kermitt2
commented
4 years ago Hi Oliver, thanks !
Not sure that it's more interesting than DeepType:
-
reported results on standard datasets at https://github.com/facebookresearch/BLINK do not seem better (from what I see all lower than DeepType), and even for TAC 2010, for similar settings (not restricted to TACKBP Reference Knowledgebase, but to the complete Wikipedia, and trained with Wikipedia), we have 90.85 for DeepType versus 90.2 according to the paper
-
using BERT is heavy and entity-fishing would loose its advantages to run with minimal memory (it can run with 4GB) on commodity hardware
-
the advantage of DeepType is that it relies on very simple sequence labelling for predictions, so it can be implemented with a light DL approach in a very fast manner
However, it would be really nice to add several approaches to reproduce and test them, and indeed BLINK would be something good to test. I think their implementation relying on SOLR would be too slow for general entity disambiguation (I initially implemented an ElasticSearch IR-approach, but this is just far too slow to process complete documents with all Wikidata for all entity mentions).
Hi Patrice, Facebook recently released BLINK https://github.com/facebookresearch/BLINK https://arxiv.org/abs/1911.03814 This is a novel approach for entity linking based on BERT transformers that apparently outperform even the already impressive DeepType (Raiman & Raiman) Maybe we could try to implement the ranking/selection of Entity fishing using this approach, what do you think?
Best regards Olivier