Research map / Label-efficient and robust learning / Semi-supervised learning
Supervised and Semi-Supervised Text Categorization using LSTM for Region Embeddings
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 |
|---|---|---|---|
| Use one-hot bidirectional LSTM with pooling instead of word-vector LSTM | LSTM text categorization · IMDB, Elec, RCV1, 20NG | error rate (%): 13.50 -> 8.14; 11.74 -> 7.33; 16.04 -> 11.17; 18.0 -> 13.32 | Helped |
| Use o-2LSTM-p instead of the strongest listed SVM 1–3grams or o-CNN baseline | Supervised text categorization · IMDB, Elec, 20NG | error rate (%): 8.39 -> 8.14; 7.64 -> 7.33; 13.64 -> 13.32 | Helped |
| Use o-2LSTM-p instead of o-CNN | Supervised text categorization · RCV1 | error rate (%): 9.17 -> 11.17 | Hurt |
| Add LSTM tv-embeddings learned from unlabeled data as additional input | o-2LSTM-p · IMDB, Elec, RCV1 | error rate (%): 8.14 -> 6.66; 7.33 -> 6.08; 11.17 -> 9.24 | Helped |
| Add CNN tv-embeddings to the LSTM model (ablation) | o-2LSTM-p with LSTM tv-embeddings · IMDB, Elec | error rate (%): 6.66 -> 5.94; 6.08 -> 5.55 | Helped |
| Add CNN tv-embeddings to the LSTM model (ablation) | o-2LSTM-p with LSTM tv-embeddings · RCV1 | error rate (%): 8.62 -> 8.52 | Helped |
| Add LSTM tv-embeddings to the CNN model (ablation) | o-CNN with CNN tv-embeddings · IMDB, Elec, RCV1 | error rate (%): 6.51 -> 6.05; 6.27 -> 5.87; 7.71 -> 7.15 | Helped |
| Use one-hot LSTM with pooling rather than word-vector LSTM with pooling (ablation) | LSTM text categorization · Elec | error rate (%): 10.90 -> 7.72 | Helped |
| Remove the input and output gates (ablation) | o-LSTM-p · Elec | error rate (%): 7.72 -> 7.68 | No clear effect |
| Remove the input and output gates (ablation) | o-LSTM-p · Elec | training time (seconds per epoch): 88 -> 48 | Helped |
| Add a backward LSTM direction (ablation) | o-LSTM-p; no i/o gates · Elec | error rate (%): 7.68 -> 7.33 | Helped |
| Add a second one-hot CNN region size | o-CNN · IMDB, Elec, 20NG | error rate (%): 8.14 -> 8.04; 7.33 -> 7.48; 13.32 -> 13.55 | Mixed |
Closest papers by meaning
- A New Method of Region Embedding for Text Classification (ICLR 2018 · similarity 0.89)
- Adaptive Region Embedding for Text Classification (AAAI 2019 · similarity 0.88)
- Revisiting LSTM Networks for Semi-Supervised Text Classification via Mixed Objective Function (AAAI 2019 · similarity 0.87)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.84)
- NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025 · similarity 0.84)
- Making Text Embedders Few-Shot Learners (ICLR 2025 · similarity 0.84)
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.