Fine-Grained Entity Recognizer
Switch branches/tags
Nothing to show
Clone or download
xiaoling close #12
adds mainClass entrance to the assembly jar
Latest commit ab0fa64 Feb 5, 2018

Fine-Grained Entity Recognizer (FIGER)

This distribution contains the source code for the experiments presented in the following research publication (PDF):

Xiao Ling and Daniel S. Weld (2012).
"Fine-Grained Entity Recognition",

Download the model file

One can test the trained model on the evaluation data or new data as they wish.

Run ./ to download the Model and save it at the root directory. Run ./ new for an updated model.

A better model has been trained and can be fetched by ./ new. Change the config value accordingly.


sbt >= 0.13.0

Replicate the experiments

To run the experiments in the AAAI-12 paper, you can proceed as follows:

$ ./ "aaai/exp.conf" &> aaai/exp.log

Run FIGER on new data

To make predictions on new data, please see package edu.washington.cs.figer.FigerSystem for example code or run:

$ sbt "runMain edu.washington.cs.figer.FigerSystem <text_file>"

Alternatively, you can change the parameter values (e.g. the input file name) in config/figer.conf and get a more structured output by running:

$ ./ "config/figer.conf"

Make a stand-alone jar

$ sbt assembly
# the actual path might be different
$ java -jar ./target/scala-2.10/figer-assembly-0.jar <text_file>

A simple web interface


$ sbt ~container:start

and go to localhost:8080/index.html for a simple web demo.

Training Data

The training data (Download link) is gzipped and serialized in Protocol Buffer. Please see entity.proto in the code package for the definitions. Each Mention represents an entity mention defined by its token offsets together with the tokens in the sentence where the mention appears. The labels field shows the Freebase types of the underlying entity for the mention. Use the mapping in the next section to get the corresponding FIGER types.

In config/figer.conf, make the following changes:

modelFile=<the output model file>
# the training file has to follow the specs from `entity.proto`. See `` for example
trainFile=<training file>

Then run ./ config/figer.conf to train a new model (It will need over 10G memory and about an hour to finish).

Mapping between Freebase MIDs and Wikipedia Titles

A mapping file from mids to titles is available here.