Bug when attacking <SPLIT> in MNLI using CLARE

[JFChi]

Describe the bug

In MNLI training example 255115, the CLARE attack generate an attack result like this, which means it attacks the split token.

you still there Fernando sorry uh<SPLIT>I thought I lost you for a minute. you still there Fernando sorry uh<SPLOTS IT>I thought I lost you for a minute.

[JFChi]

The problem lies here: https://github.com/QData/TextAttack/blob/master/textattack/shared/attacked_text.py#L415,

Since "I" is in "<SPLIT>", so it misattack the SPLIT token. Replace this line with

# get word_start
if (input_word in AttackedText.SPLIT_TOKEN) and (AttackedText.SPLIT_TOKEN in original_text):
    split_start = original_text.index(AttackedText.SPLIT_TOKEN)
    split_end = split_start + len(AttackedText.SPLIT_TOKEN)
    if original_text.index(input_word) < split_end and \
        original_text.index(input_word) >= split_start:
        # hit corner cases when input_word is in AttackedText.SPLIT_TOKEN
        word_start = original_text.replace(AttackedText.SPLIT_TOKEN, "").index(input_word) + len(AttackedText.SPLIT_TOKEN)
    else:
        word_start = original_text.index(input_word)
else:
    word_start = original_text.index(input_word)

fix the issue.

[jxmorris12]

@JFChi thanks for the fix! Can you submit a pull request and tag me?

[VijayKalmath]

@JFChi , Can you please share link to a colab notebook where we can reproduce this issue ?