Training and evaluating GROBID models
Grobid uses different sequence labelling models depending on the labeling task to be realized. For a complex extraction and parsing tasks (for instance header extraction and parsing), several models are used in cascade. The current models are the following ones:
The models are located under
grobid/grobid-home/models. Each of these models can be retrained using amended or additional training data. For production, a model is trained with all the available training data to maximize the performance. For development purposes, it is also possible to evaluate a model with part of the training data.
Train and evaluate
The sub-project grobid-trainer is be used for training. The training data is located under the grobid-trainer/resources folder, more precisely under
where MODEL is the name of the model (so for instance,
When generating a new model, a segmentation of data can be done (e.g. 80%-20%) between TEI files for training and for evaluating. This segmentation can be done following two manner:
manually: annotated data are moved into two folders, data for training have to be present under
grobid/grobid-trainer/resources/dataset/*MODEL*/corpus/, and data for evaluation under
automatically: The data present under
grobid/grobid-trainer/resources/dataset/header/corpusare randomly split following a given ratio (e.g. 0.8 for 80%). The first part is used for training and the second for evaluation.
There are different ways to generate the new model and run the evaluation, whether running the training and the evaluation of the new model separately or not, and whether to split automatically the training data or not. For any methods, the newly generated models are saved directly under grobid-home/models and replace the previous one. A rollback can be made by replacing the newly generated model by the backup record (
Train and evaluation in one command
Under the main project directory
grobid/, run the following command to execute both training and evaluation:
> ./gradlew <training goal. I.E: train_name-header>
Example of goal names:
The files used for the training are located under
grobid/grobid-trainer/resources/dataset/*MODEL*/corpus, and the evaluation files under
Examples for training the header model:
> ./gradlew train_header
Examples for training the model for names in header:
> ./gradlew train_name_header
Train and evaluation separately
Under the main project directory
grobid/, execute the following command (be sure to have built the project as indicated in Install GROBID):
Train (generate a new model):
> java -Xmx1024m -jar grobid-trainer/build/libs/grobid-trainer-<current version>-onejar.jar 0 <name of the model> -gH grobid-home
The training files considered are located under
The training of the models can be controlled using different parameters. The
nbThreads parameter in the configuration file
grobid-home/config/grobid.yaml can be increased to speed up the training. Similarly, modifying the stopping criteria can help speed up the training. Please refer this comment to know more.
> java -Xmx1024m -jar grobid-trainer/build/libs/grobid-trainer-<current version>-onejar.jar 1 <name of the model> -gH grobid-home
The considered evaluation files are located under
Automatically split data, train and evaluate:
> java -Xmx1024m -jar grobid-trainer/build/libs/grobid-trainer-<current version>-onejar.jar 2 <name of the model> -gH grobid-home -s <segmentation ratio as a number between 0 and 1, e.g. 0.8 for 80%>
For instance, training the date model with a ratio of 75% for training and 25% for evaluation:
> java -Xmx1024m -jar grobid-trainer/build/libs/grobid-trainer-<current version>-onejar.jar 2 date -gH grobid-home -s 0.75
A ratio of 1.0 means that all the data available under
grobid/grobid-trainer/resources/dataset/*MODEL*/corpus/ will be used for training the model, and the evaluation will be empty. Automatic split data, train and evaluate is for the moment only available for the following models: header, citation, date, name-citation, name-header and affiliation-address.
Several runs with different files to evaluate can be made to have a more reliable evaluation (e.g. 10 fold cross-validation). For the time being, such segmentation and iterative evaluation is not yet implemented.
For robust evaluation and reporting, n-fold cross-evaluation is commonly used, see the Wikipedia article.
GROBID implementation follows the standard approach, shuffling and dividing the annotated corpus in N equals folds, and performing N training and evaluations, where N-1 folds are used for training and the last one for evaluation. Folds are rotating for each training/evaluation, and thus each fold will be used for evaluation successively at least one time. Finally the evaluation scores for the N folds are averaged, the worst and best training/evaluations being indicated as information.
For performing a N fold evaluation:
> java -Xmx1024m -jar grobid-trainer/build/libs/grobid-trainer-<current version>-onejar.jar 3 <name of the model> -gH grobid-home -n FOLD-NUMBER
FOLD_NUMBER must be > 1.
For instance for a 10-fold evaluation of the date model:
> java -Xmx1024m -jar grobid-trainer/build/libs/grobid-trainer-<current version>-onejar.jar 3 date -gH grobid-home -n 10
Generation of training data
To generate some training datas from some input pdf, the batch grobid-core-
<current version>.onejar.jar can be used: Grobid batch (
For each pdf in input directory, GROBID generates different files because each model has separate training data, and thus uses separate files. So we have one file for header (
*.training.header.tei.xml), one for dates (
*.training.date.tei.xml), one for names, etc...
When a model uses PDF layout features, an additional feature file (for example
*.training.header for the header model) is generated without
If you wish to maintain the training corpus as gold standard, these automatically generated data have to be checked and corrected manually before being moved to the training/evaluation folder of the corresponding model. For correcting/checking these data, the guidelines presented in the next section must be followed to ensure the consistency of the whole training sets.
Annotation guidelines for creating the training data corresponding to the different GROBID models are available from the following page.