[longzhenren]

Root cause: For dicts with "arabic" in the filename, BaseRecLabelDecode sets reverse=True and decode() reverses the decoded text with pred_reverse() before calling get_word_info(). pred_reverse keeps runs of LTR characters together and reverses the token sequence, but selection (the feature-column mask) was passed through unchanged. get_word_info then pairs each character of the reversed text with the ascending column list, so every word-level box is attributed to the wrong columns.

Reproduction (return_word_box=True, reversed dict, text "Hello مرحبا"):

character	reported column	actual column
first char of reversed text (ا)	0	10
last char (space)	never reported	5

Every column after the first LTR run shifts, so downstream word-level bbox alignment for RTL models is wrong end to end.

Fix: when reversing, reorder the selected columns with the same token mapping as pred_reverse (new helper reverse_selection returning an explicit column array) and pass that to get_word_info, which now accepts either the boolean mask or an explicit column array. Non-reversed decoding is untouched.

Test: tests/unit/test_rec_postprocess.py covers the mixed LTR/RTL case (columns follow the reversed reading order, including with interleaved blanks) and the unchanged LTR behavior. Run with pytest tests/unit/test_rec_postprocess.py (5 passed).

[CLAassistant]

All committers have signed the CLA.

[longzhenren]

I have read and agree to the Contributor License Agreement