We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: MemBERT: Injecting Unstructured Knowledge into BERT

Abstract: Transformers changed modern NLP in many ways. However, they can hardly exploit domain knowledge, and like other blackbox models, they lack interpretability. Unfortunately, structured knowledge injection, in the long run, risks to suffer from a knowledge acquisition bottleneck. We thus propose a memory enhancement of transformer models that makes use of unstructured domain knowledge expressed in plain natural language. An experimental evaluation conducted on two challenging NLP tasks demonstrates that our approach yields better performance and model interpretability than baseline transformer-based architectures.
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2110.00125 [cs.CL]
  (or arXiv:2110.00125v1 [cs.CL] for this version)

Submission history

From: Federico Ruggeri [view email]
[v1] Thu, 2 Sep 2021 09:17:04 GMT (1411kb,D)

Link back to: arXiv, form interface, contact.