AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Transcormer: Transformer for Sentence Scoring with Sliding Language Modeling

Kaitao Song, Yichong Leng, Xu Tan, Yicheng Zou, Tao Qin, Dongsheng Li

NeurIPS 2022 · Architectures and efficiency › Attention and transformers · approach: Language model pretraining and induction · 3 citations (OpenAlex)

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 changedWhereMeasuredEffect
Use bidirectional sliding language modeling with triple-stream attention instead of causal language modelingTranscormer sentence scoring · IWSLT14 and WMT14 rerankingBLEU: 35.14 -> 35.24 (IWSLT14 De-En); 33.07 -> 33.10 (WMT14 De-En)Improved
Use bidirectional sliding language modeling for rescoringTranscormer sentence scoring · LibriSpeech rerankingWER: 2.30 -> 2.23 (dev-clean); 5.65 -> 5.54 (dev-other); 2.59 -> 2.49 (test-clean); 5.90 -> 5.72 (test-other)Improved
Use SLM instead of BERT MLM for linguistic acceptability scoringTranscormer · BLiMP Overallaccuracy: 84.2 -> 84.6No clear effect
Use SLM instead of RoBERTa largeTranscormer · BLiMP Overallaccuracy: 86.5 -> 84.6Hurt
Use the small configuration instead of the base configuration (ablation)Transcormer · LibriSpeech dev-cleanWER: 2.23 -> 2.48Hurt
Mask more tokens per MLM pass (k=2) (ablation)MLM rescoring · LibriSpeech dev-cleanWER: 2.30 -> 2.41Hurt
Mask more tokens per MLM pass (k=3) (ablation)MLM rescoring · LibriSpeech dev-cleanWER: 2.30 -> 2.60Hurt
Mask n/3 tokens per MLM pass (ablation)MLM rescoring · LibriSpeech dev-cleanWER: 2.30 -> 2.75Hurt
Mask n/2 tokens per MLM pass (ablation)MLM rescoring · LibriSpeech dev-cleanWER: 2.30 -> 2.80Hurt
Use stream-level instead of sentence-level pre-training (ablation)Transcormer · IWSLT14 De-EnBLEU: 35.24 -> 34.84Hurt
Use stream-level instead of sentence-level pre-training (ablation)Transcormer · LibriSpeech dev-cleanWER: 2.23 -> 2.56Hurt
Add in-domain pre-training data (ablation)Transcormer · IWSLT14 and WMT14 rerankingBLEU: 35.24 -> 35.74 (IWSLT14 De-En); 33.10 -> 33.51 (WMT14 De-En)Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy