package chunker
- Alphabetic
- Public
- All
Type Members
-
class
AssertionFilterer extends AnnotatorModel[AssertionFilterer] with HasSimpleAnnotate[AssertionFilterer] with FilteringParams with CheckLicense
Filters entities coming from ASSERTION type annotations and returns the CHUNKS.
Filters entities coming from ASSERTION type annotations and returns the CHUNKS. Filters can be set via a white list and black list on the extracted chunk, the assertion or a regular expression. White and black lists are for assertion are enabled by default. To use chunk white list,
criteria
has to be set to"isin"
. For regex,criteria
has to be set to"regex"
.Example
To see how the assertions are extracted, see the example for AssertionDLModel.
Define an extra step where the assertions are filtered
val assertionFilterer = new AssertionFilterer() .setInputCols("sentence","ner_chunk","assertion") .setOutputCol("filtered") .setCriteria("assertion") .setWhiteList("present") val assertionPipeline = new Pipeline().setStages(Array( documentAssembler, sentenceDetector, tokenizer, embeddings, nerModel, nerConverter, clinicalAssertion, assertionFilterer )) val assertionModel = assertionPipeline.fit(data) val result = assertionModel.transform(data)
Show results:
result.selectExpr("ner_chunk.result", "assertion.result").show(3, truncate=false) +--------------------------------+--------------------------------+ |result |result | +--------------------------------+--------------------------------+ |[severe fever, sore throat] |[present, present] | |[stomach pain] |[absent] | |[an epidural, PCA, pain control]|[present, present, hypothetical]| +--------------------------------+--------------------------------+ result.select("filtered.result").show(3, truncate=false) +---------------------------+ |result | +---------------------------+ |[severe fever, sore throat]| |[] | |[an epidural, PCA] | +---------------------------+
- See also
AssertionDLModel to extract the assertions
-
class
ChunkConverter extends AnnotatorModel[ChunkConverter] with HasSimpleAnnotate[ChunkConverter] with SourceTrackingMetadataParams with CheckLicense
Convert chunks from regexMatcher to chunks with a entity in the metadata.
Convert chunks from regexMatcher to chunks with a entity in the metadata. Use the identifier or field as a entity.
Example
val sampleDataset = ResourceHelper.spark.createDataFrame(Seq( (1, "My first sentence with the first rule. This is my second sentence with ceremonies rule.") )).toDF("id", "text") val documentAssembler = new DocumentAssembler().setInputCol("text").setOutputCol("document") val sentence = new SentenceDetector().setInputCols("document").setOutputCol("sentence") val regexMatcher = new RegexMatcher() .setExternalRules(ExternalResource("src/test/resources/regex-matcher/rules.txt", ReadAs.TEXT, Map("delimiter" -> ","))) .setInputCols(Array("sentence")) .setOutputCol("regex") .setStrategy(strategy) val chunkConverter = new ChunkConverter().setInputCols("regex").setOutputCol("chunk") val pipeline = new Pipeline().setStages(Array(documentAssembler, sentence, regexMatcher,chunkConverter)) val results = pipeline.fit(sampleDataset).transform(sampleDataset) results.select("chunk").show(truncate = false) +------------------------------------------------------------------------------------------------+ |col | +------------------------------------------------------------------------------------------------+ |[chunk, 23, 31, the first, [identifier -> NAME, sentence -> 0, chunk -> 0, entity -> NAME], []] | |[chunk, 71, 80, ceremonies, [identifier -> NAME, sentence -> 1, chunk -> 0, entity -> NAME], []]| +------------------------------------------------------------------------------------------------+
-
class
ChunkFilterer extends AnnotatorModel[ChunkFilterer] with HasSimpleAnnotate[ChunkFilterer] with CheckLicense with FilteringParams
ChunkFilterer can filter chunks coming from CHUNK annotations.
