clulab / eidos

Machine reading system for World Modelers

GitHub

Build Status

Eidos

Eidos is an open-domain machine reading system designed by the Computational Language Understanding (CLU) Lab at University of Arizona for the World Modelers DARPA program. It uses a cascade of Odin grammars to extract events from free text.

Eidos identifies entities like "food insecurity" along with a growing list of arguments on those entities, such as increases / decreases / quantifications. It subsequently detects events that occur between entities (directed causal events, for example) as in "food insecurity causes increased migration". The list of arguments and events is updated frequently and is documented on the JSON-LD page.

Contents

Running Eidos

Eidos in this repository is formatted as a standard sbt project, that's the Scala Build Tool, so you will need to download and install sbt to "run" the included programs as described here. (See the section on integrating for other ways to use Eidos.) sbt in turn requires Java, so that will need to have been installed as well, but that's the extent of it. Other software packages that Eidos depends on will be downloaded automatically by sbt. The contents of this repository can be downloaded with git or via the "Clone or download" button on the project's GitHub page. Assuming you have done this and any necessary unzipping, you should have a directory called eidos containing the file build.sbt among others. cd to this directory.

Using the Sample Apps

The software includes several sample applications that demonstrate its abilities. Some are more colorful and interactive and others are tailored for machine readable output. They are fairly large programs and the software is configured in file .jvmopts to supply Java with lots of memory. If you see complaints about inadequate heap space, please check the notes section for possible remedies.

Webapp

The webapp provides the most colorful, graphical, and perhaps easiest to understand output. It can be started up directly from the command line in one fell swoop

> sbt webapp/run

or from within sbt with

> sbt
sbt:eidos> webapp/run

sbt may take several minutes to bring up the application, especially the first time, as various dependencies are downloaded and Eidos itself is compiled. Numerous logging messages should keep you posted on the progress.

After starting the webapp, use a web browser to navigate to localhost:9000. There you should see something like this:

Webapp window without text

You can now submit texts for Eidos to process. Please note that the very first submission will require extra time as lazily loaded parts of the system are initialized, but subsequent texts will be processed much more quickly.

Webapp window with text

To eventually escape from sbt, you can stop the web server with Control-D and then quit the program with exit.

Web Service

When the webapp is run, it exposes a web service at port 9000 via the /process_text endpoint. It accepts a POST request and requires JSON with the following parameter:

  • text: the text you wish to submit for parsing by Eidos.

For example, we can use the Python requests library to interact with the web service with the following:

import requests

text = """Drought increases regional insecurity."""

webservice = 'http://localhost:9000'
res = requests.post('%s/process_text' %webservice, headers={'Content-type': 'application/json'}, json={'text': text})

json_dict = res.json()

Using CURL we can do the same with:

> curl \
  --header "Content-type: application/json" \
  --request POST \
  --data '{"text": "Drought increases regional insecurity."}' \
  http://localhost:9000/process_text

Interactive Shell

EidosShell is an interactive text-based program for testing the output of Eidos. Its output is less complete than the webapp, but it stands alone, no web browser is required, and it consumes fewer resources. A script is provided to run it with

> ./shell

or

> .\shell

depending on operating system, or it can be run in the usual sbt way with

> sbt "runMain org.clulab.wm.eidos.apps.EidosShell"

or

> sbt
eidos:sbt> runMain org.clulab.wm.eidos.apps.EidosShell

You should eventually be prompted for an input text or command:

(Eidos)>>> Food insecurity causes increased migration.

