Entity Relation disambiguation: Who *helps* and who *is helped*? #9034
-
|
Thank you for bringing the Transformer revolution to the masses! Do you think it is possible in spaCy to detect agent relations, helper and helpee? John saved Sam from drowning. (John=helper, Sam=helpee) The variety seems endless…: eg John went to the store for Sam. (John=helper, Sam=helpee) What NLP problem domain am I getting into here? Is this an relation extraction problem? e.g. https://www.youtube.com/watch?v=8HL-Ap5_Axo It wasn't clear to me that it falls under relation extraction. |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 3 replies
-
|
Please don't @ individuals to get their attention, it's disruptive. What you are describing is called "Semantic Role Labeling". Our standard advice is to avoid it unless you're very sure you need it because it's a hard problem (and you often don't need it). One simple approach is to use NER to identify candidates (you can assume helper/helpee are both PERSON), and then use dependencies (possibly via a DependencyMatcher) to match on common structures, like a specific verb. This has limited coverage but high precision and is easy to iterate on. For an overview of the research check Chapter 19 of Speech and Language Processing by Jurafsky and Martin. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for the direction. Roger that on @ (and apologies). |
Beta Was this translation helpful? Give feedback.
-
|
I'd mark this as the answer but github is bugging out when I try (422 error, unprocessable entity). |
Beta Was this translation helpful? Give feedback.
-
|
Huh, weird, thanks for letting us know. Discussions are pretty new and sometimes they seem to get stuck still. |
Beta Was this translation helpful? Give feedback.
Please don't @ individuals to get their attention, it's disruptive.
What you are describing is called "Semantic Role Labeling". Our standard advice is to avoid it unless you're very sure you need it because it's a hard problem (and you often don't need it).
One simple approach is to use NER to identify candidates (you can assume helper/helpee are both PERSON), and then use dependencies (possibly via a DependencyMatcher) to match on common structures, like a specific verb. This has limited coverage but high precision and is easy to iterate on.
For an overview of the research check Chapter 19 of Speech and Language Processing by Jurafsky and Martin.