Framework to homogenize and combine the output of different entity linking tools, using the level of agreement as a confidence score
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

ELVIS (Entity Linking Voting and Integration System)

Framework to homogenize and combine the output of different entity linking tools, using the level of agreement between different tools as a confidence score. It acts as a python wrapper for different Entity Linking systems.

You can run different Entity Linking tools (e.g. Tagme, Babelfy, DBpeida-Spotlight) from the same script and then convert the ouput fo the different systems into a uniform format (ELVIS format). Then, you can combine the information from the different systems into a unique output where the number of tools that agree on the identification of entities is used as a confidence score. In addition, you can convert ELVIS format to NIF format and XML format that can be loaded in GATE.

The system does a batch process on a folder of text files. runs a specific entity linking tool on every text file. For each text file, it tokenizes the text and run the entity linking tool with the whole text. Then entities are splitted in sentences in the output files. There is one output file for each input text file.

After running Entity Linking you can homogenize the output by runnin It takes data from DBpedia to improve the level of information of every entity, and to provide a common output for all the systems. It adds type and categories information, resolve redirections within DBpedia, and solve character encoding issues.

Finally, the homogenized outputs of the different tools can be combined in a new output called agreement output by running Here, the system filter the entities where more than one tool have agreed in URI and offset. This script receives a parameter called level. If level 3 is selected, only entities where the 3 tools agreeed are selected, instead if level 2 is selected, only entities where at least 2 tools agreed are selected.

The system works with three Entity Linking tools, namely Tagme, Babelfy and DBpedia Spotlight. However, it is easily extensible and other tools can be used.

Needed files

The homogenizer needs information from DBpedia. We have set up a server to query this information. However, the performance of this server is slow. If you need to process a big amount of data it highly recommended to download the DBpedia files in your machine. There is an argument when calling the homogenizer where you can choose to use local or remote DBpedia data. To use it locally, you have to download the following files from DBpedia and put them in the dbpedia folder.



In src/ you should select the path for source texts and dbpedia files. In addition you have to add your API KEYS for the different Entity Linking tools.


First you run the entity linking tool with the script for the different tools. It will create a folder in entities/ folder for the specified tool and set of texts. Inside this folder there will be a file for every file in the source folder.

Then you run the script to harmonize the output obtained from the different systems. It will create a new folder inside every tool folder with the suffix _h.

Finally, you have to run to obtain the agreement output. It will create a new folder inside the entities/agreement/ folder with the agreement output.


Delete the entities/ folder. Download DBpedia files. Configure the file, adding the API Keys of the different tools. Run the following pipeline of scripts:

python spotlight example python tagme example python babelfy example python all example python example_h 2

Then you will find the final agreement output in the entities/agreement/ folder.

Note that will take some time as it has to load a lot of information from DBpedia in memory.

To convert the homogenized ouptut of DBpedia Spotlight to NIF fro example you have to run

python spotlight example_h

Or to convert for instance the agreement output from ELVIS format to XML format GATE compatible, you must run

python agreement example_h_2


If you use this code for research purposes, please cite our paper:

Sergio Oramas, Luis Espinosa-Anke, Mohamed Sordo, Horacio Saggion, and Xavier Serra. 2016. ELMD : An Automatically Generated Entity Linking Gold Standard Dataset in the Music Domain. In In Proceedings of the 10th International Conference on Language Resources and Evaluation, LREC 2016.


This project is licensed under the terms of the MIT license.