Merge pull request #70 from JohnSnowLabs/word_embeddings

Word embeddings

build.sbt

@@ -82,7 +82,9 @@ lazy val testDependencies = Seq(
 
 lazy val utilDependencies = Seq(
   "com.typesafe" % "config" % "1.3.0",
-  "org.fusesource.leveldbjni" % "leveldbjni-all" % "1.8"
+  "org.rocksdb" % "rocksdbjni" % "5.8.0",
+  "org.slf4j" % "slf4j-api" % "1.7.25",
+  "org.apache.commons" % "commons-compress" % "1.15"
 )
 
 lazy val root = (project in file("."))

python/sparknlp/annotator.py

@@ -48,6 +48,26 @@ def setOutputCol(self, value):
         return self._set(outputCol=value)
 
 
+class AnnotatorWithEmbeddings(Params):
+    sourceEmbeddingsPath = Param(Params._dummy(),
+                                 "sourceEmbeddingsPath",
+                                 "Word embeddings file",
+                                 typeConverter=TypeConverters.toString)
+    embeddingsFormat = Param(Params._dummy(),
+                             "embeddingsFormat",
+                             "Word vectors file format",
+                             typeConverter=TypeConverters.toInt)
+    embeddingsNDims = Param(Params._dummy(),
+                            "embeddingsNDims",
+                            "Number of dimensions for word vectors",
+                            typeConverter=TypeConverters.toInt)
+
+    def setEmbeddingsSource(self, path, nDims, format):
+        self._set(sourceEmbeddingsPath=path)
+        self._set(embeddingsFormat=format)
+        return self._set(embeddingsNDims=nDims)
+
+
 class AnnotatorTransformer(JavaModel, JavaMLReadable, JavaMLWritable, AnnotatorProperties):
 
     column_type = "array<struct<annotatorType:string,begin:int,end:int,metadata:map<string,string>>>"
@@ -478,7 +498,7 @@ class NorvigSweetingModel(JavaModel, JavaMLWritable, JavaMLReadable, AnnotatorPr
 
 
 
-class NerCrfApproach(JavaEstimator, JavaMLWritable, JavaMLReadable, AnnotatorProperties):
+class NerCrfApproach(JavaEstimator, JavaMLWritable, JavaMLReadable, AnnotatorProperties, AnnotatorWithEmbeddings):
     labelColumn = Param(Params._dummy(),
                      "labelColumn",
                      "Column with label per each token",

src/main/scala/com/johnsnowlabs/ml/crf/DatasetEncoder.scala

@@ -128,4 +128,4 @@ class DatasetEncoder(val startLabel: String = "@#Start") {
       result
     }
   }
-}
+}

src/main/scala/com/johnsnowlabs/ml/crf/DatasetReader.scala

@@ -132,5 +132,3 @@ object DatasetReader {
     }
   }
 }
-
-

src/main/scala/com/johnsnowlabs/nlp/AnnotatorApproach.scala

@@ -2,7 +2,7 @@ package com.johnsnowlabs.nlp
 
 import org.apache.spark.ml.param.ParamMap
 import org.apache.spark.ml.{Estimator, Model}
-import org.apache.spark.sql.Dataset
+import org.apache.spark.sql.{Dataset, SparkSession}
 import org.apache.spark.sql.types.{ArrayType, MetadataBuilder, StructField, StructType}
 import org.apache.spark.ml.util.DefaultParamsWritable
 
@@ -24,8 +24,15 @@ abstract class AnnotatorApproach[M <: Model[M]]
 
   def train(dataset: Dataset[_]): M
 
+  def beforeTraining(spark: SparkSession): Unit = {}
+
+  def onTrained(model: M, spark: SparkSession): Unit = {}
+
   override final def fit(dataset: Dataset[_]): M = {
-    copyValues(train(dataset).setParent(this))
+    beforeTraining(dataset.sparkSession)
+    val model = copyValues(train(dataset).setParent(this))
+    onTrained(model, dataset.sparkSession)
+    model
   }
 
   override final def copy(extra: ParamMap): Estimator[M] = defaultCopy(extra)
@@ -50,5 +57,4 @@ abstract class AnnotatorApproach[M <: Model[M]]
       StructField(getOutputCol, ArrayType(Annotation.dataType), nullable = false, metadataBuilder.build)
     StructType(outputFields)
   }
-
 }

src/main/scala/com/johnsnowlabs/nlp/annotators/ner/crf/DictionaryFeatures.scala

@@ -2,8 +2,6 @@ package com.johnsnowlabs.nlp.annotators.ner.crf
 
 import com.johnsnowlabs.nlp.util.io.ResourceHelper
 
-import scala.io.Source
-
 case class DictionaryFeatures(dict: Map[String, String])
 {
  def get(tokens: Seq[String]): Seq[String] = {

