SparkNLP - Quick Start

Last updated:

Requirements & Setup

SparkNLP is built on top of Apache Spark 2.3.0 and works with any user provided Spark 2.x.x it is advised to have basic knowledge of the framework and a working environment before using Spark-NLP. Refer to its documentation to get started with Spark.

To start using the library, execute any of the following lines depending on your desired use case:

spark-shell --packages JohnSnowLabs:spark-nlp:2.0.1
pyspark --packages JohnSnowLabs:spark-nlp:2.0.1
spark-submit --packages JohnSnowLabs:spark-nlp:2.0.1

Databricks cloud cluster & Apache Zeppelin


For Python in Apache Zeppelin you may need to setup SPARK_SUBMIT_OPTIONS utilizing --packages instruction shown above like this

export SPARK_SUBMIT_OPTIONS="--packages JohnSnowLabs:spark-nlp:2.0.1"

Python Jupyter Notebook with PySpark

export SPARK_HOME=/path/to/your/spark/folder

pyspark --packages JohnSnowLabs:spark-nlp:2.0.1

Straight forward Python on jupyter notebook

Use pip to install (after you pip installed pyspark)

pip install spark-nlp==2.0.1
jupyter notebook

In this way, you will have to start SparkSession in your python program manually, this is an example

spark = SparkSession.builder \
    .master('local[4]') \
    .appName('OCR Eval') \
    .config("spark.driver.memory", "6g") \
    .config("spark.executor.memory", "6g") \
    .config("spark.jars.packages", "JohnSnowLabs:spark-nlp:2.0.1") \

S3 based standalone cluster (No Hadoop)

If your distributed storage is S3 and you don't have a standard hadoop configuration (i.e. fs.defaultFS) You need to specify where in the cluster distributed storage you want to store Spark-NLP's tmp files. First, decide where you want to put your application.conf file

import com.johnsnowlabs.uti.ConfigLoader

And then we need to put in such application.conf the following content

sparknlp {
  settings {
    cluster_tmp_dir = "somewhere in s3n:// path to some folder"

Pre-Compiled Spark-NLP for download

Pre-compiled Spark-NLP assembly fat-jar here Pre-compiled Spark-NLP assembly fat-jar with Tensorflow GPU enabled here then, run spark-shell or spark-submit with appropriate --jars /path/to/spark-nlp_2.11-2.0.1.jar to use the library in spark.

For further alternatives and documentation check out our README page in GitHub.


Spark ML provides a set of Machine Learning applications, and it's logic consists of two main components: Estimators and Transformers. The first, have a method called fit() which secures and trains a piece of data to such application, and a Transformer, which is generally the result of a fitting process, applies changes to the the target dataset. These components have been embedded to be applicable to Spark NLP. Pipelines are a mechanism that allow multiple estimators and transformers within a single workflow, allowing multiple chained transformations along a Machine Learning task. Refer to SparkML library for more information.


An annotation is the basic form of the result of a Spark-NLP operation. It's structure is made of:

  • annotatorType: which annotator generated this annotation
  • begin: the begin of the matched content relative to raw-text
  • end: the end of the matched content relative to raw-text
  • metadata: content of matched result and additional information

This object is automatically generated by annotators after a transform process. No manual work is required. But it must be understood in order to use it efficiently.


Annotators are the spearhead of NLP functionalities in SparkNLP. There are two forms of annotators:

  • Annotator Approaches: Are those who represent a Spark ML Estimator and require a training stage. They have a function called fit(data) which trains a model based on some data. They produce the second type of annotator which is an annotator model or transformer.
  • Annotator Model: They are spark models or transformers, meaning they have a transform(data) function which take a dataset and add to it a column with the result of this annotation. All transformers are additive, meaning they append to current data, never replace or delete previous information.

Both forms of annotators can be included in a Pipeline and will automatically go through all stages in the provided order and transform the data accordingly. A Pipeline is turned into a PipelineModel after the fit() stage. Either before or after can be saved and re-loaded to disk at any time.

Common Functions

  • setInputCols(column_names): Takes a list of column names of annotations required by this annotator
  • setOutputCol(column_name): Defines the name of the column containing the result of this annotator. Use this name as an input for other annotators requiring the annotations of this one.

