Research map / Architectures and efficiency / RNNs and normalization
Agreement on Target-Bidirectional LSTMs for Sequence-to-Sequence Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add right-to-left LSTM with agreement decoding to l2r LSTM | GLSTM · GM-PM | WER: 29.6 -> 23.8 | Helped |
| add right-to-left LSTM with agreement decoding to l2r LSTM | GLSTM · JP-EN | ACC: 28.3 -> 33.3 | Helped |
| add right-to-left LSTM with agreement decoding to l2r LSTM | GLSTM · EN-JP | ACC: 40.1 -> 43.8 | Helped |
| use bidirectional agreement ensembles instead of same-direction ensemble of 5 LSTMs | ELSTM · JP-EN | ACC: 34.2 -> 36.3 | Helped |
| use bidirectional agreement ensembles instead of same-direction ensemble of 5 LSTMs | ELSTM · GM-PM | WER: 22.1 -> 21.2 | Helped |
| use polynomial approximation instead of joint k-best search (ablation) | BLSTM · JP-EN | ACC: 33.3 -> 33.4 | No clear effect |
| rescore joint k-best of both directions instead of l2r k-best only (ablation) | BLSTM · JP-EN | ACC: 32.7 -> 33.3 | Helped |
| use BLSTM instead of 2-LSTM same-direction ensemble (ablation) | bidirectional agreement · JP-EN | ACC: 31.2 -> 33.3 | Helped |
| increase ensemble from 5+5 to 10+10 directional LSTMs | BELSTM · JP-EN | ACC: 36.3 -> 36.5 | Helped |
| use BELSTM on long sequences vs ELSTM(5r2l) | BELSTM · JP-EN long sequences | ACC: 28.6 -> 28.6 | No clear effect |
| add bidirectional agreement to l2r LSTM on long sequences | GLSTM(l2r) · JP-EN long sequences | ACC: 17.3 -> 25.0 | Helped |
| add bidirectional agreement to improve suffix quality | GLSTM(l2r) · JP-EN long sequences | suffix precision: 65% -> 74% | Helped |
Closest papers by meaning
- Universal Transformers (ICLR 2019 · similarity 0.88)
- Learning and Generalization in RNNs (NeurIPS 2021 · similarity 0.87)
- Wider and Deeper, Cheaper and Faster: Tensorized LSTMs for Sequence Learning (NeurIPS 2017 · similarity 0.87)
- Generalization without Systematicity: On the Compositional Skills of Sequence-to-Sequence Recurrent Networks (ICML 2018 · similarity 0.87)
- FlashRNN: Unlocking Parallel Training of Nonlinear RNNs for Large Language Models (ICLR 2026 · similarity 0.87)
- Learning Recurrent Binary/Ternary Weights (ICLR 2019 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.