WikiCorpus: select which texts fields will be tokenized: only the body, or only the title, or title and body.

[fabriciorsf]

Currently, in WikiCorpus the article title is not being tokenized, and this field is the main text about the articles from WikiNews.
So, one proposal would be to adjust it to allow the choice of which text field to tokenize, with at least 3 options (and it can be flexible):

1. only the body (current and default), or
2. only the title, or
3. title and body.

In the case of the last option (title and body), the text to be tokenized must first be composed of the title, followed by a separate (full stop) and, finally, the body.

[gojomo]

I'm not fully clear what you mean. Would your (2) "only the title" option not include the body text at all? Does (3) in practice mean just a tokenized title (& a newline or 2 as the "full stop" you mention) prepended to the body?

A compact PR with the new capability is usually the clearest way to show intent & impact. If it's of sufficient simplicity (clear, matches existing style, low risk of future maintenance costs), & of potentially wider use, it could be integrated in a future release.

OTOH, if it's something people can easily do for themselves with a few lines of wrapping code over what is already returned, such functionality might be better kept as an example usage or outside recipe.

As a more-general note: it's been a few years since I was feeding Wikipedia text to Gensim, but when I did, I was sufficiently frustrated with the limits of WikiCorpus (& general difficulties of starting with idiosyncratic wiki-syntax text) that I used the Kiwix 'zim' format dumps instead - which have already rendered the wikitext into HTML, which in my experience at the time…

• was straightforward to strip to plaintext blocks, or pull useful subheadings/etc, using something like beautifulsoup
• retained some valuable tokens that the crude WikiCorpus processing lost/mangled

[fabriciorsf]

To illustrate,, see the example below:

<page>
<title>Alves to leave FC Barcelona</title>
....
<revision>
....
<text ...>{{date|June 4, 2016}}
{{infobox|FC Barcelona}}
On Thursday, [[FC Barcelona]] technical secretary {{w|Roberto Fernández Bonillo|Robert Fernández}} announced [[Brazil]]lian [[football (soccer)|football]] defender [[Dani Alves]] would leave the club this summer as a free agent.
[[File:2015 UEFA Super Cup 107.jpg|left|thumb|File photo of Dani Alves{{image|Олег Дубина}}]]
Alves spent eight seasons with the [[Catalonia]]ns when the club signed him from {{w|Sevilla FC}}. Making his debut for Barça in 2008, Alves played 391 matches and he played the second most games of any foreign player in the Catalan jersey after [[Lionel Messi]]. In his last league appearance for Barça, he provided an assist to [[Luis Suárez]] which was Alves's 100th assist in [[La Liga]].
....
FC Barcelona has invited him to address club supporters at the 2016&amp;ndash;17 season start in farewell.
</text>
</revision>
</page>

I need to extract and tokenize the title and text together (or just the title).

So, a proposed solution would be to make the modification below in the process_article function, adding a parameter fields=None (fields: str or tuple of str, optional; Names of fields to consider as textual content ('text' and/or 'title'). If None, only text fields are considered.)

    fields = tuple(fields.split()) if isinstance(fields, str) else fields
    fields = set(fields) if fields else None
    TEXT_OR_BODY = set(("text", "body"))
    text, title, pageid = args
    text_content = filter_wiki(text) if (fields is None) or (fields.intersection(TEXT_OR_BODY)) else ""
    text_content = f"{title} . {text_content}" if ('title' in fields) else text_content
    text_content = text_content.strip()
    result = tokenizer_func(text_content, token_min_len, token_max_len, lower)
    return result, title, pageid

This change is in the forked repo https://github.com/LINE-PESC/gensim
Can I submit a pull request?

[gojomo]

Sorry I didn't notice your reply earlier.

Note: you can always submit a proposal PR, because a compact (& ideally complete) PR is the easiest way to see the precise impact of whatever you're proposing. It's far easier to understand than an incomplete snippet (which here only indirectly refers to the change-in-function-signature and doesn't dhow how its doc-comment would need updating) & reference to other codebase.

In your snippet, I don't get the meaning of 'body', which doesn't seem to be a relevant label in nearby code or the underlying XML. I'm unsure why it's allowing an open-ended list of fields (and as either a string or asequence) when (a) only 'text', 'body', and 'title' are ever relevant; (b) 'text' and 'bod' appear to be functionally synonyms.

To the extent the snippet implies your need, it seems just-as-easily met outside process_article(), because it returns enough for you to do what you want.

For example, instead of calling a modified more-complex process_article() to do…

tokens, title, pageid = process_article(..., fields="title text")

…wouldn't it be sufficient to get the same end-result to just, using the more-simple original process_article():

tokens, title, pageid = process_article(...)
tokens = tokenizer(title, ...) + ('.') + tokens

(I could be missing something.)