Quickly annotate some text

Basic Pipeline

A basic pipeline is a downloadable Spark ML pipeline ready to compute some annotations for you right away. Keep in mind this is trained with generic data and is not meant for production use, but should give you an insight of how the library works at a glance.

Downloading and using a pretrained pipeline

Basic pipelines allow to inject either common strings or Spark dataframes. Either way, SparkNLP will execute the right pipeline type to bring back the results. Let's retrieve it:

import com.johnsnowlabs.nlp.pretrained.pipelines.en.BasicPipeline

val annotations = BasicPipeline().annotate(Array("We are very happy about SparkNLP", "And this is just another sentence"))

annotations.foreach(println(_, "\n"))
(Map(lemma -> List(We, be, very, happy, about, SparkNLP), document -> List(We are very happy about SparkNLP), normal -> List(We, are, very, happy, about, SparkNLP), pos -> ArrayBuffer(PRP, VBP, RB, JJ, IN, NNP), token -> List(We, are, very, happy, about, SparkNLP)))
(Map(lemma -> List(And, this, be, just, another, sentence), document -> List(And this is just another sentence), normal -> List(And, this, is, just, another, sentence), pos -> ArrayBuffer(CC, DT, VBZ, RB, DT, NN), token -> List(And, this, is, just, another, sentence)))

Using a pretrained pipeline with spark dataframes

Since this is Apache Spark, we should make use of our DataFrame friends. Pretraned pipelines allow us to do so with the same function, just adding the target string column. The result here will be a dataframe with many annotation columns.

import spark.implicits._

val data = Seq("hello, this is an example sentence").toDF("mainColumn")

BasicPipeline().annotate(data, "mainColumn").show()
|                text|            document|               token|              normal|               lemma|                 pos|
|hello, this is an...|[[document, 0, 33...|[[token, 0, 4, he...|[[token, 0, 4, he...|[[token, 0, 4, he...|[[pos, 0, 4, UH, ...|

Manipulating pipelines

To add a bit of challenge, the output of the previous DataFrame was in terms of Annotation objects. What if we want to deal with just the resulting annotations? We can use the Finisher annotator, retrieve the BasicPipeline, and add them together in a Spark ML Pipeline. Note BasicPipeline expect the target column to be named "text".

import com.johnsnowlabs.nlp.Finisher

val finisher = new Finisher().
    setInputCols("token", "normal", "lemma", "pos")

val basicPipeline = BasicPipeline().pretrained()

val pipeline = new Pipeline().

    transform(data.withColumnRenamed("mainColumn", "text")).
|                text|      finished_token|     finished_normal|      finished_lemma|        finished_pos|
|hello, this is an...|[hello, ,, this, ...|[hello, this, is,...|[hello, ,, this, ...|[UH, DT, VBZ, DT,...|

Train your own pipeline

Annotator types

Every annotator has a type. Those annotators that share a type, can be used interchangeably, meaning you could you use any of them when needed. For example, when a token type annotator is required by another annotator, such as a sentiment analysis annotator, you can either provide a normalized token or a lemma, as both are of type token.

Necessary imports

Since version 1.5.0 we are making necessary imports easy to reach, base._ will include general Spark NLP transformers and concepts, while annotator._ will include all annotators that we currently provide. We also need SparkML pipelines.

import com.johnsnowlabs.nlp.base._
import com.johnsnowlabs.nlp.annotator._

DocumentAssembler: Getting data in

In order to get through the NLP process, we need to get raw data annotated. There is a special transformer that does this for us: the DocumentAssembler, it creates the first annotation of type Document which may be used by annotators down the road

val documentAssembler = new DocumentAssembler().

Sentence detection and tokenization

In this quick example, we now proceed to identify the sentences in each of our document lines. SentenceDetector requires a Document annotation, which is provided by the DocumentAssembler output, and it's itself a Document type token. The Tokenizer requires a Document annotation type, meaning it works both with DocumentAssembler or SentenceDetector output, in here, we use the sentence output.

val sentenceDetector = new SentenceDetector().

val regexTokenizer = new Tokenizer().

