Named Entity Disambiguation

Named Entity Disambiguation (NED) is the task of predicting whether a particular entity is mentioned in a sentence. For example, given the sentence “Dronning Margrethe er kendt for sin store arkæologiske interesse.” , is the entity “Margrethe II of Denmark” (corresponding to the Wikidata QID Q102139) mentioned in the sentence?

NED is the binary version of Named Entity Linking (NEL) which consists in attaching a QID to an entity. Given a sentence and a list of QIDs (potentialy mentioned in the text), NED can be used to find the most probable mentioned entity.

NED is usually used in combination with NER.

Model	Train Data	Maintainer	Tags	DaNLP
XLMR	DaNED	Alexandra Institute	0 / 1	✔

Use cases

One entity name can refer to two different persons/locations/organizations depending on the context. Then, in combination with NER, NED can be used for finding entities and their correct reference, and linking them to a description page. The result can be used for differentiating between well-known persons and anonymous individuals in order to, for example, semi-anonymize a text.

Models

🔧 XLMR

The XLM-R NED model is based on the pre-trained XLM-Roberta, a transformer-based multilingual masked language model (Conneau et al. 2020), and finetuned on the combination of the DaWikiNED dataset and the training part of the DaNED dataset. The model has been developed as part of a Master student project (ITU) by Hiêu Trong Lâm and Martin Wu under the supervision of Maria Jung Barrett (ITU) and Ophélie Lacroix (DaNLP).

The XLM-R NED model can be loaded with the load_xlmr_ned_model() method. (You can also find the model on our HuggingFace page.) Please note that the model take maximum 512 tokens as input at a time. Longer text sequences will be truncated.

from danlp.models import load_xlmr_ned_model

xlmr = load_xlmr_ned_model()

sentence = "Karen Blixen vendte tilbage til Danmark, hvor hun boede resten af sit liv på Rungstedlund, som hun arvede efter sin mor i 1939"

# to check if the entity "Karen Blixen" correspond to the QID Q182804
# you have to first generate the knowledge graph (KG) context of this QID
# use the get_kg_context_from_wikidata_qid function from danlp.utils to get the KG
# or the DaNED or DaWikiNED dataset to get the corresponding KG string (see doc below)
# (the following example has been truncated -- use the full KG)
kg_context = "udmærkelser modtaget Kritikerprisen udmærkelser modtaget Tagea Brandts Rejselegat udmærkelser modtaget Ingenio ..."

label = xlmr.predict(sentence, kg_context)

For more details about how to generate a KG context adapted to the model, see the DaNED and DaWikiNED documentation.

📈 Benchmarks

The benchmarks has been performed on the test part of the DaNED dataset. None of the models have been trained on this test part. See F1 scores below (QID is mentioned == label 1 – QID is not mentioned == label 0) :

Model	QID is mentioned	QID is not mentioned	AVG	Sentences per second (CPU*)
XLMR	83.77	91.02	87.40	~1

*Sentences per second is based on a Macbook Pro with Apple M1 chip.

The evaluation script ned_benchmarks.py can be found here.

🎓 References

• Maria Jung Barrett, Hiêu Trong Lâm, Martin Wu, Ophélie Lacroix, Barbara Plank and Anders Søgaard. Resources and Evaluations for Danish Entity Resolution. In CRAC 2021.