Understanding _REGEX_REPLACE

[sladenheim]

Feature you are interested in and your specific question(s):
In the _clean_utterance() function there is a _REGEX_REPLACE tuple which consists of a sub-tuple

(re.compile(r"[^\[\./!]\?"), " ? ")

I noticed that this regex takes strings like "hello?" and replaces it with "hell ?".

What you are trying to accomplish with this feature or functionality:
I would expect the value to return as "hello ? ".

Additional context:
This is causing incorrect calculations for the type to token ratio.

Do you have any feedback on why this regex is used and if this is expected behavior?

[jacksonllee]

Hello! For datasets that come directly from the CHILDES / Talkbank ecosystem or those that strictly follow their CHAT data format, transcribed data in the form of "hello?" would not be expected, and instead would be "hello ?" with a space before the question mark. (This being said, I've recently become aware that some official CHILDES / Talkbank datasets might have these spaces missing after their semi-regular dataset updates, but this seems to occur only when there are non-ASCII characters around the spot where the space would be expected.) Where does your dataset come from?

[sladenheim]

Thanks for the response! I am working with .cha files coming from the CLAN software. The .cha files are transcripts of classroom video recordings. I'm not that knowledgeable about how CLAN works and whether it does or does not enforce the CHILDES/Talkbank CHAT format so that questions marks in sentences have preceding white space.
Is this the only punctuation that follows this convention (e.g., do periods work the same way?). This helps clarify that what is happening is the expected behavior. Do you think it is worthwhile to build in functionality to handle cases where this convention is violated? Although if it violates the convention then it might be better to just raise a warning.

[JohaCM]

Thank Jackson, for your response. We are using existing transcripts created in CLAN, but they are not part of CHILDES / Talkbank datasets.