Research map / Label-efficient and robust learning / Self-supervised and contrastive
Multi-Task Self-Supervised Learning for Disfluency Detection
Shaolei Wang, Wangxiang Che, Qi Liu, Pengda Qin, Ting Liu, William Yang 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 |
|---|---|---|---|
| combine tagging and sentence-classification pre-training rather than using either task alone (ablation) | Our self-supervised model · Switchboard test, Full | F1: 87.7 / 87.0 -> 90.2 | Improved |
| combine tagging and sentence-classification pre-training rather than using either task alone (ablation) | Our self-supervised model · Switchboard test, 1000 sents | F1: 82.3 / 80.7 -> 86.1 | Improved |
| add self-supervised pre-training instead of random initialization (ablation) | Transformer-based disfluency detector · Switchboard test, Full | F1: 78.1 -> 90.2 | Improved |
| add self-supervised pre-training instead of random initialization (ablation) | Transformer-based disfluency detector · Switchboard test, 1000 sents | F1: 62.1 -> 86.1 | Improved |
| increase model size across the tested configurations (ablation) | Our self-supervised model · Switchboard dev, 1000 sents | F1: 78.1 -> 79.7 -> 81.8 -> 82.9 -> 84.5 -> 86.3 | Improved |
| increase model size across the tested configurations (ablation) | Our self-supervised model · Switchboard dev, Full | F1: 85.7 -> 86.8 -> 88.2 -> 89.1 -> 89.6 -> 90.3 | Improved |
| concatenate representations from BERT and the self-supervised model | Switchboard disfluency detector · Switchboard test, Full | F1: 90.2 / 90.1 -> 91.4 | Improved |
| concatenate representations from BERT and the self-supervised model | Switchboard disfluency detector · Switchboard test, 1000 sents | F1: 86.1 / 82.4 -> 87.8 | Improved |
| pre-train on increasing amounts of automatically generated pseudo data | Our self-supervised model · Switchboard | F1 | Improved |
| directly merge pseudo disfluent sentences with gold training data instead of using the proposed pre-training process | Transition-based and transformer-based baselines · Switchboard | F1 | Hurt |
Closest papers by meaning
- Uncertainty-aware Self-training for Few-shot Text Classification (NeurIPS 2020 · similarity 0.86)
- DABS: a Domain-Agnostic Benchmark for Self-Supervised Learning (NeurIPS 2021 · similarity 0.84)
- Silencer: From Discovery to Mitigation of Self-Bias in LLM-as-Benchmark-Generator (NeurIPS 2025 · similarity 0.83)
- Self Iterative Label Refinement via Robust Unlabeled Learning (NeurIPS 2025 · similarity 0.83)
- Self-contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation (ICLR 2024 · similarity 0.83)
- DenoSent: A Denoising Objective for Self-Supervised Sentence Representation Learning (AAAI 2024 · similarity 0.83)
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.
