By David Talby, CTO Usermind.
Apache Spark is a general-purpose cluster computing framework, with native support for distributed SQL, streaming, graph processing, and machine learning. Now, the Spark ecosystem also has an Spark Natural Language Processing library.
Get it on GitHub or begin with the quickstart tutorial.
The John Snow Labs NLP Library is under the Apache 2.0 license, written in Scala with no dependencies on other NLP or ML libraries. It natively extends the Spark ML Pipeline API. You will benefit from:
- Unmatched runtime performance, since processing is done directly on Spark DataFrames without any copying and taking full advantage of Spark’s caching, execution planning and optimized binary __data science pipeline.
“With John Snow Labs NLP, we’re delivering on the promise to enable customers to take advantage of the latest open source technology and academic breakthroughs in data science, all within a high performance, enterprise-grade code base.”, said the founding team. In addition, “John Snow Labs NLP encompasses a wide range of highly efficient Natural Language Understanding tools for text mining, question answering, chat bots, fact extraction, topic modelling or Search, running at a scale and performance that has not been available to date.”
The framework provides the concepts of annotators, and comes out of the box with:
- Entity Extractor
- Date Extractor
- Part of Speech Tagger
- Named Entity Recognition
- Sentence boundary detection
- Sentiment analysis
- Spell checker
In addition, given the tight integration with Spark ML, there is a lot more you can use right away when building your NLP pipelines. This includes word embeddings, topic modeling, stop word removal, a variety of feature engineering functions (tf-idf, n-grams, similarity metrics, …) and using NLP annotations as features in machine learning workflows. If you’re not familiar with these terms, this guide to understanding NLP tasks is a good start.
Our virtual team has been building commercial software that heavily depends on natural language understanding for several years now. As such, we have hands-on experience with spaCy, CoreNLP, OpenNLP, Mallet, GATE, Weka, UIMA, nltk, gensim, Negex, word2vec, GloVe, and a few others.
We are big fans, and the many places where we’ve imitated these libraries are intended as the sincere form of flattery that they are. But we’ve also banged our heads too many times against their limitations – when we’ve had to deliver scalable, high-performance, high-accuracy software for real production use.
The first of three top-level requirements we tackled is runtime performance. You’d think this was largely a solved problem with the advent of spaCy and its public benchmarks which reflect a well thought-out and masterfully implemented set of tradeoffs. However, when building Spark applications on top of it, you’d still get unreasonably subpar throughput.
To understand why, consider that an NLP pipeline is always just a part of a bigger data processing pipeline: For example, question answering involves loading training, data, transforming it, applying NLP annotators, building features, training the value extraction models, evaluating the results (train/test split or cross-validation), and hyper-parameter estimation.
Splitting your data processing framework (Spark) from your NLP frameworks means that most of your processing time gets spent serializing and copying strings.
A great parallel is TensorFrames – which greatly improves the performance of running TensorFlow workflows on Spark data frames. This image is credited to Tim Hunter’s excellent TensorFrames overview:
Both Spark and TensorFlow are optimized to the extreme for performance and scale. However, since DataFrames live in the JVM and TensorFlow runs in a Python process, any integration between the two frameworks means that every object has to be serialized, go through inter-process communication (!) in both ways, and copied at least twice in memory. TensorFrames public benchmarks report a 4x speedup by just copying the data within the JVM process (and much more when using GPUs).
We see the same issue when using spaCy with Spark: Spark is highly optimized for loading & transforming data, but running an NLP pipeline requires copying all the data outside the Tungsten optimized format, serializing it, pushing it to a Python process, running the NLP pipeline (this bit is lightning fast), and then re-serializing the results back to the JVM process. This naturally kills any performance benefits you would get from Spark’s caching or execution planner, requires at least twice the memory, and doesn’t improve with scaling. Using CoreNLP eliminates the copying to another process, but still requires copying all text from the data frames and copying the results back in.
So our first order of business is to perform the analysis directly on the optimized data frames, as Spark ML already does (credit: ML Pipelines introduction post by Databricks):
Our second core requirement was frictionless reuse of existing Spark libraries. Part of it is our own pet peeve – why does every NLP library out there have to build its own topic modeling and word embedding implementations? The other part is pragmatic – we’re a small team under tight deadlines and need to make the most of what’s already there.
