search

lavis-nlp / jerex

PyTorch code for JEREX: Joint Entity-Level Relation Extractor
MIT License
63 stars 15 forks source link

What's the relation, if there is any relation, between Jerex Model and Spert Model ? #6

Closed raphael10-collab closed 3 years ago

raphael10-collab commented 3 years ago

What's the relation, if there is any relation, between Jerex Model and Spert Model (https://github.com/lavis-nlp/spert) ?

At first sight, they look similar, and may be overlapping. But I guess I'm pretty wrong

markus-eberts commented 3 years ago

SpERT and JEREX do indeed overlap to some extend. Both models extract entities (or entity mentions) and relations between them. We also borrowed some ideas from SpERT when we developed JEREX (-> mention representations, localized context, general training regime...).

However, the two models differ in two main aspects: While SpERT detects relations between mentions of entities in sentences, JEREX extracts relations between entity clusters (containing multiple mentions) in whole documents. In JEREX, only annotations on an entity(-cluster) level are required (and not between individual mentions as in SpERT). The model than aggregates signals from different parts of the document via multi-instance learning. This way, relations that require reasoning over multiple sentences can be extracted. Document-level and entity(-cluster)-level relation extraction is an active research field and there is still a lot of room for improvement.

Whether SpERT or JEREX is more suited depends on your use case. If you are interested in more details, feel free to have a look at our papers.