser/word-breaker.cc
index 8108523ccd53..4fabc23f4b56 100644
--- a/xapian-core/queryparser/word-breaker.cc
+++ b/xapian-core/queryparser/word-breaker.cc
@@ -103,7 +103,7 @@ is_unbroken_script(unsigned p)
 	// FE30..FE4F; CJK Compatibility Forms
 	0xFE30 - 1, 0xFE4F,
 	// FF00..FFEF; Halfwidth and Fullwidth Forms
-	0xFF00 - 1, 0xFFEF,
+	//0xFF00 - 1, 0xFFEF,
 	// 1AFF0..1AFFF; Kana Extended-B
 	// 1B000..1B0FF; Kana Supplement
 	// 1B100..1B12F; Kana Extended-A

If we're fixing it this way we should check this list for other
instances of this (and doing this would probably reveal if that
assumption is valid).

Cheers,
    Olly