SentenceTokenizer incorrectly processes punctuation marks within words

[nataliashitova]

Explanation

The currently used SentenceTokenizer generates wrong results when a punctuation mark such as ! or ? or . are used within a word (e.g., in a company name).

Examples

Example 1

The following text (see Yoast/wordpress-seo#13726)

The free App FRITZ!App WLAN helps to find the ideal locations when setting up a repeater.

gets incorrectly parsed into the following sentences

0: "The free App FRITZ!"
1: "App WLAN helps to find the ideal locations when setting up a repeater."

Example 2

The same text as in Example 1 but with a . instead of the !

The free App FRITZ.App WLAN helps to find the ideal locations when setting up a repeater.

gets correcty parsed into one sentence

0: "The free App FRITZ.App WLAN helps to find the ideal locations when setting up a repeater."

Example 3

The same text as in Example 2 but the entire word FRITZ.APP capitalized

The free App FRITZ.APP WLAN helps to find the ideal locations when setting up a repeater.

gets incorrectly parsed into the following sentences

0: "The free App FRITZ."
1: "APP WLAN helps to find the ideal locations when setting up a repeater."

Why does it happen?

The problem in Example 1 occurs because the SentenceTokenizer splits text on !, ?, ; and ... without checking if the cut-off part begins as a proper sentence should (e.g., with a space and a capital letter). Here is the rule where this check should take place.

Note that such a check is implemented for the situation when the text is split on a .. Specifically, the rule checks if the second letter of the cut-off remainder text is a capital letter, or a number, etc.
However, the SentenceTokenizer does not check that the first letter of the remainder text is a space. Which is a reason why the problem in Example 3 occurs.

Things to consider

1. A fix for both problems seems to be pretty straight-forward to implement.
2. A few users complained about these issues.
3. The currently used SentenceTokenizer will not be used in its current form when the tree-based text parser is implemented, because the said tokenizer relies on HTML tags.
4. We will still need a variant of a sentence tokenizer to be able to operate with sentences in researches. The work on implementing fixes to the current sentence tokenizer will not necessarily be lost therefore.