Using SparkML Pipeline

Now we want to put all this together and retrieve the results, we use a Pipeline for this. We also include another special transformer, called Finisher to show tokens in a human language. We use an emptyDataFrame in fit() since none of the pipeline stages have a training stage.

val testData = Seq("Lorem ipsum dolor sit amet, " +
    "consectetur adipiscing elit, sed do eiusmod tempor " +
    "incididunt ut labore et dolore magna aliqua.").toDF("text")

val finisher = new Finisher().

val pipeline = new Pipeline().

    transform(Seq("hello, this is an example sentence").toDF("text")).

Using LightPipeline

LightPipeline is a Spark-NLP specific Pipeline class equivalent to SparkML's Pipeline. The difference is that it's execution does not hold to Spark principles, instead it computes everything locally (but in parallel) in order to achieve fast results when dealing with small amounts of data. This means, we do not input a Spark Dataframe, but a string or an Array of strings instead, to be annotated. To create Light Pipelines, you need to input an already trained (fit) Spark ML Pipeline. It's transform() stage is converted into annotate() instead.

import com.johnsnowlabs.nlp.base._

val trainedModel =[String].toDF("text"))

val lightPipeline = new LightPipeline(trainedModel)

lightPipeline.annotate("Hello world, please annotate my text")

Utilizing Spark-NLP OCR PDF Converter

Installing Spark-NLP OCRHelper

First, either build from source or download the following standalone jar module (works both from Spark-NLP python and scala): Spark-NLP-OCR And add it to your Spark environment (with --jars or spark.driver.extraClassPath and spark.executor.extraClassPath configuration) Second, if your PDFs don't have a text layer (this depends on how PDFs were created), the library will use Tesseract 4.0 on background. Tesseract will utilize native libraries, so you'll have to get them installed in your system. You can also use the following coordinates for Maven: com.johnsnowlabs.nlp:spark-nlp_2.11:2.0.1

Creating Spark datasets from PDF (For SparkML Pipeline)

You can use OcrHelper to directly create spark dataframes from PDF. This will hold entire documents in single rows, meant to be later processed by a SentenceDetector. This way, you won't be breaking the content in rows as if you were reading a standard document. Metadata column will include page numbers and file name information per row.


val data = OcrHelper.createDataset(spark, "/pdfs/", "text", "metadata")

val documentAssembler = new DocumentAssembler().setInputCol("text").setMetadataCol("metadata")


Creating an Array of Strings from PDF (For LightPipeline)

Another way, would be to simply create an array of strings. This is useful for example if you are parsing a small amount of pdf files and would like to use LightPipelines instead. See an example below.


val raw = OcrHelper.createMap("/pdfs/")

val documentAssembler = new DocumentAssembler().setInputCol("text").setOutputCol("document")

val sentenceDetector = new SentenceDetector().setInputCols("document").setOutputCol("sentence")

val lightPipeline = new LightPipeline(new Pipeline().setStages(Array(documentAssembler, sentenceDetector)).fit(Seq.empty[String].toDF("text")))


Training annotators

Training methodology

Training your own annotators is the most key concept when dealing with real life scenarios. Any of the annotators provided above, such as pretrained pipelines and models, will rarely ever apply to a specific use case. Dealing with real life problems will require training your own models. In Spark-NLP, training annotators will vary depending on the annotators. Currently, we support three ways:

  1. Training from an external source: Most of our annotators train from an external file or folder passed to the annotator as a param. You will see such ones as setCorpus() or setDictionary() param setter methods, allowing you to configure the input to use. You can set Spark-NLP to read them as Spark datasets or LINE_BY_LINE which is usually faster for small enough corpora
  2. Some annotators are capable of training through the dataset passed to fit(). This is the standard behavior in Spark ML, but unfortunately in NLP, not all annotators use structured data to train. Hence, we provide external resource capabilities. This form of training might be used when 'setCorpus' is not set or any form for setLabelCol() is set instead. Check the reference to see which annotators allow training this one.
  3. Last but not least, some of our annotators are Deep Learning based. These models may be trained either on Tensorflow and read utilizing internal tools, or converted into Spark-NLP models (Using tensorflow graphs) and allowing them to be trainable from Spark-NLP as well.

More examples in Scala and Python

You can checkout the reference for pretrained models, and to check which params you can set in order to train your own.

Where to go next?

Documentation and reference

Detailed information about Spark-NLP concepts, annotators and more may be found HERE

More examples in Scala and Python

We are working on examples to show you how the library may be used in different scenarios, take a look HERE