The first text processed will again take extra time as the lazily loaded parts are initialized, but the eventual pertinent output should look substantially similar to

sentence #0
Food insecurity causes increased migration .
Tokens: (0,Food,NNP), (1,insecurity,NN), (2,causes,VBZ), (3,increased,VBN), (4,migration,NN), (5,.,.)
roots: 2
outgoing:
        0:
        1: (0,compound)
        2: (1,nsubj) (4,dobj) (5,punct)
        3:
        4: (3,amod)
        5:
incoming:
        0: (1,compound)
        1: (2,nsubj)
        2:
        3: (4,amod)
        4: (2,dobj)
        5: (2,punct)


timeExpressions:

entities:
List(Concept, Entity) => Food insecurity
         ------------------------------
         Rule => simple-np
         Type => TextBoundMention
         ------------------------------
         Concept, Entity => Food insecurity
         ------------------------------


List(Concept, Entity) => increased migration
         ------------------------------
         Rule => simple-np++Increase_ported_syntax_6_verb
         Type => TextBoundMention
         ------------------------------
         Concept, Entity => increased migration
          * Attachments: Increase(increased,None)
         ------------------------------



events:
List(Causal, DirectedRelation, EntityLinker, Event) => Food insecurity causes increased migration
         ------------------------------
         Rule => ported_syntax_1_verb-Causal
         Type => EventMention
         ------------------------------
         trigger => causes
         cause (Concept, Entity) => Food insecurity
         effect (Concept, Entity) => increased migration
          * Attachments: Increase(increased,None)

         ------------------------------


==================================================

To exit the program, enter :exit as the menu indicates. Another exit will close sbt.

ExtractAndExport

If the texts to be processed can be placed into one or more files, especially if the files are located in the current directory and end with .txt, then the ExtractAndExport app is particularly handy. The expected location of the files, the extension, and several other parameters are actually specified in a configuration file, reference.conf, which can be edited. (Other apps from the section below allow command line specification of more of the parameters.) However, if default values are satisfactory, one command will process the *.txt files into *.txt.jsonld, *.txt.mitre.tsv, and *.txt.serialized with output in corresponding formats.

> sbt "runMain org.clulab.wm.eidos.apps.ExtractAndExport"

or of course

> sbt
std:eidos> runMain org.clulab.wm.eidos.apps.ExtractAndExport

Other Apps

Other apps are located in directory src/main/scala/org/clulab/wm/eidos/apps. ExtractFromDirectory, for instance, is similar to the previously described app, but it allows specification of directories on the command line and only outputs JSON-LD format. The command is

> sbt "runMain org.clulab.wm.eidos.apps.ExtractFromDirectory path/to/input/directory path/to/output/directory"

or its two-line counterpart. Again, files in the input directory should end with .txt and the extracted mentions from each file will be saved in corresponding JSON-LD files. Note that you cannot use tildes (~) in the invocation in lieu of the home directory for this and most Java-like programs.

Using the Scala API

One is of course not limited to using the existing apps: one can write new programs which make use of Eidos's Scala/Java interface. New programs can be contained in the same sbt project or a different one, in Java projects, in Python projects, or elsewhere. See the section on integration for tips on how this is accomplished. This description assumes that a program is added to the collection of existing apps in their apps directory or somewhere nearby. There are two major output formats provided through the Scala API.

Prettified Display

To produce a pretty display of the extracted mentions, use EidosSystem to extract an annotatedDocument from the text and then display its mentions.

package org.yourself.eidosClient

import org.clulab.wm.eidos.EidosSystem
import org.clulab.wm.eidos.utils.DisplayUtils.displayMention

object YourClassName extends App {
  val text = "Water trucking has decreased due to the cost of fuel."

  // Initialize the reader
  val reader = new EidosSystem()

  // Extract the mentions
  val annotatedDocument = reader.extractFromText(text)

  // Display in a pretty way
  annotatedDocument.odinMentions.foreach(displayMention)
}

When called with

sbt:eidos> runMain org.yourself.eidosClient.YourClassName

it should produce output similar to this:

List(NounPhrase, Entity) => Water trucking
	------------------------------
	Rule => simple-np++Decrease_ported_syntax_2_verb
	Type => TextBoundMention
	------------------------------
	NounPhrase, Entity => Water trucking
	  * Attachments: Decrease(decreased,None)
	------------------------------

List(NounPhrase, Entity) => cost of fuel
	------------------------------
	Rule => simple-np
	Type => TextBoundMention
	------------------------------
	NounPhrase, Entity => cost of fuel
	------------------------------

List(Causal, DirectedRelation, EntityLinker, Event) => Water trucking has decreased due to the cost of fuel
	------------------------------
	Rule => dueToSyntax2-Causal
	Type => EventMention
	------------------------------
	trigger => due
	cause (NounPhrase, Entity) => cost of fuel
	effect (NounPhrase, Entity) => Water trucking
	  * Attachments: Decrease(decreased,None)
	------------------------------

This description is based on the ExtractFromText app.