When we started thinking about a Spark NLP library, we first asked Databricks to point us to whoever is already building one. When the answer came there there isn’t one, the next ask was to help us make sure the design and API of the library fully meet Spark ML’s API guidelines. The result of this collaboration is that the library is a seamless extension of Spark ML, so that for example you can build this kind of pipeline:
In this code, the document assembler, tokenizer, and stemmer come from the Spark NLP library – the com.jsl.nlp.* package. The TF hasher, IDF and labelDeIndex all come from MLlib’sorg.apache.spark.ml.feature.* package. The dtree stage is a spark.ml.classification.DecisionTreeClassifier.
All these stages run within one pipeline that is configurable, serializable and testable in the exact same way. They also run on a data frame without any copying of data (unlike spark-corenlp), enjoying Spark’s signature in-memory optimizations, parallelism and distributed scale out.
What this means is the John Snow Labs NLP library comes with fully distributed, heavily tested and optimized topic modeling, word embedding, n-gram generation, and cosine similarity out of the box. We didn’t have to build them though – they come with Spark.
Most importantly, it means that your NLP and ML pipelines are now unified. The above code sample is typical, in the sense that it’s not “just” an NLP pipeline – NLP is used to generate features which are then used to train a decision tree. This is typical of question answering tasks. A more complex example would also apply named entity recognition, filtered by POS tags and coreference resolution; train a random forest, taking into account both NLP-based features and structured features from other sources; and use grid search for hyper-parameter optimization. Being able to use a unified API pays dividends whenever you need to test, reproduce, serialize or publish such a pipeline – even beyond the performance and reuse benefits.
Our third core requirement is delivering a mission-critical, enterprise-grade NLP library. We make our living building production software. Many of the most popular NLP packages today have academic roots – which shows in design trade-offs that favor ease of prototyping over runtime performance, breadth of options over simple minimalist API’s, and downplaying of scalability, error handling, frugal memory consumption and code reuse.
The John Snow Labs NLP library is written in Scala. It includes Scala and Python APIs for use from Spark. It has no dependency on any other NLP or ML library. For each type of annotator, we do an academic literature review to find the state of the art, have a team discussion and decide which algorithm(s) to implement. Implementations are evaluated on three criteria:
- Accuracy – there’s no point in a great framework, if it has sub-par algorithms or models.
- Performance – runtime should be on par or better than any public benchmark. No one should have to give up accuracy because annotators don’t run fast enough to handle a streaming use case, or don’t scale well in a cluster setting.
- Trainability or Configurability – NLP is an inherently domain-specific problem. Different grammars and vocabularies are used in social media posts vs. academic papers vs. SEC filings vs. electronic medical records vs. newspaper articles.
The library is already in use in enterprise projects – which means that the first level of bugs, refactoring, unexpected bottlenecks and serialization issues have been resolved. Unit test coverage and reference documentation are at a level that made us comfortable to make the code open source.
John Snow Labs is the company leading and sponsoring the development of the Spark NLP library. The company provides commercial support, indemnification and consulting for it. This provides the library with long-term financial backing, a funded active development team, and a growing stream of real-world projects that drives robustness and roadmap prioritization.
If you need NLP for your current project, head to the John Snow Labs NLP for Apache Spark homepage or quickstart guide and give it a try. Prebuilt maven central (Scala) and pip install (Python) versions are available. Send us questions or feedback to firstname.lastname@example.org or via Twitter, LinkedIn or Facebook.
Let us know what functionality you need next.
Here are some of the requests we’re getting, and are looking for more feedback to design and prioritize:
- Provide a SparkR client
- Provide “Spark-free” Java and Scala versions
- Add a state of the art annotator for coreference resolution
- Add a state of the art annotators for polarity detection
- Add a state of the art annotator for temporal reasoning
- Publish sample applications for common use cases such as question answering, text summarization or information retrieval
- Train and publish models for new domains or languages
- Publish reproducible, peer reviewed accuracy and performance benchmarks
If you’d like to extend or contribute to the library, start by cloning the John Snow Labs NLP for Spark GitHub repository. We use pull requests and GitHub’s issue tracker to manage code changes, bugs and features. The library is still in its early days and we highly appreciate contribution and feedback of any kind.
Bio: David Talby is a Consulting CTO at Usermind, specializing in applying big data and data science in healthcare.
- [eBook] A Gentle Introduction to Apache Spark(tm)
- Building a Wikipedia Text Corpus for Natural Language Processing
- Search Millions of Documents for Thousands of Keywords in a Flash