ChunkFilterer can filter chunks coming from CHUNK annotations. Filters can be set via white list and black list or a regular expression. White list criteria is enabled by default. To use regex,
criteria
has to be set toregex
. Additionally, It can filter chunks according to the confidence of the chunk in the metadata.Example
Filtering POS tags
First pipeline stages to extract the POS tags are defined
val data = Seq("Has a past history of gastroenteritis and stomach pain, however patient ...").toDF("text") val docAssembler = new DocumentAssembler().setInputCol("text").setOutputCol("document") val sentenceDetector = new SentenceDetector().setInputCols("document").setOutputCol("sentence") val tokenizer = new Tokenizer().setInputCols("sentence").setOutputCol("token") val posTagger = PerceptronModel.pretrained() .setInputCols("sentence", "token") .setOutputCol("pos") val chunker = new Chunker() .setInputCols("pos", "sentence") .setOutputCol("chunk") .setRegexParsers(Array("(<NN>)+"))
Then the chunks can be filtered via a white list. Here only terms with "gastroenteritis" remain.
val chunkerFilter = new ChunkFilterer() .setInputCols("sentence","chunk") .setOutputCol("filtered") .setCriteria("isin") .setWhiteList("gastroenteritis") val pipeline = new Pipeline().setStages(Array( docAssembler, sentenceDetector, tokenizer, posTagger, chunker, chunkerFilter)) result.selectExpr("explode(chunk)").show(truncate=false) +---------------------------------------------------------------------------------+ |col | +---------------------------------------------------------------------------------+ |{chunk, 11, 17, history, {sentence -> 0, chunk -> 0}, []} | |{chunk, 22, 36, gastroenteritis, {sentence -> 0, chunk -> 1}, []} | |{chunk, 42, 53, stomach pain, {sentence -> 0, chunk -> 2}, []} | |{chunk, 64, 70, patient, {sentence -> 0, chunk -> 3}, []} | |{chunk, 81, 110, stomach pain now.We don't care, {sentence -> 0, chunk -> 4}, []}| |{chunk, 118, 132, gastroenteritis, {sentence -> 0, chunk -> 5}, []} | +---------------------------------------------------------------------------------+ result.selectExpr("explode(filtered)").show(truncate=false) +-------------------------------------------------------------------+ |col | +-------------------------------------------------------------------+ |{chunk, 22, 36, gastroenteritis, {sentence -> 0, chunk -> 1}, []} | |{chunk, 118, 132, gastroenteritis, {sentence -> 0, chunk -> 5}, []}| +-------------------------------------------------------------------+
-
class
ChunkFiltererApproach extends AnnotatorApproach[ChunkFilterer] with HasFeatures with FilteringParams with CheckLicense
Trains a ChunkFilterer annotator.
Trains a ChunkFilterer annotator. ChunkFiltererApproach can filter chunks coming from CHUNK annotations. Filters can be set via white list and black list or a regular expression. White list criteria is enabled by default. To use regex,
criteria
has to be set toregex
. Additionally, It can filter chunks according to the confidence of the chunk in the metadata.Example
Filtering POS tags
First pipeline stages to extract the POS tags are defined
val data = Seq("Has a past history of gastroenteritis and stomach pain, however patient ...").toDF("text") val docAssembler = new DocumentAssembler().setInputCol("text").setOutputCol("document") val sentenceDetector = new SentenceDetector().setInputCols("document").setOutputCol("sentence") val tokenizer = new Tokenizer().setInputCols("sentence").setOutputCol("token") val posTagger = PerceptronModel.pretrained() .setInputCols("sentence", "token") .setOutputCol("pos") val chunker = new Chunker() .setInputCols("pos", "sentence") .setOutputCol("chunk") .setRegexParsers(Array("(<NN>)+"))
Then the chunks can be filtered via a white list. Here only terms with "gastroenteritis" remain.