Export to JSON-LD

To export extractions, and in fact causal graphs, as JSON-LD, take the annotatedDocument, create a JLDCorpus, serialize it, and then print a pretty string version.

package org.yourself.eidosClient

import scala.collection.Seq
import org.clulab.serialization.json.stringify
import org.clulab.wm.eidos.EidosSystem
import org.clulab.wm.eidos.serialization.jsonld.JLDCorpus

object YourClassName extends App {
  val text = "Water trucking has decreased due to the cost of fuel."

  // Initialize the reader
  val reader = new EidosSystem()

  // Extract the mentions
  val annotatedDocument = reader.extractFromText(text)

  // Export to JSON-LD
  val corpus = new JLDCorpus(Seq(annotatedDocument), reader.loadableAttributes.adjectiveGrounder)
  val mentionsJSONLD = corpus.serialize()
  println(stringify(mentionsJSONLD, pretty = true))
}

The output that results when such code is called with

sbt:eidos> runMain org.yourself.eidosClient.YourClassName

can be quite long. The JSON-LD export format for Eidos is defined on a wiki page.

Connecting Eidos

Output of Eidos can already be connected to other tools for visualization and modelling. Follow their links for details.

INDRA

Eidos reading output (JSON-LD) can be visualized using INDRA and Jupyter notebooks. Here is an example: Link to Delphi

Delphi

Events extracted using Eidos can be converted to INDRA Influence statements, which are bespoke data structures designed for modeling causal networks.

Example usage:

>>> from indra.sources import eidos
>>> ep = eidos.process_text("Water trucking has decreased due to the cost of fuel.")
>>> ep.statements
[Influence(cost of fuel(), Water trucking(negative))]

Delphi is a framework built on top of INDRA that assembles causal fragments extracted by Eidos into a causal analysis graph. This causal analysis graph is then converted to a dynamic Bayes network and used to make probabilistic predictions.

Working with Eidos

Compiling

This is a standard sbt project, so the usual command, sbt compile, can be used for compilation. A plugin is included which can be used to assemble a fat jar file with the command sbt assembly. Each of these commands results in files being written to the directory target. sbt runMain can run files that are (temporarily) stored in that location.

The project can be readily imported into IntelliJ IDEA with the Scala plugin and compiled, run, and debugged from there. In addition, the Eclipse plugin is configured for use. The project can be converted for use with a Scala-fortified Eclipse with the command

> sbt eclipse

Configuring

There are two main configuration files for Eidos: reference.conf and eidos.conf in the directory src/main/resources. The former is meant to contain default values that aren't very likely to change. The latter is where users are advised to make adjustments. These values in particular are often changed:

  • useGrounding - formerly useW2V, turns on grounding, which uses Word2Vec (W2V)
  • useCacheForOntologies - enables use of smaller and faster cached ontology files
  • useCacheForW2V - enables use of smaller and faster vector files

Grounding no longer requires additional installation and in fact it is enabled by default. The large vector file used for this purpose is declared as a library dependency and will be installed automatically. One disadvantage of these settings is that the program will start up more slowly and require extra memory. One way to avoid this is to turn off grounding by setting useGrounding to false. If you want to be relatively fast and still ground, follow instructions below for optimizing to create cached versions of these files and then turn on the two values useCacheForOntologies and useCacheForW2V.

