Skip to content
This repository has been archived by the owner. It is now read-only.


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

GraphAware Natural Language Processing Has Been Retired

As of May 2021, this repository has been retired.

GraphAware Natural Language Processing

This Neo4j plugin offers Graph Based Natural Language Processing capabilities.

The main module, this module, provide a common interface for underlying text processors as well as a Domain Specific Language built atop stored procedures and functions making your Natural Language Processing workflow developer friendly.

It comes in 2 versions, Community (open-sourced) and Enterprise with the following NLP features :

Feature Matrix

Community Edition Enterprise Edition
Text information Extraction
Multi-languages in the same database
Custom NamedEntityRecognition model builder
ConceptNet5 Enricher
Microsoft Concept Enricher
Keyword Extraction
TextRank Summarization
Topics Extraction
Word Embeddings (Word2Vec)
Similarity Computation
PDF Parsing
Apache Spark Binding for Distributed Algorithms
Doc2Vec implementation
User Interface
ML Prediction capabilities
Entity Merging

Two NLP processor implementations are available, respectively Stanford NLP and OpenNLP (OpenNLP receives less frequent updates, StanfordNLP is recommended).


From version you need to download the language models, see below

From the GraphAware plugins directory, download the following jar files :

  • neo4j-framework (the JAR for this is labeled "graphaware-server-enterprise-all")
  • neo4j-nlp
  • neo4j-nlp-stanfordnlp
  • The language model to be downloaded from

and copy them in the plugins directory of Neo4j.

Take care that the version numbers of the framework you are using match with the version of Neo4J you are using. This is a common setup problem. For example, if you are using Neo4j 3.4.0 and above, all of the JARs you download should contain 3.4 in their version number.

plugins/ directory example :

-rw-r--r--  1 ikwattro  staff    58M Oct 11 11:15 graphaware-nlp-
-rw-r--r--@ 1 ikwattro  staff    13M Aug 22 15:22 graphaware-server-community-all-
-rw-r--r--  1 ikwattro  staff    16M Oct 11 11:28 nlp-stanfordnlp-
-rw-r--r--@ 1 ikwattro  staff   991M Oct 11 11:45 stanford-english-corenlp-2018-10-05-models.jar

Append the following configuration in the neo4j.conf file in the config/ directory:


Start or restart your Neo4j database.

Note: both concrete text processors are quite greedy - you will need to dedicate sufficient memory for to Neo4j heap space.

Additionally, the following indexes and constraints are suggested to speed performance:

CREATE INDEX ON :Tag(value);

Or use the dedicated procedure :

CALL ga.nlp.createSchema()

Define which language you will use in this database :

CALL ga.nlp.config.setDefaultLanguage('en')

Quick Documentation in Neo4j Browser

Once the extension is loaded, you can see basic documentation on all available procedures by running this Cypher query:

CALL dbms.procedures() YIELD name, signature, description
WHERE name =~ 'ga.nlp.*'
RETURN name, signature, description ORDER BY name asc;

Getting Started

Text extraction

Pipelines and components

The text extraction phase is done with a Natural Language Processing pipeline, each pipeline has a list of enabled components.

For example, the basic tokenizer pipeline has the following components :

  • Sentence Segmentation
  • Tokenization
  • StopWords Removal
  • Stemming
  • Part Of Speech Tagging
  • Named Entity Recognition

It is mandatory to create your pipeline first :

CALL ga.nlp.processor.addPipeline({textProcessor: 'com.graphaware.nlp.processor.stanford.StanfordTextProcessor', name: 'customStopWords', processingSteps: {tokenize: true, ner: true, dependency: false}, stopWords: '+,result, all, during', 
threadNumber: 20})

The available optional parameters (default values are in brackets):

