ibm-research / knowgl-large

KnowGL: Knowledge Generation and Linking from Text

The knowgl-large model is trained by combining Wikidata with an extended version of the training data in the REBEL dataset. Given a sentence, KnowGL generates triple(s) in the following format:

[(subject mention # subject label # subject type) | relation label | (object mention # object label # object type)]

If there are more than one triples generated, they are separated by $ in the output. More details in Rossiello et al. (AAAI 2023) .

The model achieves state-of-the-art results for relation extraction on the REBEL dataset. See results in Mihindukulasooriya et al. (ISWC 2022) .

The generated labels (for the subject, relation, and object) and their types can be directly mapped to Wikidata IDs associated with them.

Citation

@inproceedings{DBLP:conf/aaai/RossielloCMCG23,
  author       = {Gaetano Rossiello and
                  Md. Faisal Mahbub Chowdhury and
                  Nandana Mihindukulasooriya and
                  Owen Cornec and
                  Alfio Massimiliano Gliozzo},
  title        = {KnowGL: Knowledge Generation and Linking from Text},
  booktitle    = {{AAAI}},
  pages        = {16476--16478},
  publisher    = {{AAAI} Press},
  year         = {2023}
}

@inproceedings{DBLP:conf/semweb/Mihindukulasooriya22,
  author    = {Nandana Mihindukulasooriya and
               Mike Sava and
               Gaetano Rossiello and
               Md. Faisal Mahbub Chowdhury and
               Irene Yachbes and
               Aditya Gidh and
               Jillian Duckwitz and
               Kovit Nisar and
               Michael Santos and
               Alfio Gliozzo},
  title     = {Knowledge Graph Induction Enabling Recommending and Trend Analysis:
               {A} Corporate Research Community Use Case},
  booktitle = {{ISWC}},
  series    = {Lecture Notes in Computer Science},
  volume    = {13489},
  pages     = {827--844},
  publisher = {Springer},
  year      = {2022}
}