mengxr / spark-corenlp

a Stanford CoreNLP wrapper for Spark ML pipeline API

Stanford CoreNLP wrapper for Apache Spark

This package wraps Stanford CoreNLP annotators as Spark DataFrame functions following the simple APIs introduced in Stanford CoreNLP 3.7.0.

This package requires Java 8 and CoreNLP to run. Users must include CoreNLP model jars as dependencies to use language models.

All functions are defined under com.databricks.spark.corenlp.functions.

• cleanxml: Cleans XML tags in a document and returns the cleaned document.
• tokenize: Tokenizes a sentence into words.
• ssplit: Splits a document into sentences.
• pos: Generates the part of speech tags of the sentence.
• lemma: Generates the word lemmas of the sentence.
• ner: Generates the named entity tags of the sentence.
• depparse: Generates the semantic dependencies of the sentence and returns a flattened list of (source, sourceIndex, relation, target, targetIndex, weight) relation tuples.
• coref: Generates the coref chains in the document and returns a list of (rep, mentions) chain tuples, where mentions are in the format of (sentNum, startIndex, mention).
• natlog: Generates the Natural Logic notion of polarity for each token in a sentence, returned as up, down, or flat.
• openie: Generates a list of Open IE triples as flat (subject, relation, target, confidence) tuples.
• sentiment: Measures the sentiment of an input sentence on a scale of 0 (strong negative) to 4 (strong positive).

Users can chain the functions to create pipeline, for example:

import org.apache.spark.sql.functions._
import com.databricks.spark.corenlp.functions._

val input = Seq(
  (1, "<xml>Stanford University is located in California. It is a great university.</xml>")
).toDF("id", "text")

val output = input
  .select(cleanxml('text).as('doc))
  .select(explode(ssplit('doc)).as('sen))
  .select('sen, tokenize('sen).as('words), ner('sen).as('nerTags), sentiment('sen).as('sentiment))

output.show(truncate = false)

+----------------------------------------------+------------------------------------------------------+--------------------------------------------------+---------+
|sen                                           |words                                                 |nerTags                                           |sentiment|
+----------------------------------------------+------------------------------------------------------+--------------------------------------------------+---------+
|Stanford University is located in California .|[Stanford, University, is, located, in, California, .]|[ORGANIZATION, ORGANIZATION, O, O, O, LOCATION, O]|1        |
|It is a great university .                    |[It, is, a, great, university, .]                     |[O, O, O, O, O, O]                                |4        |
+----------------------------------------------+------------------------------------------------------+--------------------------------------------------+---------+

Dependencies

Because CoreNLP depends on protobuf-java 3.x but Spark 2.3 depends on protobuf-java 2.x, we release spark-corenlp as an assembly jar that includes CoreNLP as well as its transitive dependencies, except protobuf-java being shaded. This might cause issues if you have CoreNLP or its dependencies on the classpath.

To use spark-corenlp, you need one of the CoreNLP language models:

# Download one of the language models. 
wget http://repo1.maven.org/maven2/edu/stanford/nlp/stanford-corenlp/3.9.1/stanford-corenlp-3.9.1-models.jar
# Run spark-shell 
spark-shell --packages databricks/spark-corenlp:0.3.1-s_2.11 --jars stanford-corenlp-3.9.1-models.jar

Acknowledgements

Many thanks to Jason Bolton from the Stanford NLP Group for API discussions.