  • name: desired name of a new pipeline
  • textProcessor: to which text processor should the new pipeline be added
  • processingSteps: pipeline configuration (available in both Stanford and OpenNLP unless stated otherwise)
    • tokenize (default: true): perform tokenization
    • ner (default: true): Named Entity Recognition
    • sentiment (default: false): run sentiment analysis on sentences
    • coref (default: false): Coreference Resolution (identify multiple mentions of the same entity, such as "Barack Obama" and "he")
    • relations (default: false): run relations identification between two tokens
    • dependency (default: false, StanfordNLP only): extract typed dependencies (ex.: amod - adjective modifier, conj - conjunct, ...)
    • cleanxml (default: false, StanfordNLP only): remove XML tags
    • truecase (default: false, StanfordNLP only): recognizes the "true" case of tokens (how they would be capitalized in well-edited text)
    • customNER: list of custom NER model identifiers (as a string, model identifiers separated by “,”)
  • stopWords: specify words that are required to be ignored (if the list starts with +, the following words are appended to the default stopwords list, otherwise the default list is overwritten)
  • threadNumber (default: 4): for multi-threading
  • excludedNER: (default: none) specify a list of NE to not be recognized in upper case, for example for excluding NER_Money and NER_O on the Tag nodes, use ['O', 'MONEY']

To set a pipeline as a default pipeline:

CALL ga.nlp.processor.pipeline.default(<your-pipeline-name>)

To delete a pipeline, use this command:

CALL ga.nlp.processor.removePipeline(<pipeline-name>, <text-processor>)

To see details of all existing pipelines:

CALL ga.nlp.processor.getPipelines()


Let's take the following text as example :

Scores of people were already lying dead or injured inside a crowded Orlando nightclub,
and the police had spent hours trying to connect with the gunman and end the situation without further violence.
But when Omar Mateen threatened to set off explosives, the police decided to act, and pushed their way through a
wall to end the bloody standoff.

Simulate your original corpus

Create a node with the text, this node will represent your original corpus or knowledge graph :

CREATE (n:News)
SET n.text = "Scores of people were already lying dead or injured inside a crowded Orlando nightclub,
and the police had spent hours trying to connect with the gunman and end the situation without further violence.
But when Omar Mateen threatened to set off explosives, the police decided to act, and pushed their way through a
wall to end the bloody standoff.";

Perform the text information extraction

The extraction is done via the annotate procedure which is the entry point to text information extraction

MATCH (n:News)
CALL ga.nlp.annotate({text: n.text, id: id(n)})
YIELD result
RETURN result

Available parameters of annotate procedure:

  • text: text to annotate represented as a string
  • id: specify ID that will be used as id property of the new AnnotatedText node
  • textProcessor (default: "Stanford", if not available than the first entry in the list of available text processors)
  • pipeline (default: tokenizer)
  • checkLanguage (default: true): run language detection on provided text and check whether it's supported

This procedure will link your original :News node to an :AnnotatedText node which is the entry point for the graph based NLP of this particular News. The original text is broken down into words, parts of speech, and functions. This analysis of the text acts as a starting point for the later steps.

annotated text

Running a batch of annotations

If you have a big set of data to annotate, we recommend to use APOC :

CALL apoc.periodic.iterate(
"MATCH (n:News) RETURN n",
"CALL ga.nlp.annotate({text: n.text, id: id(n)})
YIELD result MERGE (n)-[:HAS_ANNOTATED_TEXT]->(result)", {batchSize:1, iterateList:true})

It is important to keep the batchSize and iterateList options as mentioned in the example. Running the annotation procedure in parallel will create deadlocks.

Enrich your original knowledge

We implement external knowledge bases in order to enrich the knowledge of your current data.

As of now, two implementations are available :

  • ConceptNet5
  • Microsoft Concept Graph

This enricher will extend the meaning of tokens (Tag nodes) in the graph.

MATCH (n:Tag)
CALL ga.nlp.enrich.concept({enricher: 'conceptnet5', tag: n, depth:1, admittedRelationships:["IsA","PartOf"]})
YIELD result
RETURN result

The available parameters (default values are in brackets):

  • tag: tag to be enriched
  • enricher ("conceptnet5"): choose microsoft or conceptnet5
  • depth (2): how deep to go in concept hierarchy
  • admittedRelationships: choose desired concept relationships types, please refer to the ConceptNet Documentation for details
  • pipeline: choose pipeline name to be used for cleansing of concepts before storing them to your DB; your system default pipeline is used otherwise
  • filterByLanguage (true): allow only concepts of languages specified in outputLanguages; if no languages are specified, the same language as tag is required
  • outputLanguages ([]): return only concepts with specified languages
  • relDirection ("out"): desired direction of relationships in concept hierarchy ("in", "out", "both")
  • minWeight (0.0): minimal admitted concept relationship weight
  • limit (10): maximal number of concepts per tag
  • splitTag (false): if true, tag is first tokenised and then individual tokens enriched