val chunkerFilter = new ChunkFiltererApproach() .setInputCols("sentence","chunk") .setOutputCol("filtered") .setCriteria("isin") .setWhiteList("gastroenteritis") val pipeline = new Pipeline().setStages(Array( docAssembler, sentenceDetector, tokenizer, posTagger, chunker, chunkerFilter)) result.selectExpr("explode(chunk)").show(truncate=false) +---------------------------------------------------------------------------------+ |col | +---------------------------------------------------------------------------------+ |{chunk, 11, 17, history, {sentence -> 0, chunk -> 0}, []} | |{chunk, 22, 36, gastroenteritis, {sentence -> 0, chunk -> 1}, []} | |{chunk, 42, 53, stomach pain, {sentence -> 0, chunk -> 2}, []} | |{chunk, 64, 70, patient, {sentence -> 0, chunk -> 3}, []} | |{chunk, 81, 110, stomach pain now.We don't care, {sentence -> 0, chunk -> 4}, []}| |{chunk, 118, 132, gastroenteritis, {sentence -> 0, chunk -> 5}, []} | +---------------------------------------------------------------------------------+ result.selectExpr("explode(filtered)").show(truncate=false) +-------------------------------------------------------------------+ |col | +-------------------------------------------------------------------+ |{chunk, 22, 36, gastroenteritis, {sentence -> 0, chunk -> 1}, []} | |{chunk, 118, 132, gastroenteritis, {sentence -> 0, chunk -> 5}, []}| +-------------------------------------------------------------------+
-
class
ChunkKeyPhraseExtraction extends BertSentenceEmbeddings with CheckLicense
Extracts key phrases from texts.
Extracts key phrases from texts.
ChunkKeyPhraseExtraction
usesBertSentenceEmbeddings
to determine the most relevant key phrases describing a text with the use of two approaches:- By using cosine similarities between the embedding representation of the chunks and the embedding representation of the corresponding sentences/documents.
- By using the Maximal Marginal Relevance (MMR) algorithm (set with the
setDivergence
method) to determine the most relevant key phrases. If theselectMostDifferent
parameter is set, return the key phrases that are the most different from each other (avoid too similar key phrases). The model compares the chunks against the corresponding sentences/documents and selects the chunks which are most representative of the broader text context (i.e., the document or the sentence they belong to). This allows, for example, to obtain a brief understanding of a document by selecting the most relevant phrases. The input to the model consists of chunk annotations and sentence or document annotation. The input chunks can be generated in various ways: - Using
NGramGenerator
, which allows to obtain ranked n-gram chunks from the text (can be used to identify new entities). - Using
YakeKeywordExtractor
, allowing to rank the keywords extracted using the YAKE algorithm. - Using
TextMatcher
, which allows to rank the desired chunks from the annotator. - Using
NerConverter
, which allows to extract ranked named entities (which entities are the most relevant in the sentence/document). The model operates either at sentence (selecting the most descriptive chunks from the sentence they belong to) or at document level. In the latter case, the key phrases are selected to represent all the input document annotations.
This model is a subclass of BertSentenceEmbeddings and shares all parameters with it. It can load any pretrained BertSentenceEmbeddings model. Available models can be found at Models Hub.
val embeddings = ChunkKeyPhraseExtraction.pretrained() .setInputCols("sentence", "chunk") .setOutputCol("key_phrase_chunks")
The default model is
"sbert_jsl_medium_uncased"
, if no name is provided.Sources :
The use of MMR, diversity-based reranking for reordering documents and producing summaries
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
Example
