Our paper has been accepted to #EMNLP2024 main conference! Our 'relevance thesaurus' reveals neural IR models show bias in prioritizing certain car brands over others.
Curious about the implications? Check out our work: arxiv.org/abs/2410.03584
aclweb.org/anthology/D16-104…
Conclusion : transfer leanring in NLP has benefits, if they are semantically similar.
1. If you transfer, the output layer must be post-trained for target domain
2. Multi-Task Learning is best, Initializing with trained parameters is ok