Skip to content
/ talen Public

A way to do annotations for NER. TALEN: Tool for Annotation of Low-resource ENtities

License

Notifications You must be signed in to change notification settings

CogComp/talen

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TALEN: Tool for Annotation of Low-resource ENtities

A lightweight web-based tool for annotating word sequences.

Screenshot of web interface

Installation

Requires Java 8 and Maven. Run:

$ ./scripts/run.sh

This will start the server on port 8009. Point a browser to localhost:8009. The port number is specified in application.properties.

This reads from config/users.txt, which has a username and password pair on each line. You will log in using one of those pairs, and then that username is tied to your activities in that session. All annotations that you do will be written to a path called <orig>-annotation-<username>, where <orig> is the original path specified in the config file, and <username> is what you chose as username.

Suppose you do some annotations, then leave the session, and come back again. If you log in with the same username as the previous session, it will reload all of the annotations right where you left off, so no work is lost.

Usage

You make annotations by clicking on words and selecting a label. If you want to remove a label, right click on a word.

To annotate a phrase, highlight the phrase, ending with the mouse in the middle of the last word. The standard box will show up, and you can select the correct label. To dismiss the annotation box, click on the word it points to.

A document is saved by pressing the Save button. If you navigate away using the links on the top of the page, the document is not saved.

Configuration

There are two kinds of config files, corresponding to the two annotation methods (see below). The document-based method looks for config files that start with 'doc-' and the sentence-based method looks for config files that start with 'sent-'.

See the example config files for the minimally required set of options.

Annotation Methods

There are two main annotation methods supported: document-based, and sentence-based.

Document-based

The document-based method is a common paradigm. You point the software to a folder of documents and each is displayed in turn, and you annotate them.

Sentence-based

The sentence-based method is intended to allow a rapid annotation process. First, you need to build an index using TextFileIndexer.java, then you supply some seed names in the config file. The system searches for these seed names in the index, and returns a small number of sentences containing them. The annotator is encouraged to annotate these correctly, and also annotate any other names which may appear. These new names then join the list of seed names, and annotation continues.

For example, if the seed name is 'Pete Sampras', then we might hope that 'Andre Agassi' will show up in the same sentence. If the annotator chooses to annotate 'Andre Agassi' also, then the system will retrieve new sentences containing 'Andre Agassi'. Presumably these sentences will contain entities such as 'Wimbledon' and 'New York City'. In principle, this will continue until some cap on the number of entities has been reached.

Using the sentence-based

First, you need to download a corpus. We have used Hindi for this. Run:

$ (If you don't already have nltk) sudo pip install -U nltk 
$ python -m nltk.downloader indian

Now convert this:

$ cd data
$ python data/getindian.py
$ cd ..

You'll notice that this created files in data/txt/hindi and in data/tajson/hindi. Now build the index:

$ mvn dependency:copy-dependencies
$ ./scripts/buildindex.sh data/tajson/hindi/ data/index_hindi 

That's it! There is already a config file called config/sent-Hindi.txt that should get you started.

Non-speaker Helps

One major focus of the software is to allow non-speakers of a language to annotate text. Some features are: inline dictionary replacement, morphological awareness and coloring, entity propagation, entity suggestions, hints based on frequency and mutual information.

How to build an index

Use buildindex.sh to build a local index for the sentence based mode. The indexdir variable will be put in the sentence-based config file. This, in turn calls TextFileIndexer.java.

Command line tool

We also ship a lightweight command line tool for TALEN. This tool will read a folder of JSON TextAnnotations (more formats coming soon) and spin up a Java-only server, serving static HTML versions of each document. This will be used only for examination and exploration.

Install it as follows:

$ ./scripts/install-cli.sh
$ export PATH=$PATH:$HOME/software/talen/

(You can change the INSTALLDIR in install-cli.sh if you want it installed somewhere else). Now it is installed, you can run it from any folder in your terminal:

$ talen-cli FolderOfTAFiles

This will serve static HTML documents at localhost:PORT (default PORT is 8008). You can run with additional options:

$ talen-cli FolderOfTAFiles -roman -port 8888

Where the -roman option uses the ROMANIZATION view in the TextAnnotation for text (if available), and the -port option uses the specified port.

Mechanical Turk

Although the main function of this software is a server based system, there is also a lightweight version that runs entirely in Javascript, for the express purpose of creating Mechanical Turk jobs.

The important files are mturkTemplate.html and annotate-local.js. The latter is a version of annotate.js, but the code to handle adding and removing spans is included in the Javascript instead of sent to a Java controller. This is less powerful (because we have NLP libraries written in Java, not Javascript), but can be run with no server.

All the scripts needed to create this file are included in this repository. It was created as follows:

$ python scripts/preparedata.py preparedata data/txt tmp.csv
$ python scripts/preparedata.py testfile tmp.csv docs/index.html

mturkTemplate.html has a lot of extra stuff (instructions, annotator test, etc) which can all be removed if desired. I found it was useful for mturk tasks. When you create the mturk task, there will be a submit button, and the answer will be put into the #finalsubmission field. The output string is a Javascript list of token spans along with label.

Citation

If you use this in your research paper, please cite us!

@inproceedings{talen2018,
    author = {Stephen Mayhew, Dan Roth},
    title = {TALEN: Tool for Annotation of Low-resource ENtities},
    booktitle = {ACL System Demonstrations},
    year = {2018},
}

Read the paper here: http://cogcomp.org/papers/MayhewRo18.pdf