Tags have now a IS_RELATED_TO relationships to other enriched concepts.

annotated text

List of available procedures

Keyword Extraction

MATCH (a:AnnotatedText)
CALL{annotatedText: a, stopwords: '+,other,email', useDependencies: true})
YIELD result RETURN result

annotatedText is a mandatory parameter which refers to the annotated document that is required to be analyzed.

Available optional parameters (default values are in brackets):

  • keywordLabel (Keyword): label name of the keyword nodes
  • useDependencies (true): use universal dependencies to enrich extracted keywords and key phrases by tags related through COMPOUND and AMOD relationships
  • dependenciesGraph (false): use universal dependencies for creating tag co-occurrence graph (default is false, which means that a natural word flow is used for building co-occurrences)
  • cleanKeywords (true): run cleaning procedure
  • topXTags (1/3): set a fraction of highest-rated tags that will be used as keywords / key phrases
  • respectSentences (false): respect or not sentence boundaries for co-occurrence graph building
  • respectDirections (false): respect or not directions in co-occurrence graph (how the words follow each other)
  • iterations (30): number of PageRank iterations
  • damp (0.85): PageRank damping factor
  • threshold (0.0001): PageRank convergence threshold
  • removeStopwords (true): use a stopwords list for co-occurrence graph building and final cleaning of keywords
  • stopwords: customize stopwords list (if the list starts with +, the following words are appended to the default stopwords list, otherwise the default list is overwritten)
  • admittedPOSs: specify which POS labels are considered as keyword candidates; needed when using different language than English
  • forbiddenPOSs: specify list of POS labels to be ignored when constructing co-occurrence graph; needed when using different language than English
  • forbiddenNEs: specify list of NEs to be ignored

For a detailed TextRank algorithm description, please refer to our blog post about Unsupervised Keyword Extraction.

Using universal dependencies for keyword enrichment (useDependencies option) can result in keywords with unnecessary level of detail, for example a keyword space shuttle logistics program. In many use cases we might be interested to also know that given document speaks generally about space shuttle (or logistic program). To do that, run post-processing with one of these options:

  • direct - each key phrase of n number of tags is checked against all key phrases from all documents with 1 < m < n number of tags; if the former contains the latter key phrase, then a DESCRIBES relationship is created from the m-keyphrase to all annotated texts of the n-keyphrase
  • subgroups - the same procedure as for direct, but instead of connecting higher level keywords directly to AnnotatedTexts, they are connected to the lower level keywords with HAS_SUBGROUP relationships
// Important note: create subsequent indices to optimise the post-process method performance
CREATE INDEX ON :Keyword(numTerms)
CREATE INDEX ON :Keyword(value)

CALL{keywordLabel: "Keyword", method: "subgroups"})
YIELD result
RETURN result

keywordLabel is an optional argument set by default to "Keyword".

The postprocess operation by default is processing on all keywords, which can be very heavy on large graphs. You can specify the annotatedText on which to apply the postprocess operation with the annotatedText argument :

MATCH (n:AnnotatedText) WITH n LIMIT 100
CALL{annotatedText: n, method:'subgroups'}) YIELD result RETURN count(n)

Example for running it efficiently on the full set of Keywords with APOC :

CALL apoc.periodic.iterate(
'MATCH (n:AnnotatedText) RETURN n',
'CALL{annotatedText: n, method:"subgroups"}) YIELD result RETURN count(n)',
{batchSize: 1, iterateList:false}

TextRank Summarization

Similar approach to the keyword extraction can be employed to implement simple summarization. A densely connect graph of sentences is created, with Sentence-Sentence relationships representing their similarity based on shared words (number of shared words vs sum of logarithms of number of words in a sentence). PageRank is then used as a centrality measure to rank the relative importance of sentences in the document.

