Build and Convert a Spark NLP Pipeline into PMML in Apache Zeppelin

[#1]

This article is designed to extend my articles Twitter Sentiment using Spark Core NLP in Apache Zeppelin and Connecting Solr to Spark - Apache Zeppelin Notebook

I have included the complete notebook on my Github site, which can be found on my GitHub site.

Step 1 - Follow the tutorial in the provide articles above, and establish an Apache Solr collection called "tweets"

Step 2 - Verify the version of Apache Spark being used, and visit the Solr-Spark connector site. The key is to match the version of Spark the version of the Solr-Spark connector. In the example below, the version of Spark is 2.2.0, and the connector version is 3.4.4

%spark2
sc
sc.version
</p><p>
Step 3 - Include the Solr-Spark dependency in Zeppelin. Important note: This needs to be run before
the Spark Context has been initialized</p>
<pre>%dep 
z.load("com.lucidworks.spark:spark-solr:jar:3.4.4")
//Must be used before SparkInterpreter (%spark2) initialized
//Hint: put this paragraph before any Spark code and restart Zeppelin/Interpreter

Step 4 - Download the Stanford CoreNLP libraries found on here: <a href="http://nlp.stanford.edu/software/stanford-corenlp-full-2018-02-27.zip">http://nlp.stanford.edu/software/stanford-corenlp-full-2018-02-27.zip</a>

 Upzip the download and move it to the /tmp directory. Note: This can be accomplished
on the command line or the following Zeppelin paragraph will work as well
%sh 
wget /tmp/stanford-corenlp-full-2018-02-27.zip http://nlp.stanford.edu/software/stanford-corenlp-full-2018-02-27.zip unzip /tmp/stanford-corenlp-full-2018-02-27.zip

 Step 5 - In Zeppelin's
Interpreters configurations for Spark, include the following artifact:
/tmp/stanford-corenlp-full-2018-02-27/stanford-corenlp-3.9.1-models.jar

Step 6 - Include the following Spark dependencies for Stanford CoreNLP and Spark CoreNLP. Important
note: This needs to be run before the Spark Context has been initialized
%dep
z.load("edu.stanford.nlp:stanford-corenlp:3.9.1")
//In Spark Interper Settings Add the following artifact
// /tmp/stanford-corenlp-full-2018-02-27/stanford-corenlp-3.9.1-models.jar
%dep
z.load("databricks:spark-corenlp:0.2.0-s_2.11")

 Step 7  Include the following Spark dependencies for 
 JPMML-SparkML and JPMML-Model. Important note: This needs to be run before the Spark Context has been initialized.
%dep
z.load("org.jpmml:jpmml-sparkml:1.4.5")

</p><pre>%dep
z.load("org.jpmml:pmml-model:1.4.5")

 Step 8 - Run Solr query and return results into Spark DataFrame. Note: Zookeeper host might need to use full names:
 "zkhost" -> "host-1.domain.com:2181,host-2.domain.com:2181,host-3.domain.com:2181/solr"
%spark2 
val options = Map( "collection" -> 
 "Tweets", "zkhost" -> "localhost:2181/solr", 
 // "query" -> "Keyword, 'More Keywords'" 
) 
val df = spark.read.format("solr").options(options).load df.cache()
</p><p>
 Step 9 - Review results of the Solr query</p><pre>%spark2 
df.count() 
df.printSchema() 
df.take(1)
</p><p>
 Step 10 - Filter the
Tweets in the Spark DataFrame to ensure the Tweet text isn't null Once filter
has been completed, add the sentiment value to the tweets.</p>
<pre>%spark2 
import org.apache.spark.sql.functions._ 
import org.apache.spark.sql.types._
import com.databricks.spark.corenlp.functions._ 
val df_TweetSentiment = df.filter("text_t is not null").select($"text_t", sentiment($"text_t").as('sentimentScore))

  Step 11 - Valid
results
%spark2 
df_TweetSentiment.cache() 
df_TweetSentiment1.printSchema() 
df_TweetSentiment1.take(1) 
df_TweetSentiment1.count()

 Step 12 - Build Stages
to build features that will be fed into a Logistic Regression model for
classification

   Stage 1 -Regex Tokenizer will be used to separate
each word into individual "tokens"

   Stage 2 -Count Vectorizer will count the number
of occurrences each token occurs in the text corpus

   Stage 3 -Term frequency-inverse document
frequency (TF-IDF) is a feature vectorization method widely used in text mining
to reflect the importance of a term to a document in the corpus.

   Stage 4 -Logistic
Regression for classification to predict sentiment score
%spark2
import org.apache.spark.ml.feature.{HashingTF, IDF, Tokenizer, RegexTokenizer, CountVectorizer, CountVectorizerModel}
import org.apache.spark.ml.classification.LogisticRegression
val tokenizer = new RegexTokenizer()
.setInputCol("text_t")
.setOutputCol("words")
.setPattern("\W+")
.setGaps(true)
val wordsData = tokenizer.transform(df_TweetSentiment)
val cvModel = new CountVectorizer()
.setInputCol("words").setOutputCol("rawFeatures")
.setMinDF(4)
.fit(wordsData)
val featurizedData = cvModel.transform(wordsData)
val idf = new IDF()
.setInputCol("rawFeatures")
.setOutputCol("features")
val idfModel = idf.fit(featurizedData)
val rescaledData = idfModel.transform(featurizedData)
rescaledData.select("sentimentScore", "features").show()
val lr = new LogisticRegression()
.setMaxIter(10)
.setRegParam(0.001)
.setLabelCol("sentimentScore")

 Step 13 - Build Spark Pipeline from Stages
%spark2
import org.apache.spark.ml.{Pipeline, PipelineModel}
import org.apache.spark.ml.classification.LogisticRegression
import org.apache.spark.ml.linalg.Vector
import org.apache.spark.sql.Row
import org.apache.spark.sql.types.StructType
val pipeline = new Pipeline()
.setStages(Array(tokenizer, cvModel, idfModel, lr))
val PipeLineModel = pipeline.fit(df_TweetSentiment)
//Save Pipeline to Disk (Optional)
PipeLineModel.write.overwrite().save("/tmp/spark-IDF-model")
val schema = df_TweetSentiment.schema
Step 14 - Export Spark Pipeline to PMML using JPMML-SparkML
%spark2
import org.jpmml.sparkml.PMMLBuilder
import java.io.File
val pmml = new PMMLBuilder(schema, PipeLineModel)
val file = pmml.buildFile(new File("/tmp/TweetPipeline.pmml"))

[#2]

Hi Ian! Thanks for posting this article.

At the "pmml.buildFile" I'm getting the following error: java.lang.IllegalArgumentException: iperbole_

Any ideas? Thank you very much!