A powerful retriever+pre-trained generator (eg. DPR+T5) often relies on spurious cues / generates hallucinations. Our 𝕖𝕧𝕚𝕕𝕖𝕟𝕥𝕚𝕒𝕝𝕚𝕥𝕪-guided generator learns to focus and generate on the right passages and shows large improvements in QA/fact verification/dialogue👇
2
39
248
0
50
Download Image
Retrieval-augmented seq2seq models (e.g., RAG / DPR+BART, FiD … etc) have shown state-of-the-art results across many different knowledge-intensive tasks. Yet their training regime (train with input-output only) can result in reliance on spurious cues. 👇
Link to the paper: akariasai.github.io/files/evidenti… Please feel free to contact us if you have any questions or suggestions!! Work done during my internship at AI2 under fantastic mentors @nlpmattg @HannaHajishirzi