The need to bridge between the unstructured data on the document Web and the structured data on the Data Web has led to the development of a considerable number of annotation tools. Those tools are hard to compare since published results are calculated on diverse datasets and measured in different units.
We present GERBIL, a general entity annotation system based on the BAT-Framework. GERBIL offers an easy-to-use web-based platform for the agile comparison of annotators using multiple datasets and uniform measuring approaches. To add a tool to GERBIL, all the end user has to do is to provide a URL to a REST interface to its tool which abides by a given specification. The integration and benchmarking of the tool against user-specified datasets is then carried out automatically by the GERBIL platform. Currently, our platform provides results for 9 annotators and 19 datasets with more coming. Internally, GERBIL is based on the Natural Language Programming Interchange Format (NIF) and provide Java classes for implementing APIs for datasets and annotators to NIF.
If you want to know more, please have a look at our novel paper about GERBIL currently undersubmission at SWJ http://www.semantic-web-journal.net/content/gerbil-%E2%80%93-benchmarking-named-entityrecognition-and-linking-consistently
We also used GERBIL to develop a novel benchmarking system for question answering (QA). Have a look http://www.semantic-web-journal.net/content/benchmarking-question-answering-systems
|FOX||✔||OKE Task 1|
|FRED||✔||OKE Task 1|
|FREME||✔||OKE Task 1|
|CETUS||✔||OKE Task 2|
The following table lists the annotators that are currently available and the experiment types they support. Note that some of the A2KB annotators support the D2KB experiment by offering an own API method. Other A2KB annotators can be chosen for a D2KB experiment as well as described in the wiki. However, since the comparison might not be fair, we marked these annotators with (✔) in the table. The same is done for Entity Typing.
|OKE Task 1||OKE Task 2|
The following table lists the datasets that are currently available and the experiment types they support.
|A2KB, C2KB, D2KB,
|Entity Typing||OKE Task 1||OKE Task 2|
|OKE 2015 Task 1 evaluation dataset||✔||✔||✔|
|OKE 2015 Task 1 example set||✔||✔||✔|
|OKE 2015 Task 1 gold standard sample||✔||✔||✔|
|OKE 2015 Task 2 evaluation dataset||✔|
|OKE 2015 Task 2 example set||✔|
|OKE 2015 Task 2 gold standard sample||✔|
Long term stability
The idea of GERBIL emerged in September 2014 when a couple of articles released at the same time claimed to be state-of-the-art. Especially, those approaches were not easily comparable due to their heterogeneous set-up, dataset use and evaluation metrics. Thus, we decided to build GERBIL and extend the BAT-Framework to break the barriers for people not able to write source code. GERBIL is still a young project and thus we are trying to explore the borders of our endeavour. As GERBIL has been launched within two PhD projects funded by European Social Fund we are confident that it will be a long lasting web service. The fallback is our working group AKSW which currently already hosts more than 30 open source projects. Finally, GERBIL is open source software which can be maintained and hosted by anybody.
Furthermore, the research and development unit of the University Leipzig Computation Center keeps daily backups to ensure long-term quotability.
With this project we aim at establishing a highly available, easy quotable and liable focal point for NER and NED evaluations. Additionally, we build our framework to be rapidly extensible and adaptable for future uses.
The survey data from our paper can be found at GERBIL's GitHub repository.
- Ciro Baron (University Leipzig, Germany)
- Andreas Both (R&D, Unister GmbH, Germany)
- Martin Brümmer (University Leipzig, Germany)
- Diego Ceccarelli (Unversity Pisa, Italy)
- Marco Cornolti (University of Pisa, Italy)
- Didier Cherix (R&D, Unister GmbH, Germany)
- Bernd Eickmann (R&D, Unister GmbH, Germany)
- Paolo Ferragina (University of Pisa, Italy)
- Christiane Lemke (R&D, Unister GmbH, Germany)
- Andrea Moro (Sapienza University of Rome, Italy)
- Roberto Navigli (Sapienza University of Rome, Italy)
- Francesco Piccinno (University of Pisa, Italy)
- Giuseppe Rizzo (EURECOM, France)
- Harald Sack (HPI Potsdam, Germany)
- René Speck (Institute for Applied Informatics, Germany)
- Raphaël Troncy (EURECOM, France)
- Jörg Waitelonis (HPI Potsdam, Germany)
- Lars Wesemann (R&D, Unister GmbH, Germany)