abzif / babzel Goto Github PK

I'd like to use the pre-trained models for lemmatization of text from Java program.
I see, I'd need a complex analysis chain if I am using Solr. I'd like to skip the unnecessary complexity of Solr.
I took a look at the tool's command implementation and this is what I am guessing:

1. Make a TextNormalizer like: var normalizer = new LuceneTextNormalizer;
2. Run the input text through the normalizer: String normalized = normalizer.normalizeText(inputText);
3. Then use sentenceDetector, posTaggar and lemmatizer as used in ToolsCmd#verifyModels.

Am I understanding this correctly?

Another question is, do you have these artifacts (JARs) published in a public repo? Or do I have to compile and install them locally?

This is just a question.
Is the model listed as Chinese for Simplified (script) Chinese, Traditional Chinese, or support both?

End of sentence characters characters are not correct for far-east languages. Sentence-detector is of low quality.
Add some class which computes appropriate EOS characters for a language (or computes default set !.?)

I found out "tee-shirts" in French don't get lemmatized to "tee-shirt". I thought this is may be happening just because the "tee-shirt" is a loan word. To verify this hypothesis, I tried a few French compounds listed in https://www.colanguage.com/french-compound-nouns

But none of these French native compound words don't get lemmatized properly either.

As seen, the first noun element stays in the plural form and only the second noun gets lemmatized.

I also noticed that "chefs-d'oeuvre" gets tokenized strangely. This compound word is tokenized into two tokens "chefs-d'" and "oeuvre".

Two sets of models are produced for a language (simple and lucene). They differ only how raw text are pre-processed. This is unnecessarily complicated and requires twice as much time to train models.

Implement common text pre-processor which can be used from java code/solr/elastic. It may consist of:

• lowercasing
• normalization
• accent drop / ligature expansion

I am seeing inconsistent tokenization for "T-shirt" and probably for any hyphen separated words.
Below, the italic is the input and the bold is the output

$ bin/opennlp TokenizerME en-tokenizer.onlpm
Loading Tokenizer model ... done (0.123s)
yellow t-shirt
yellow t - shirt
yellow t-shirt!
yellow t- shirt !

"t-shirt" is sometimes tokenized as three tokens and sometimes two tokens.

I tested with openNLP 1.9.1 and 2.1.0 and they show the same results.

The model distributed in opennlp.apache.org, opennlp-en-ud-ewt-tokens-1.0-1.9.3.bin, doesn't have this problem. "t-shirt" is always tokenized as "t", "-", "shirt". I tested other words like "ice-cream", "truck-driver", "sign-in", "warm-up", and they are consistent in that "-" is a separate token.

Is there any cure on this?