Questions on training and using corenlp-chinese-smoothnlp-0.1-with-dependencies.jar

[zegzag]

I've trained the corenlp-chinese-smoothnlp-0.1-with-dependencies.jar on a new corps ./sentiment_output_train_combined.txt, which sentiment category is 6. (terminal): java -mx8g -cp corenlp-chinese-smoothnlp-0.1-with-dependencies.jar edu.stanford.nlp.sentiment.SentimentTraining -numHid 20 -trainPath sentimen_output_train_combined.txt -train -model model.ser.gz -numClasses 6 -classNames "Very negative,Negative,Neutral,Positive,Very positive,Unknown" And I got the training model ./model.ser.gz But when I used it, it seemed that the result came from the original inner model but not the updated model, with means that the updated model didn't work. (terminal):java -jar corenlp-chinese-smoothnlp-0.1-with-dependencies.jar sentiment.model model.ser.gz When I use it in python notebook, the result seemed to be unchanged The sentimentDistribution of the test text on the new model is totally the same with "NLP_Utils/demo.ipynb, and its sentiment category is 5, but not 6.

[zegzag]

The model can be found in /nlp.creditx/ZhangZe

[victorzhrn]

Did you try:

wget --post-data '今天天气不错' 'localhost:9000/?properties={"annotators": "tokenize,sentiment", 
"sentiment.model":"/nlp.creditx/ZhangZe.model.ser.gz","outputFormat": "json","pipelineLanguage":"zh"}' -O -

I would most likely to recommend this way of specifying the model addr. The reason why the model file is not properly loaded might be: When I was making the maven project, I made changes to fileIO.class to specify the reading input streams from "resource" of the maven project, so that, unlikely official CoreNLP project, you do not need to download the main skeleton and the model files separately.

Please keep an eye on the maven project which I will hopefully update it in the next few days. It would show more details on how the jar is compiled and how you may adjust different file IO.

Thx

[zegzag]

Nice! It works with wget.