The Knowledge Graph Conference Icon
The Knowledge Graph Conference
  • 馃彔Home
  • 馃搮Events
  • 馃懁Members
  • 馃數Announcements
  • 馃數Ask
  • 馃數Ask The Ontologists
  • 馃數Events
  • 馃數Jobs
  • 馃數Promotions
  • 馃數Share
Powered by Tightknit
Ask
Ask

Seeking Guidance on Triples Extraction with Language Models

Avatar of Mark KramMark Kram
路Mar 27, 2023 10:14 AM

Hello, is there anyone here with experience in applying Language models and other Machine Learning Techniques to extract triples from text? I have a question regarding the mapping of the extracted predicates to my defined predicates in the ontology.

馃憤1

2 comments

路 Sorted by Oldest
  • Avatar of Panos A.
    Panos A.
    路

    HI Mark, I have done some work on this, what is your question?

  • Avatar of Mark Kram
    Mark Kram
    路

    Hi Panos! I am comparing two approaches with LLMs. I have some domain specific text documents and defined entities and relations that I want to have extracted. For the first approach, I finetune a pretrained model for triple extraction, however as data is highly unbalanced the accuracy for some triples is just not good. For the second approach, I use an Open Information Extraction (OIE) model which is not finetuned on my relations. Instead, it outputs any kind of triples out of a text. This approach recognizes all possible relations but does not apply "relation-linking". My question is, is there a way to map the relations from the OIE model to my defined relations. Maybe with ambeddings and similarity? For example, I get the relation "Alice, works at, Adidas" from the OIE model and want to map the relation "works for" to my defined relation "Is Employed At"(based on my ontology). Hope I made myself clear. If not we can exchange thrpuh PM. Thanks!