@inproceedings{AspillagaEtAl:ACL:2021,
  author = {C. Aspillaga and M. Mendoza and A. Soto},
  title = {Inspecting the concept knowledge graph encoded by modern language models},
  booktitle = {{ACL}},
  year = {2021},
  abstract = {The field of natural language understanding has experienced exponential progress in the last few years, with impressive results in several tasks. This success has motivated researchers to study the underlying knowledge encoded by these models. Despite this, attempts to understand their semantic capabilities have not been successful, often leading to non-conclusive, or contradictory conclusions among different works. Via a probing classifier, we extract the underlying knowledge graph of nine of the most influential language models of the last years, including word embeddings, text generators, and context encoders. This probe is based on concept relatedness, grounded on WordNet. Our results reveal that all the models encode this knowledge, but suffer from several inaccuracies. Furthermore, we show that the different architectures and training strategies lead to different model biases. We conduct a systematic evaluation to discover specific factors that explain why some concepts are challenging. We hope our insights will motivate the development of models that capture concepts more precisely.},
url = {https://aclanthology.org/2021.findings-acl.263.pdf},
}

{"_id":"bGwMqr2HLEgncCWze","bibbaseid":"aspillaga-mendoza-soto-inspectingtheconceptknowledgegraphencodedbymodernlanguagemodels-2021","author_short":["Aspillaga, C.","Mendoza, M.","Soto, A."],"bibdata":{"bibtype":"inproceedings","type":"inproceedings","author":[{"firstnames":["C."],"propositions":[],"lastnames":["Aspillaga"],"suffixes":[]},{"firstnames":["M."],"propositions":[],"lastnames":["Mendoza"],"suffixes":[]},{"firstnames":["A."],"propositions":[],"lastnames":["Soto"],"suffixes":[]}],"title":"Inspecting the concept knowledge graph encoded by modern language models","booktitle":"ACL","year":"2021","abstract":"The field of natural language understanding has experienced exponential progress in the last few years, with impressive results in several tasks. This success has motivated researchers to study the underlying knowledge encoded by these models. Despite this, attempts to understand their semantic capabilities have not been successful, often leading to non-conclusive, or contradictory conclusions among different works. Via a probing classifier, we extract the underlying knowledge graph of nine of the most influential language models of the last years, including word embeddings, text generators, and context encoders. This probe is based on concept relatedness, grounded on WordNet. Our results reveal that all the models encode this knowledge, but suffer from several inaccuracies. Furthermore, we show that the different architectures and training strategies lead to different model biases. We conduct a systematic evaluation to discover specific factors that explain why some concepts are challenging. We hope our insights will motivate the development of models that capture concepts more precisely.","url":"https://aclanthology.org/2021.findings-acl.263.pdf","bibtex":"@inproceedings{AspillagaEtAl:ACL:2021,\n author = {C. Aspillaga and M. Mendoza and A. Soto},\n title = {Inspecting the concept knowledge graph encoded by modern language models},\n booktitle = {{ACL}},\n year = {2021},\n abstract = {The field of natural language understanding has experienced exponential progress in the last few years, with impressive results in several tasks. This success has motivated researchers to study the underlying knowledge encoded by these models. Despite this, attempts to understand their semantic capabilities have not been successful, often leading to non-conclusive, or contradictory conclusions among different works. Via a probing classifier, we extract the underlying knowledge graph of nine of the most influential language models of the last years, including word embeddings, text generators, and context encoders. This probe is based on concept relatedness, grounded on WordNet. Our results reveal that all the models encode this knowledge, but suffer from several inaccuracies. Furthermore, we show that the different architectures and training strategies lead to different model biases. We conduct a systematic evaluation to discover specific factors that explain why some concepts are challenging. We hope our insights will motivate the development of models that capture concepts more precisely.},\nurl = {https://aclanthology.org/2021.findings-acl.263.pdf},\n}\n\n","author_short":["Aspillaga, C.","Mendoza, M.","Soto, A."],"key":"AspillagaEtAl:ACL:2021","id":"AspillagaEtAl:ACL:2021","bibbaseid":"aspillaga-mendoza-soto-inspectingtheconceptknowledgegraphencodedbymodernlanguagemodels-2021","role":"author","urls":{"Paper":"https://aclanthology.org/2021.findings-acl.263.pdf"},"metadata":{"authorlinks":{}}},"bibtype":"inproceedings","biburl":"https://asoto.ing.puc.cl/AlvaroPapers.bib","dataSources":["3YPRCmmijLqF4qHXd","QjT2DEZoWmQYxjHXS"],"keywords":[],"search_terms":["inspecting","concept","knowledge","graph","encoded","modern","language","models","aspillaga","mendoza","soto"],"title":"Inspecting the concept knowledge graph encoded by modern language models","year":2021,"downloads":1}