src/main/scala/com/johnsnowlabs/nlp/annotators/ner/crf/FeatureGenerator.scala

@@ -1,7 +1,8 @@
 package com.johnsnowlabs.nlp.annotators.ner.crf
 
 import com.johnsnowlabs.ml.crf._
-import com.johnsnowlabs.nlp.annotators.common.{TaggedSentence, WordEmbeddings}
+import com.johnsnowlabs.nlp.annotators.common.TaggedSentence
+import com.johnsnowlabs.nlp.embeddings.WordEmbeddings
 
 import scala.collection.mutable
 

src/main/scala/com/johnsnowlabs/nlp/annotators/ner/crf/NerCrfApproach.scala

@@ -9,6 +9,7 @@ import com.johnsnowlabs.nlp.annotators.common.NerTagged
 import com.johnsnowlabs.nlp.annotators.pos.perceptron.PerceptronApproach
 import com.johnsnowlabs.nlp.annotators.sbd.pragmatic.SentenceDetectorModel
 import com.johnsnowlabs.nlp.datasets.CoNLL
+import com.johnsnowlabs.nlp.embeddings.AnnotatorWithWordEmbeddings
 import org.apache.spark.ml.Pipeline
 import org.apache.spark.ml.param.{DoubleParam, IntParam, Param, StringArrayParam}
 import org.apache.spark.ml.util.{DefaultParamsReadable, Identifiable}
@@ -17,7 +18,9 @@ import org.apache.spark.sql.{DataFrame, Dataset}
 /*
   Algorithm for training Named Entity Recognition Model.
    */
-class NerCrfApproach(override val uid: String) extends AnnotatorApproach[NerCrfModel]{
+class NerCrfApproach(override val uid: String)
+  extends AnnotatorWithWordEmbeddings[NerCrfApproach, NerCrfModel] {
+
   def this() = this(Identifiable.randomUID("NER"))
 
   override val description = "CRF based Named Entity Recognition Tagger"
@@ -116,7 +119,8 @@ class NerCrfApproach(override val uid: String) extends AnnotatorApproach[NerCrfM
 
     val dictPaths = get(dicts).getOrElse(Array.empty[String])
     val dictFeatures = DictionaryFeatures.read(dictPaths.toSeq)
-    val crfDataset = FeatureGenerator(dictFeatures).generateDataset(trainDataset)
+    val crfDataset = FeatureGenerator(dictFeatures, embeddings)
+      .generateDataset(trainDataset)
 
     val params = CrfParams(
       minEpochs = getOrDefault(minEpochs),

src/main/scala/com/johnsnowlabs/nlp/annotators/ner/crf/NerCrfModel.scala

@@ -4,6 +4,7 @@ import com.johnsnowlabs.ml.crf.{LinearChainCrfModel, SerializedLinearChainCrfMod
 import com.johnsnowlabs.nlp.AnnotatorType._
 import com.johnsnowlabs.nlp.annotators.common.{IndexedTaggedWord, NerTagged, PosTagged, TaggedSentence}
 import com.johnsnowlabs.nlp.annotators.common.Annotated.{NerTaggedSentence, PosTaggedSentence}
+import com.johnsnowlabs.nlp.embeddings.ModelWithWordEmbeddings
 import com.johnsnowlabs.nlp.{Annotation, AnnotatorModel}
 import org.apache.hadoop.fs.Path
 import org.apache.spark.ml.param.StringArrayParam
@@ -14,8 +15,7 @@ import org.apache.spark.sql.{Encoders, Row}
 /*
   Named Entity Recognition model
  */
-class NerCrfModel(override val uid: String)
-  extends AnnotatorModel[NerCrfModel] {
+class NerCrfModel(override val uid: String) extends ModelWithWordEmbeddings[NerCrfModel]{
 
   def this() = this(Identifiable.randomUID("NER"))
 
@@ -36,7 +36,7 @@ class NerCrfModel(override val uid: String)
   def setEntities(toExtract: Array[String]): NerCrfModel = set(entities, toExtract)
 
   /**
-    Predicts Named Entities in input sentences
+  Predicts Named Entities in input sentences
     * @param sentences POS tagged sentences.
     * @return sentences with recognized Named Entities
     */
@@ -45,8 +45,9 @@ class NerCrfModel(override val uid: String)
 
     val crf = model.get
 
+    val fg = FeatureGenerator(dictionaryFeatures, embeddings)
     sentences.map{sentence =>
-      val instance = FeatureGenerator(dictionaryFeatures).generate(sentence, crf.metadata)
+      val instance = fg.generate(sentence, crf.metadata)
       val labelIds = crf.predict(instance)
       val words = sentence.indexedTaggedWords
         .zip(labelIds.labels)
@@ -116,6 +117,9 @@ object NerCrfModel extends DefaultParamsReadable[NerCrfModel] {
       instance
         .setModel(crfModel.deserialize)
         .setDictionaryFeatures(dictFeatures)
+
+      instance.deserializeEmbeddings(path, sparkSession.sparkContext)
+      instance
     }
   }
 