import spark.implicits._ import com.johnsnowlabs.nlp.base.DocumentAssembler import com.johnsnowlabs.nlp.annotator.SentenceDetector import com.johnsnowlabs.nlp.embeddings.BertSentenceEmbeddings import com.johnsnowlabs.nlp.EmbeddingsFinisher import org.apache.spark.ml.Pipeline val documentAssembler = new DocumentAssembler() .setInputCol("text") .setOutputCol("document") val tokenizer = new Tokenizer() .setInputCols("document") .setOutputCol("tokens") val stopWordsCleaner = StopWordsCleaner.pretrained() .setInputCols("tokens") .setOutputCol("clean_tokens") .setCaseSensitive(false) val nGrams = new NGramGenerator() .setInputCols(Array("clean_tokens")) .setOutputCol("ngrams") .setN(3) val chunkKeyPhraseExtractor = ChunkKeyPhraseExtraction .pretrained() .setTopN(2) .setDivergence(0.7f) .setInputCols(Array("document", "ngrams")) .setOutputCol("key_phrases") val pipeline = new Pipeline() .setStages(Array( documentAssembler, tokenizer, stopWordsCleaner, nGrams, chunkKeyPhraseExtractor)) val sampleText = "Her Diabetes has become type 2 in the last year with her Diabetes." + " He complains of swelling in his right forearm." val testDataset = Seq("").toDS.toDF("text") val result = pipeline.fit(emptyDataset).transform(testDataset) result .selectExpr("explode(key_phrases) AS key_phrase") .selectExpr( "key_phrase.result", "key_phrase.metadata.DocumentSimilarity", "key_phrase.metadata.MMRScore") .show(truncate=false) +--------------------------+-------------------+------------------+ |result |DocumentSimilarity |MMRScore | +--------------------------+-------------------+------------------+ |complains swelling forearm|0.6325718954229369 |0.1897715761677257| |type 2 year |0.40181028931546364|-0.189501077108947| +--------------------------+-------------------+------------------+
- See also
BertEmbeddings for token-level embeddings
BertSentenceEmbeddings for sentence-level embeddings
Annotators Main Page for a list of transformer based embeddings
- class ChunkMapperApproach extends AnnotatorApproach[ChunkMapperModel] with CheckLicense with ChunkMapperFuzzyMatchingParams
- class ChunkMapperFilterer extends AnnotatorModel[ChunkMapperFilterer] with HasSimpleAnnotate[ChunkMapperFilterer] with CheckLicense
- class ChunkMapperModel extends AnnotatorModel[ChunkMapperModel] with HasSimpleAnnotate[ChunkMapperModel] with CheckLicense with ChunkMapperFuzzyMatchingParams
-
class
ChunkSentenceSplitter extends AnnotatorModel[ChunkSentenceSplitter] with HasSimpleAnnotate[ChunkSentenceSplitter] with CheckLicense
Split the document using the chunks that you provided,and put in the metadata the chunk entity.
Split the document using the chunks that you provided,and put in the metadata the chunk entity. The first piece of document to the first chunk will have the entity as header.
Is useful if you identify the titles and subtitles using some ner and after that you can split by paragraph
Example
val data = Seq(text,text).toDS.toDF("text") val documentAssembler = new DocumentAssembler().setInputCol("text").setOutputCol("doc") val regexMatcher = new RegexMatcher().setInputCols("doc").setOutputCol("chunks").setExternalRules("src/test/resources/chunker/title_regex.txt",",") val chunkSentenceSplitter = new ChunkSentenceSplitter().setInputCols("chunks","doc").setOutputCol("paragraphs") val pipeline = new Pipeline().setStages(Array(documentAssembler,regexMatcher,chunkSentenceSplitter)) val result = pipeline.fit(data).transform(data).select("paragraphs") result.show(truncate = false)
- class DocMapperApproach extends ChunkMapperApproach
- class DocMapperModel extends ChunkMapperModel
- trait ReadChunkKeyPhraseExtractionTensorflowModel extends ReadTensorflowModel
- trait ReadablePretrainedChunkKeyPhraseExtractionModel extends ParamsAndFeaturesReadable[ChunkKeyPhraseExtraction] with HasPretrained[ChunkKeyPhraseExtraction]
- trait ReadablePretrainedChunkMapperModel extends ParamsAndFeaturesReadable[ChunkMapperModel] with HasPretrained[ChunkMapperModel]
- trait ReadablePretrainedDocMapperModel extends ParamsAndFeaturesReadable[DocMapperModel] with HasPretrained[DocMapperModel]
Value Members
- object AssertionFilterer extends ParamsAndFeaturesReadable[AssertionFilterer] with Serializable
- object ChunkFilterer extends ParamsAndFeaturesReadable[ChunkFilterer] with Serializable
- object ChunkFiltererApproach extends DefaultParamsReadable[ChunkFiltererApproach] with Serializable
- object ChunkKeyPhraseExtraction extends ReadablePretrainedChunkKeyPhraseExtractionModel with ReadChunkKeyPhraseExtractionTensorflowModel with Serializable
- object ChunkMapperFilterer extends ParamsAndFeaturesReadable[ChunkMapperFilterer] with Serializable
- object ChunkMapperModel extends ParamsAndFeaturesReadable[ChunkMapperModel] with ReadablePretrainedChunkMapperModel with Serializable
- object ChunkSentenceSplitter extends ParamsAndFeaturesReadable[ChunkSentenceSplitter] with Serializable
- object DocMapperModel extends ParamsAndFeaturesReadable[DocMapperModel] with ReadablePretrainedDocMapperModel with Serializable