Reach

What is it?

Reach stands for Reading and Assembling Contextual and Holistic Mechanisms from Text. In plain English, Reach is an information extraction system for the biomedical domain, which aims to read scientific literature and extract cancer signaling pathways. Reach implements a fairly complete extraction pipeline, including: recognition of biochemical entities (proteins, chemicals, etc.), grounding them to known knowledge bases such as Uniprot, extraction of BioPAX-like interactions, e.g., phosphorylation, complex assembly, positive/negative regulations, and coreference resolution, for both entities and interactions.

Reach is developed using Odin, our open-domain information extraction framework, which is released within our processors repository.

Please scroll down to the bottom of this page for additional resources, including a Reach output visualizer, REST API, and datasets created with Reach.

Licensing

This project is, and will always be, free for research purposes. However, starting with version 1.2, we are using a license that restricts its use for commercial purposes. Please contact us for details.

Changes

1.4.0 - Require separate Processors Server (Reach is the client). Add CLUProcessor type.
1.4.0 - Update to use Processors 6.1.5 (for refactored Client/Server). Update run scripts.
much more...

Authors

Reach was created by the following members of the CLU lab at the University of Arizona:

Citations

If you use Reach, please cite one of the following papers:

@Article{Escarcega:2018,
  author={Valenzuela-Esc{\'a}rcega, Marco A and Babur, {\"O}zg{\"u}n and Hahn-Powell, Gus and Bell, Dane and Hicks, Thomas and Noriega-Atala, Enrique and Wang, Xia and Surdeanu, Mihai and Demir, Emek and Morrison, Clayton T},
  title={Large-scale Automated Machine Reading Discovers New Cancer Driving Mechanisms},
  journal={Database: The Journal of Biological Databases and Curation},
  url={http://clulab.cs.arizona.edu/papers/escarcega2018.pdf},
  doi={10.1093/database/bay098},
  year={2018},
  publisher={Oxford University Press}
}

@inproceedings{Valenzuela+:2015aa,
  author    = {Valenzuela-Esc\'{a}rcega, Marco A. and Gustave Hahn-Powell and Thomas Hicks and Mihai Surdeanu},
  title     = {A Domain-independent Rule-based Framework for Event Extraction},
  organization = {ACL-IJCNLP 2015},
  booktitle = {Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing: Software Demonstrations (ACL-IJCNLP)},
  url = {http://www.aclweb.org/anthology/P/P15/P15-4022.pdf},
  year      = {2015},
  pages = {127--132},
  Note = {Paper available at \url{http://www.aclweb.org/anthology/P/P15/P15-4022.pdf}},
}

More publications from the Reach project are available here.

Installation

This software requires Java 1.8, Scala 2.11, and CoreNLP 3.x or higher.

The jar is available on Maven Central. To use, simply add the following dependency to your pom.xml:

<dependency>
   <groupId>org.clulab</groupId>
   <artifactId>reach_2.11</artifactId>
   <version>1.4.0</version>
</dependency>

The equivalent SBT dependencies are:

libraryDependencies ++= Seq(
    "org.clulab" %% "reach" % "1.4.0"
)

How to compile the source code

This is a standard sbt project, so use the usual commands (i.e. sbt compile, sbt assembly, etc.) to compile. Add the generated jar files under target/ to your $CLASSPATH, along with the other necessary dependency jars. Take a look at build.sbt to see which dependencies are necessary at runtime.

Running Reach

Processing a directory of `.nxml` papers

The most common use of Reach is to process a directory containing one or more papers in the proper formats. A Wiki page documents the supported input formats.

Some configuration is necessary before running Reach. Please refer to the Running Reach Wiki page for detailed information on configuring and running Reach.

The Interactive Shell

An interactive shell can be run from the command line to process small fragments of entered text. The shell is useful for reviewing and understanding the operation of Reach, including NER, entity, and event processing and rule debugging. To start a Reach shell, run the runReachShell.sh script:

runReachShell.sh

At the shell prompt enter :help to get a list of available commands.

The sieve-based assembly system

Reach now provides a sieve-based system for assembly of event mentions. While still under development, the system currently has support for (1) exact deduplication for both entity and event mentions, (2) unification of mentions through coreference resolution, and (3) the reporting of intra and inter-sentence causal precedence relations (ex. A causally precedes B) using linguistic features, and (4) a feature-based classifier for causal precedence. Future versions will include additional sieves for causal precedence and improved approximate deduplication.

For more details on the sieve-based assembly system, please refer to the following paper:

@inproceedings{GHP+:2016aa,
  author       = {Gus Hahn-Powell and
Dane Bell and
Marco A. Valenzuela-Esc\'{a}rcega and Mihai Surdeanu},
  title        = {This before That: Causal Precedence in the Biomedical Domain},
  booktitle    = {Proceedings of the 2016 Workshop on Biomedical Natural Language Processing},
  organization = {Association for Computational Linguistics}
  year         = {2016}
  Note         = {Paper available at \url{https://arxiv.org/abs/1606.08089}}
}

The sieve-based assembly system can be run over a directory of .nxml and/or .csv files:

sbt "runMain org.clulab.reach.RunReachCLI"

In src/main/resources/application.conf, you will need to...

set outputTypes to ["assembly-tsv"]
set your input directory of papers via papersDir
set your output directory via outDir

Currently, two .tsv files are produced for assembly results within each paper:

results meeting MITRE's (March 2016) requirements
results without MITRE's constraints

Two additional output files are produced for assembly results across all papers:

results meeting MITRE's (March 2016) requirements
results without MITRE's constraints

The interactive Assembly shell

You can run interactively explore assembly output for various snippets of text using the assembly shell:

sbt "runMain org.clulab.assembly.AssemblyShell"

Modifying the code

Reach builds upon our Odin event extraction framework. If you want to modify event and entity grammars, please refer to Odin's Wiki page for details. Please read the included Odin manual for details on the rule language and the Odin API.

Reach web services

We have developed a series of web services on top of the Reach library. All are freely available here.

Reach datasets

We have generated multiple datasets by reading publications from the open-access PubMed subset using Reach. All datasets are freely available here.

Funding

The development of Reach was funded by the DARPA Big Mechanism program under ARO contract W911NF-14-1-0395.

clu-ling/reach