Two previous settings are no longer used:

  • useTimeNorm - activates time processing functions
  • useGeoNorm - activates geolocation functions

Their functionalty has been incorporated into the setting for entityFinders

entityFinders = ["gazetteer", "rulebased", "geonorm", "timenorm"]

in which they are configured to be active. To deactivate them, remove them from the list. They no longer require any additional configuration. All the code and resources are retrieved from projects timenorm and geonorm, which are separate from Eidos and declared in built.sbt as a library dependencies.

With useGrounding set to true and geonorm and timenorm included, your output should look more like this:

Eidos with Grounding

Optimizing

Processing the vector file and ontologies used in grounding can consume multiple minutes of time, so if you want to run Eidos more than once with grounding, then it's useful to cache a preprocessed version of all of them. This requires a large amount of memory, possibly 8 or 10GB, so please read the notes below. If the configuration file eidos.conf is adjusted appropriately (useGrounding is turned on) and there is enough memory, then the command

> sbt "runMain org.clulab.wm.eidos.apps.CacheOntologies"

should write serialized versions of the configured vector file and known ontologies to the directory ./cache/. To then use the cached copies, set useCacheForOntologies = true in eidos.conf. useCacheForW2V is set to track the same value, but can be independently controlled if need be. The program should work significantly faster afterwards.

Translating

Eidos is being translated into Portuguese! The functionality is in the alpha stages. To try it out, switch language in eidos.conf from english to portuguese. Here is a sneak preview:

Eidos in Portuguese!

Integrating

If you want to use Eidos in its default configuration, it is available to project management software like Maven and sbt in prepackaged form, independently of the GitHub repository. Code like that below can be used to declare the dependency. The extra resolver is used for glove word vectors and models from the processors project.

resolvers += "Artifactory" at "http://artifactory.cs.arizona.edu:8081/artifactory/sbt-release"

libraryDependencies ++=
  Seq(
    "org.clulab"    %% "eidos"          % "0.2.2"
  )

These instructions have otherwise assumed use of sbt which does significant management of Java's $CLASSPATH settings, directing Java both to the compiled class files and the jar files of library dependencies. The easiest way to help Java use Eidos without direct aid of sbt is to "assemble" a single jar file for Eidos:

> sbt assembly

This command should result in a large jar file being placed in a subdirectory of target based on the version of Scala configured and the current Eidos version number. The file might be target/scala-2.12/eidos-assembly-0.2.2-SNAPSHOT.jar. To access Eidos from Java, add the assembled jar file to your $CLASSPATH, project, or command line.

> java -Xmx6g -classpath target/scala-2.12/eidos-assembly-0.2.2-SNAPSHOT.jar org.yourself.eidosClient.YourClassName

Another option is to "publish" the project locally, either in an Ivy repository as sbt prefers with publishLocal, or a local Maven repository with publishM2. This will result in a smaller jar file for Eidos classes and a record of the library dependencies.

> sbt publishLocal

Eidos can then be used in other projects managed by sbt (or Maven) with the same instructions used for non-local storage.

libraryDependencies ++=
  Seq(
    "org.clulab"    %% "eidos"          % "0.2.2-SNAPSHOT"
  )

Notes

The default size of the memory allocation pool for the JVM is usually 1/4 of the computer's physical memory, but Eidos may require more RAM than that. It is currently being developed and tested with a 6GB limit or more if timenorm or grounding is configured. Performance can suffer significantly if the Java garbage collector is activated frequently.

For those using sbt, the file .jvmopts is included with the source code to arrange for more memory. No other changes should be necessary. Alternatively, an environment variable specific to sbt can be used:

SBT_OPTS=-Xmx6g

IDEs and other development tools are generally unaware of .jvmopts, but can be configured via an environment variable instead.

JAVA_TOOL_OPTIONS=-Xmx6g

Other situations may require a more general setting.

_JAVA_OPTIONS=-Xmx6g

The procedure for defining these variables is dependent on operating system and shell.

License

While we will soon be licensed as Apache, currently one dependency has a GPL license. This will be removed very soon and the license will be updated.