TrueCaseAnnotator's truecase.bias not having any effect

[dzenilee]

On https://stanfordnlp.github.io/CoreNLP/truecase.html, truecase.bias is described as:

Biases to choose certain behaviors. You can use this to adjust the proclivities of the truecaser. The truecaser classes are: UPPER, LOWER, INIT_UPPER, and O (for mixed case words like McVey).

It has the default values of INIT_UPPER:-0.7,UPPER:-0.7,O:0, and I'd like to change these, but it looks like the only way to do that (using stanza) is by modifying self.client.start_cmd directly in the constructor. This is what I'm doing:

from stanza.server import CoreNLPClient

class TrueCaseAnnotator(object):
    def __init__(self, classpath=CLASSPATH, bias="INIT_UPPER:-1,UPPER:-1,O:0"):  # note updated values 
        self.client = CoreNLPClient(
            annotators=["tokenize,ssplit,truecase"],
            classpath=classpath,
            output_format='json',
        )
        self.client.start_cmd.append("-bias")
        self.client.start_cmd.append(bias)

And this correctly logs the command with the adjusted bias:

2021-02-04 21:12:51 INFO: Starting server with command: java -Xmx5G -cp /app/artifacts/stanford-corenlp-4.2.0/* edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 60000 -threads 5 -maxCharLength 100000 -quiet False -serverProperties corenlp_server-462b8ca2f7024c78.props -annotators tokenize,ssplit,truecase -preload -outputFormat json -bias INIT_UPPER:-1,UPPER:-1,O:0

Assuming the range for these variables is -1 to 1 and an INIT_UPPER value of -1 means never capitalize the initial word of a sentence, I expect the first word to come out lowercase at all times. However, this is not what's happening.

First, can you confirm -1 to 1 is the correct range? (I've also tried other numbers like -100, 0.2 and 1, and nothing changes).

I did notice that classBias (with emphasis on class) is always set to the default values (INIT_UPPER:-0.7,UPPER:-0.7,O:0) no matter what I pass for the bias parameter (see logs below). Is it possible classBias is overwriting the value of bias? Where is that getting set and what's the difference between the two anyway?

2021-02-04 21:29:55 INFO: Starting server with command: java -Xmx5G -cp /app/artifacts/stanford-corenlp-4.2.0/* edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 60000 -threads 5 -maxCharLength 100000 -quiet False -serverProperties corenlp_server-ceededd66e0947d7.props -annotators tokenize,ssplit,truecase -preload -outputFormat json -bias INIT_UPPER:-1,UPPER:-1,O:0
[main] INFO CoreNLP - --- StanfordCoreNLPServer#main() called ---
[main] INFO CoreNLP - Server default properties:
            (Note: unspecified annotator properties are English defaults)
            annotators = tokenize,ssplit,truecase
            bias = INIT_UPPER:1,UPPER:-100,O:0
            inputFormat = text
            outputFormat = json
            prettyPrint = false
            threads = 5
[main] INFO CoreNLP - Threads: 5
[main] INFO edu.stanford.nlp.pipeline.StanfordCoreNLP - Adding annotator tokenize
[main] INFO edu.stanford.nlp.pipeline.StanfordCoreNLP - Adding annotator ssplit
[main] INFO edu.stanford.nlp.pipeline.StanfordCoreNLP - Adding annotator truecase
[main] INFO edu.stanford.nlp.sequences.SeqClassifierFlags - classBias=INIT_UPPER:-0.7,UPPER:-0.7,O:0
[main] INFO edu.stanford.nlp.sequences.SeqClassifierFlags - loadClassifier=edu/stanford/nlp/models/truecase/truecasing.fast.caseless.qn.ser.gz
[main] INFO edu.stanford.nlp.sequences.SeqClassifierFlags - mixedCaseMapFile=edu/stanford/nlp/models/truecase/MixDisambiguation.list
[main] INFO edu.stanford.nlp.ie.AbstractSequenceClassifier - Loading classifier from edu/stanford/nlp/models/truecase/truecasing.fast.caseless.qn.ser.gz ... done [5.9 sec].
[main] INFO CoreNLP - Starting server...
[main] INFO CoreNLP - StanfordCoreNLPServer listening at /0.0.0.0:9000

[AngledLuffa]

I'm not sure that's the correct flag... I think you want it to be "-truecase.bias" and not "-bias"

[dzenilee]

Thanks, @AngledLuffa. I think that was it (I can see some changes in the output!). But can you still explain what range of values INIT_UPPER, UPPER, and O can take? Setting specifically INIT_UPPER to -1 (what I think may be the lower bound) has no effect on the output even after updating the param name to truecase.bias, i.e., it doesn't lowercase the initial letter of the sentence. It could be that I don't understand how these bias settings work and am expecting something that shouldn't happen, so more extensive documentation would be appreciated!

[AngledLuffa]

Hopefully these comments can help explain it:

https://github.com/stanfordnlp/CoreNLP/blob/master/src/edu/stanford/nlp/ie/crf/CRFBiasedClassifier.java