[OPENNLP-1366] Training of MaxEnt Model with large corpora fails with java.io.UTFDataFormatException

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.9.4, 2.0.0
• Fix Version/s: 2.1.0
• Component/s: Lemmatizer
• Labels: pull-request-available
• Assignee: Richard Zowalla
• Reporter: Richard Zowalla
• Created: 11/Apr/22 14:21
• Updated: 09/Dec/22 10:45
• Resolved: 07/Nov/22 13:10

Description

As written on [dev@opennlp.a.o|https://lists.apache.org/thread/vc5lfzj81tco703noqxpvy8sfj8fw8b1], we are working on training a large opennlp maxent model for lemmatizing
German texts. We use a wikipedia tree bank from Tübingen.

This consumes > 2 TB of RAM during training but will finish after some time. However, writing this model will result in a  java.io.UTFDataFormatException 

However, training such a big model isn't feasable for debugging. Gladly, a similar with a smaller dataset is found on Stackoverflow: https://stackoverflow.com/questions/70064477/opennlp-lemmatizertrainer-utfdataformatexception-encoded-string-too-long 

It contains the OpenNLP CLI command to train a lemmatizer on a much smaller dataset.

The stacktrace is raced while writing a String as UTF in DataOutputStream, which has a hard-coded size limitation in the JDK (for reasons behind my knowledge )

Stacktrace:

java.io.UTFDataFormatException: encoded string too long: 383769 bytes
        at java.base/java.io.DataOutputStream.writeUTF(DataOutputStream.java:364)
        at java.base/java.io.DataOutputStream.writeUTF(DataOutputStream.java:323)
        at opennlp.tools.ml.maxent.io.BinaryGISModelWriter.writeUTF(BinaryGISModelWriter.java:71)
        at opennlp.tools.ml.maxent.io.GISModelWriter.persist(GISModelWriter.java:97)
        at opennlp.tools.ml.model.GenericModelWriter.persist(GenericModelWriter.java:75)
        at opennlp.tools.util.model.ModelUtil.writeModel(ModelUtil.java:71)
        at opennlp.tools.util.model.GenericModelSerializer.serialize(GenericModelSerializer.java:36)
        at opennlp.tools.util.model.GenericModelSerializer.serialize(GenericModelSerializer.java:29)
        at opennlp.tools.util.model.BaseModel.serialize(BaseModel.java:597)
        at opennlp.tools.cmdline.CmdLineUtil.writeModel(CmdLineUtil.java:182)
        at opennlp.tools.cmdline.lemmatizer.LemmatizerTrainerTool.run(LemmatizerTrainerTool.java:77)
        at opennlp.tools.cmdline.CLI.main(CLI.java:256) 

Issue Links

• is duplicated by: OPENNLP-1218 All binary implementation of AbstractModelWriter and DataReader throws java.io.UTFDataFormatException when large dataset is used for training
• links to: How to handle LemmatizerTrainer 'UTFDataFormatException: encoded string too long'?