To run this algorithm:

MATCH (a:AnnotatedText)
CALL{annotatedText: a}) YIELD result
RETURN result

Available parameters:

  • annotatedText
  • iterations (30): number of PageRank iterations
  • damp (0.85): PageRank damping factor
  • threshold (0.0001): PageRank convergence threshold

The summarisation procedure saves new properties to Sentence nodes: summaryRelevance (PageRank value of given sentence) and summaryRank (ranking; 1 = highest ranked sentence). Example query for retrieving summary:

match (n:Kapitel)-[:HAS_ANNOTATED_TEXT]->(a:AnnotatedText)
where id(n) = 233
match (a)-[:CONTAINS_SENTENCE]->(s:Sentence)
with a, count(*) as nSentences
match (a)-[:CONTAINS_SENTENCE]->(s:Sentence)-[:HAS_TAG]->(t:Tag)
with a, s, count(distinct t) as nTags, (CASE WHEN nSentences*0.1 > 10 THEN 10 ELSE toInteger(nSentences*0.1) END) as nLimit
where nTags > 4
with a, s, nLimit
order by s.summaryRank
with a, collect({text: s.text, pos: s.sentenceNumber})[..nLimit] as summary
unwind summary as sent
return sent.text
order by sent.pos

Sentiment Detection

You can also determine whether the text presented is positive, negative, or neutral. This procedure requires an AnnotatedText node, which is produced by ga.nlp.annotate above.

MATCH (t:MyNode)-[]-(a:AnnotatedText) 
CALL ga.nlp.sentiment(a) YIELD result 
RETURN result;

This procedure will simply return "SUCCESS" when it is successful, but it will apply the :POSITIVE, :NEUTRAL or :NEGATIVE label to each Sentence. As a result, when sentiment detection is complete, you can query for the sentiment of sentences as such:

MATCH (s:Sentence)
RETURN s.text, labels(s)

Language Detection

CALL ga.nlp.detectLanguage("What language is this in?") 
YIELD result return result

NLP based filter

CALL ga.nlp.filter({text:'On 8 May 2013,
    one week before the Pakistani election, the third author,
    in his keynote address at the Sentiment Analysis Symposium, 
    forecast the winner of the Pakistani election. The chart
    in Figure 1 shows varying sentiment on the candidates for 
    prime minister of Pakistan in that election. The next day, 
    the BBC’s Owen Bennett Jones, reporting from Islamabad, wrote 
    an article titled Pakistan Elections: Five Reasons Why the 
    Vote is Unpredictable, in which he claimed that the election 
    was too close to call. It was not, and despite his being in Pakistan, 
    the outcome of the election was exactly as we predicted.', filter: 'Owen Bennett Jones/PERSON, BBC, Pakistan/LOCATION'}) YIELD result 
return result

Cosine similarity computation

Once tags are extracted from all the news or other nodes containing some text, it is possible to compute similarities between them using content based similarity. During this process, each annotated text is described using the TF-IDF encoding format. TF-IDF is an established technique from the field of information retrieval and stands for Term Frequency-Inverse Document Frequency. Text documents can be TF-IDF encoded as vectors in a multidimensional Euclidean space. The space dimensions correspond to the tags, previously extracted from the documents. The coordinates of a given document in each dimension (i.e., for each tag) are calculated as a product of two sub-measures: term frequency and inverse document frequency.

MATCH (a:AnnotatedText) 
//WHERE ...
WITH collect(a) as nodes
CALL{input: <list_of_annotated_texts>[, query: <tfidf_query>, relationshipType: "CUSTOM_SIMILARITY", ...]}) YIELD result
RETURN result

Available parameters (default values are in brackets):

  • input: list of input nodes - AnnotatedTexts
  • relationshipType (SIMILARITY_COSINE): type of similarity relationship, use it along with query
  • query: specify your own query for extracting tf and idf in form ... RETURN id(Tag), tf, idf
  • propertyName (value): name of an existing node property (array of numerical values) which contains already prepared document vector


Word2vec is a shallow two-layer neural network model used to produce word embeddings (words represented as multidimensional semantic vectors) and it is one of the models used in ConceptNet Numberbatch.