@@ -136,7 +140,8 @@ object NerCrfModel extends DefaultParamsReadable[NerCrfModel] {
       val dictPath = new Path(path, "dict").toString
       val dictLines = model.dictionaryFeatures.dict.toSeq.map(p => p._1 + ":" + p._2)
       Seq(dictLines).toDS.write.mode("overwrite").parquet(dictPath)
+
+      model.serializeEmbeddings(path, sparkSession.sparkContext)
     }
   }
 }
-

src/main/scala/com/johnsnowlabs/nlp/datasets/CoNLL.scala

@@ -74,7 +74,7 @@ case class CoNLL(targetColumn: Int = 3, annotatorType: String) {
 
   def pack(sentences: Seq[TaggedSentence]): Seq[Annotation] = {
     if (annotatorType == AnnotatorType.NAMED_ENTITY)
-        NerTagged.pack(sentences)
+      NerTagged.pack(sentences)
     else
       PosTagged.pack(sentences)
   }
@@ -99,4 +99,4 @@ case class CoNLL(targetColumn: Int = 3, annotatorType: String) {
     val seq = readLines(lines).map(p => (p._1, pack(p._2)))
     seq.toDF(textColumn, labelColumn)
   }
-}
+}

src/main/scala/com/johnsnowlabs/nlp/datasets/CoNLL2003NerReader.scala

@@ -0,0 +1,75 @@
+package com.johnsnowlabs.nlp.datasets
+
+import java.io.File
+
+import com.johnsnowlabs.ml.crf.{CrfDataset, DatasetMetadata, InstanceLabels, TextSentenceLabels}
+import com.johnsnowlabs.nlp.AnnotatorType
+import com.johnsnowlabs.nlp.annotators.common.TaggedSentence
+import com.johnsnowlabs.nlp.annotators.ner.crf.{DictionaryFeatures, FeatureGenerator}
+import com.johnsnowlabs.nlp.embeddings.{WordEmbeddings, WordEmbeddingsFormat, WordEmbeddingsIndexer}
+
+/**
+  * Helper class for to work with CoNLL 2003 dataset for NER task
+  * Class is made for easy use from Java
+  */
+class CoNLL2003NerReader(wordEmbeddingsFile: String,
+                         wordEmbeddingsNDims: Int,
+                         embeddingsFormat: WordEmbeddingsFormat.Format,
+                         dictionaryFile: String) {
+
+  private val nerReader = CoNLL(3, AnnotatorType.NAMED_ENTITY)
+  private val posReader = CoNLL(1, AnnotatorType.POS)
+
+  private var wordEmbeddings: Option[WordEmbeddings] = None
+
+  if (wordEmbeddingsFile != null) {
+    require(new File(wordEmbeddingsFile).exists())
+
+    var fileDb = wordEmbeddingsFile + ".db"
+
+    if (!new File(fileDb).exists()) {
+      embeddingsFormat match {
+        case WordEmbeddingsFormat.Text =>
+          WordEmbeddingsIndexer.indexText(wordEmbeddingsFile, fileDb)
+        case WordEmbeddingsFormat.Binary =>
+          WordEmbeddingsIndexer.indexBinary(wordEmbeddingsFile, fileDb)
+        case WordEmbeddingsFormat.SparkNlp =>
+          fileDb = wordEmbeddingsFile
+      }
+
+    }
+
+    if (new File(fileDb).exists()) {
+      wordEmbeddings = Some(WordEmbeddings(fileDb, wordEmbeddingsNDims))
+    }
+  }
+
+  private val dicts = if (dictionaryFile == null) Seq.empty[String] else Seq(dictionaryFile)
+
+  private val fg = FeatureGenerator(
+    DictionaryFeatures.read(dicts),
+    wordEmbeddings
+  )
+
+  private def readDataset(file: String): Seq[(TextSentenceLabels, TaggedSentence)] = {
+    val labels = nerReader.readDocs(file).flatMap(_._2)
+      .map(sentence => TextSentenceLabels(sentence.tags))
+
+    val posTaggedSentences = posReader.readDocs(file).flatMap(_._2)
+    labels.zip(posTaggedSentences)
+  }
+
+  def readNerDataset(file: String, metadata: Option[DatasetMetadata] = None): CrfDataset = {
+    val lines = readDataset(file)
+    if (metadata.isEmpty)
+      fg.generateDataset(lines)
+    else {
+      val labeledInstances = lines.map { line =>
+        val instance = fg.generate(line._2, metadata.get)
+        val labels = InstanceLabels(line._1.labels.map(l => metadata.get.label2Id.getOrElse(l, -1)))
+        (labels, instance)
+      }
+      CrfDataset(labeledInstances, metadata.get)
+    }
+  }
+}