Research map / Architectures and efficiency / RNNs and normalization
MorphTE: Injecting Morphology in Tensorized Embeddings
Guobing Gan, Peng Zhang, Sunzhu Li, Xiuqing Lu, Benyou Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use MorphTE instead of original embeddings at about 20× compression | Transformer · IWSLT'14 De-En and OPUS-100 En-It, En-Es, En-Ru | BLEU: 34.5 -> 34.9; 32.9 -> 32.9; 39.1 -> 39.1; 31.6 -> 31.9 | Mixed |
| Use MorphTE instead of Word2ket at about 20× compression | Transformer · De-En, En-It, En-Es, En-Ru translation datasets | BLEU: +0.7, +0.6, 0.0, +0.6 BLEU | Improved |
| Use MorphTE instead of original embeddings | RE2 · WikiQA | MAP: 0.6798 -> 0.6834 | No clear effect |
| Use MorphTE instead of original embeddings | RE2 · WikiQA | MRR: 0.6970 -> 0.7051 | Improved |
| Use MorphTE instead of original embeddings | RE2 · SNLI | Accuracy: 0.8492 -> 0.8497 | No clear effect |
| Set MorphTE order to 3 rather than 2 or 4 (ablation) | MorphTE · IWSLT'14 De-En | BLEU: 34.4 -> 34.9; 34.4 -> 34.9 | Improved |
| Replace morphology-based segmentation with randomSeg (ablation) | MorphTE · De-En, En-It, En-Es, En-Ru translation datasets | BLEU: 34.9 -> 33.8; 32.9 -> 31.7; 39.1 -> 38.6; 31.9 -> 30.4 | Hurt |
| Replace Word2ket's individual vectors with randomly shared vectors rather than morpheme-based sharing (ablation) | Word2ket · De-En, En-It, En-Es, En-Ru translation datasets | BLEU: 34.2 -> 34.0; 32.3 -> 32.0; 39.1 -> 38.3; 31.3 -> 30.6 | Hurt |
| Increase embedding size while using MorphTE (ablation) | Transformer · IWSLT'14 De-En, embedding sizes 216, 512, 1000 | BLEU: 34.1 -> 34.2; 34.5 -> 34.9; 32.8 -> 33.4 | Mixed |
Closest papers by meaning
- All Word Embeddings from One Embedding (NeurIPS 2020 · similarity 0.88)
- Improving Word Embeddings with Convolutional Feature Learning and Subword Information (AAAI 2017 · similarity 0.87)
- Understanding Composition of Word Embeddings via Tensor Decomposition (ICLR 2019 · similarity 0.87)
- From Fully Trained to Fully Random Embeddings: Improving Neural Machine Translation with Compact Word Embedding Tables (AAAI 2022 · similarity 0.86)
- A Latent Morphology Model for Open-Vocabulary Neural Machine Translation (ICLR 2020 · similarity 0.86)
- A Simple but Tough-to-Beat Baseline for Sentence Embeddings (ICLR 2017 · similarity 0.86)
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.
