How to remove non-valid unicode characters from strings in java

In a way, both answers provided by Mukesh Kumar and GsusRecovery are helping, but not fully correct.

document.replaceAll("[^\\u0009\\u000a\\u000d\\u0020-\\uD7FF\\uE000-\\uFFFD]", "");

seems to replace all invalid characters. But CoreNLP seems to not support even more. I manually figured them out by running the parser on my whole corpus, which led to this:

document.replaceAll("[\\uD83D\\uFFFD\\uFE0F\\u203C\\u3010\\u3011\\u300A\\u166D\\u200C\\u202A\\u202C\\u2049\\u20E3\\u300B\\u300C\\u3030\\u065F\\u0099\\u0F3A\\u0F3B\\uF610\\uFFFC]", "");

So right now I am running two replaceAll() commands before handing the document to the parser. The complete code snippet is

// remove invalid unicode characters
String tmpDoc1 = document.replaceAll("[^\\u0009\\u000a\\u000d\\u0020-\\uD7FF\\uE000-\\uFFFD]", "");
// remove other unicode characters coreNLP can't handle
String tmpDoc2 = tmpDoc1.replaceAll("[\\uD83D\\uFFFD\\uFE0F\\u203C\\u3010\\u3011\\u300A\\u166D\\u200C\\u202A\\u202C\\u2049\\u20E3\\u300B\\u300C\\u3030\\u065F\\u0099\\u0F3A\\u0F3B\\uF610\\uFFFC]", "");
DocumentPreprocessor tokenizer = new DocumentPreprocessor(new StringReader(tmpDoc2));
for (List<HasWord> sentence : tokenizer) {
    List<TaggedWord> tagged = tagger.tagSentence(sentence);
    GrammaticalStructure gs = parser.predict(tagged);
    System.err.println(gs);
}

This is not necessarily a complete list of unsupported characters, though, which is why I opened an issue on GitHub.

Please note that CoreNLP automatically removes those unsupported characters. The only reason I want to preprocess my corpus is to avoid all those error messages.

UPDATE Nov 27ths

Christopher Manning just answered the GitHub Issue I opened. There are several ways to handle those characters using the class edu.stanford.nlp.process.TokenizerFactory;. Take this code example to tokenize a document:

DocumentPreprocessor tokenizer = new DocumentPreprocessor(new StringReader(document));
TokenizerFactory<? extends HasWord> factory=null;
factory=PTBTokenizer.factory();
factory.setOptions("untokenizable=noneDelete");
tokenizer.setTokenizerFactory(factory);

for (List<HasWord> sentence : tokenizer) {
    // do something with the sentence
}

You can replace noneDeletein line 4 with other options. I am citing Manning:

"(...) the complete set of six options combining whether to log a warning for none, the first, or all, and whether to delete them or to include them as single character tokens in the output: noneDelete, firstDelete, allDelete, noneKeep, firstKeep, allKeep."

That means, to keep the characters without getting all those error messages, the best way is to use the option noneKeep. This way is way more elegant than any attempt to remove those characters.

Remove specific unwanted chars with:

document.replaceAll("[\\uD83D\\uFFFD\\uFE0F\\u203C\\u3010]", "");

If you found others unwanted chars simply add with the same schema to the list.

UPDATE:

The unicode chars are splitted by the regex engine in 7 macro-groups (and several sub-groups) identified by a one letter (macro-group) or two letters (sub-group).

Basing my arguments on your examples and the unicode classes indicated in the always good resource Regular Expressions Site i think you can try a unique only-good-pass approach such as this:

document.replaceAll("[^\\p{L}\\p{N}\\p{Z}\\p{Sm}\\p{Sc}\\p{Sk}\\p{Pi}\\p{Pf}\\p{Pc}\\p{Mc}]","")

This regex remove anything that is not:

• \p{L}: a letter in any language
• \p{N}: a number
• \p{Z}: any kind of whitespace or invisible separator
• \p{Sm}\p{Sc}\p{Sk}: Math, Currency or generic marks as single char
• \p{Mc}*: a character intended to be combined with another character that takes up extra space (vowel signs in many Eastern languages).
• \p{Pi}\p{Pf}\p{Pc}*: Opening quote, Closing quote, words connectors (i.e. underscore)

*: i think these groups can be eligible to be removed as well for the purpose of CoreNPL.

This way you only need a single regex filter and you can handle groups of chars (with the same purpose) instead of single cases.