To add source model (vectors) into a Lucene index

CALL<path_to_source_dir>, <path_to_index>, <identifier>)
  • <path_to_source_dir> is a full path to the directory with source vectors to be indexed
  • <path_to_index> is a full path where the index will be stored
  • <identifier> is a custom string that uniquely identifies the model

To list available models:


The model can now be used to compute cosine similarities between words:

WITH'äpple', 'swedish-numberbatch') AS appleVector,'frukt', 'swedish-numberbatch') AS fruitVector
RETURN, fruitVector) AS similarity
  • 1st parameter: word
  • 2nd parameter: model identifier

Or you can ask directly for a word2vec of a node which has a word stored in property value:

MATCH (n1:Tag), (n2:Tag)
WITH, <model_name>) AS vector1,, <model_name>) AS vector2
RETURN, vector2) AS similarity

We can also permanently store the word2vec vectors to Tag nodes:

CALL{query:'MATCH (t:Tag) RETURN t', modelName:'swedish-numberbatch'})
  • query: query which returns tags to which embedding vectors should be attached
  • modelName: model to use

You can also get the nearest neighbors with the following procedure :

CALL'analyzed', 10, 'fasttext') YIELD word, distance RETURN word, distance

For large models, for example full fasttext for english, approximately 2 million words, it will be inefficient to compute the nearest neighbors on the fly.

You can load the model into memory in order to have faster nearest neighbors ( fasttext 1M word vectors generally takes 27 seconds if needed to read from disk, ~300ms in memory) :

Make sure to have efficient heap memory dedicated to Neo4j :


Load the model into memory :

CALL<maxNeighbors>, <modelName>)

And retrieve it with


Using other models

You can use any word embedding model as long as the following is true :

  • Every line contain the word + the vector
  • The file has a .txt extension

For example, you can load the models from fasttext and just rename the file from .vec to .txt :

Parsing PDF Documents

CALL ga.nlp.parser.pdf("file:///Users/ikwattro/_graphs/nlp/import/myfile.pdf") YIELD number, paragraphs

The procedure return rows with columns number being the page number and paragraphs being a List<String> of paragraph texts.

You can also pass an http or https url to the procedure for loading a file from a remote location.

Exclude content from the pdf

In some cases, pdf documents have some recurrent useless content like page footers etc, you can excluded them from the parsing by passing a list of regexes defining the parts to exclude :

CALL ga.nlp.parser.pdf("myfile.pdf", ["^[0-9]$","^Licensed to"])

Use a different user Agent than TIKA

TIKA can be recognized as crawler and be denied access to some sites containing pdf's. You can override this by passing a UserAgent option :

CALL ga.nlp.parser.pdf($url, [], {UserAgent: 'Mozilla/5.0 (Windows; U; Win98; en-US; rv:1.7.2) Gecko/20040803'})


Parsing raw content from a file

RETURN ga.nlp.parse.raw(<path-to-file>) AS content

Storing only certain Tag/Tokens

In certain situations, it would be useful to store only certain values instead of the full graph, note though that it might reduce the ability to extract insights ( textRank ) for eg :

CALL ga.nlp.processor.addPipeline({
processingSteps:{tokenize:true, ner:true}})
CALL ga.nlp.annotate({text:"Hello, my name is John and I worked at IBM.", id:"test-123", pipeline:"whitelist", checkLanguage:false})
YIELD result
RETURN result

Parsing WebVTT

WebVTT is the format for Web Video Text Tracks, such as Youtube Transcripts of videos :

CALL ga.nlp.parser.webvtt("url-to-transcript.vtt") YIELD startTime, endTime, text

Listing files from directory(ies)

CALL ga.nlp.utils.listFiles(<path-to-directory>, <extensionFilter>)

// eg:

CALL ga.nlp.utils.listFiles("/Users/ikwattro/dev/papers", ".pdf") YIELD filePath RETURN filePath

The above procedure list files of the current directory only, if you need to walk the children directories as well, use walkdir :

CALL ga.nlp.utils.walkdir("/Users/ikwattro/dev/papers", ".pdf") YIELD filePath RETURN filePath

Additional Procedures


List stored models and their paths


Remove and re-create a pipeline with the same configuration ( useful when using static ner files that have been changed for eg )


Copyright (c) 2013-2019 GraphAware

GraphAware is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program. If not, see