![]() ![]() Whether you're planning on getting into industry or whether you're planning on staying in academia I would strongly recommend reading around and familiarising yourself with what is now considered "old school". If I can get ~0.93 micro F1 on a text classification problem using bag-of-words features and a logistic regression model that will happily chug through 100k inferences/min on a $25/month virtual server, it is unlikely my customer will want to pay $500/month for the same throughput and 0.96 micro F1 using a fine-tuned huggingface BERTForClassification model. Regarding transformers and older methods "no longer" being useful: whilst some companies in industry (typically the well funded incumbents like FAANG and unicorns) are obsessed with transformers, the rest of the industry is decidedly /NOT/ blinded by the transformers trend.Īt my company the philosophy is to start with simple models and move towards more complex modelling approaches only if you have to. ![]() Both NER and POS are useful upstream tasks that help with co-reference resolution and entity linking. Likewise POS tagging for identifying verb chunks and noun chunks for the purpose of metadata enrichment or to improve document retrieval is quite common. ![]() NER - perhaps in combination with some form of co-reference resolution can be useful in and of itself for some use cases: for example clients might want to group/filter documents by which people and organisations are mentioned most within them. In industry we're mostly pragmatic engineers who aren't aiming for SOTA but "whatever works and is cheapest". ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |