Training Data Generation

Choose data for both mention recognition and knowledge-base linking. General NER annotations identify spans and types; linking annotations add entity identities. A specialized domain may require its own label schema and knowledgeable annotators.

There are two approaches you can adopt to gather training data for the entity linking problem.

1. Open-source datasets

2. Manual labeling

You can use one or both depending on the particular task for which we have to perform entity linking.

Open-source datasets

If the task is not extremely domain-specific and does not require very specific tags, you can avail open-source datasets as training data. For example, if you were asked to perform entity linking for a simple chatbot, you could utilize the general-purpose, open-source dataset CoNLL-2003 for named-entity recognition.

The CoNLL-2003 task paper describes separate English and German datasets, each with training, development, and test splits. English uses Reuters news; German uses the ECI corpus. The English labeled splits contain 946, 216, and 231 articles respectively. The source’s 10,788-document/1.3-million-word claim does not describe these task splits.

Original CoNLL IOB convention: