%0 Conference Paper %1 chang-etal-2023-revisiting-architectures %A Chang, Haw-Shiuan %A Yao, Zonghai %A Gon, Alolika %A Yu, Hong %A McCallum, Andrew %B Findings of the Association for Computational Linguistics: ACL 2023 %C Toronto, Canada %D 2023 %E Rogers, Anna %E Boyd-Graber, Jordan %E Okazaki, Naoaki %I Association for Computational Linguistics %K nlp pointer reading %P 12707--12730 %R 10.18653/v1/2023.findings-acl.805 %T Revisiting the Architectures like Pointer Networks to Efficiently Improve the Next Word Distribution, Summarization Factuality, and Beyond %U https://aclanthology.org/2023.findings-acl.805 %X Is the output softmax layer, which is adopted by most language models (LMs), always the best way to compute the next word probability? Given so many attention layers in a modern transformer-based LM, are the pointer networks redundant nowadays? In this study, we discover that the answers to both questions are no. This is because the softmax bottleneck sometimes prevents the LMs from predicting the desired distribution and the pointer networks can be used to break the bottleneck efficiently. Based on the finding, we propose several softmax alternatives by simplifying the pointer networks and accelerating the word-by-word rerankers. In GPT-2, our proposals are significantly better and more efficient than mixture of softmax, a state-of-the-art softmax alternative. In summarization experiments, without very significantly decreasing its training/testing speed, our best method based on T5-Small improves factCC score by 2 points in CNN/DM and XSUM dataset, and improves MAUVE scores by 30\% in BookSum paragraph-level dataset.

@inproceedings{chang-etal-2023-revisiting-architectures, abstract = {Is the output softmax layer, which is adopted by most language models (LMs), always the best way to compute the next word probability? Given so many attention layers in a modern transformer-based LM, are the pointer networks redundant nowadays? In this study, we discover that the answers to both questions are no. This is because the softmax bottleneck sometimes prevents the LMs from predicting the desired distribution and the pointer networks can be used to break the bottleneck efficiently. Based on the finding, we propose several softmax alternatives by simplifying the pointer networks and accelerating the word-by-word rerankers. In GPT-2, our proposals are significantly better and more efficient than mixture of softmax, a state-of-the-art softmax alternative. In summarization experiments, without very significantly decreasing its training/testing speed, our best method based on T5-Small improves factCC score by 2 points in CNN/DM and XSUM dataset, and improves MAUVE scores by 30{\%} in BookSum paragraph-level dataset.}, added-at = {2024-04-24T09:03:55.000+0200}, address = {Toronto, Canada}, author = {Chang, Haw-Shiuan and Yao, Zonghai and Gon, Alolika and Yu, Hong and McCallum, Andrew}, biburl = {https://www.bibsonomy.org/bibtex/2646d4d300054935bfefaed9500bb52eb/tobias.koopmann}, booktitle = {Findings of the Association for Computational Linguistics: ACL 2023}, doi = {10.18653/v1/2023.findings-acl.805}, editor = {Rogers, Anna and Boyd-Graber, Jordan and Okazaki, Naoaki}, interhash = {2753121cd7756ce0429b1d810a79b8a5}, intrahash = {646d4d300054935bfefaed9500bb52eb}, keywords = {nlp pointer reading}, month = jul, pages = {12707--12730}, publisher = {Association for Computational Linguistics}, timestamp = {2024-04-24T09:03:55.000+0200}, title = {Revisiting the Architectures like Pointer Networks to Efficiently Improve the Next Word Distribution, Summarization Factuality, and Beyond}, url = {https://aclanthology.org/2023.findings-acl.805}, year = 2023 }