Spark NLP 分类器总是预测同一个类

如何解决Spark NLP 分类器总是预测同一个类

我正在使用 Spark NLP 训练分类模型。我关注了this tutorial，以下大部分代码都来自那里。

这是我的训练脚本：

from pyspark.sql import SparkSession
from pyspark.ml import Pipeline

from sparknlp.annotator import *
from sparknlp.common import *
from sparknlp.base import *

import pandas as pd

import sparknlp
spark = sparknlp.start(gpu=True)

# has only 2 columns: category and description
DF = spark.read \
      .option("header",True) \
      .csv("data.csv")

(trainingData,testData) = DF.randomSplit([0.7,0.3],seed = 100)

document_assembler = DocumentAssembler() \
      .setInputCol("description") \
      .setoutputCol("document")

sent_embeddings = BertSentenceEmbeddings.pretrained("sent_biobert_clinical_base_cased","en") \
      .setInputCols("document") \
      .setoutputCol("sentence_embeddings")

classsifierdl = ClassifierDLApproach()\
      .setInputCols("sentence_embeddings")\
      .setoutputCol("class")\
      .setLabelColumn("category")\
      .setMaxEpochs(5)\
      .setLr(0.5)\
      .setDropout(0.5)\
      .setEnableOutputLogs(True)

clf_pipeline = Pipeline(
    stages=[document_assembler,sent_embeddings,classsifierdl])

clf_pipelineModel = clf_pipeline.fit(trainingData)

from sklearn.metrics import classification_report,accuracy_score
df = clf_pipelineModel.transform(testData).select('category','description',"class.result").toPandas()
df['result'] = df['result'].apply(lambda x: x[0])
print(classification_report(df.category,df.result))
print(accuracy_score(df.category,df.result))

然而，分类器总是预测同一个类。我哪里弄错了？谢谢。