Research map / Large language models / Language, safety and interpretability
Improving Natural Language Processing Tasks with Human Gaze-Guided Neural Attention
Ekta Sood, Simon Tannert, Philipp Mueller, Andreas Bulling
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 |
|---|---|---|---|
| Integrate TSM gaze predictions into task attention and jointly adapt the TSM | Paraphrase generation · Quora Question Pairs, 100K | BLEU-4: 17.9 -> 28.82 | Improved |
| Integrate TSM gaze predictions into task attention and jointly adapt the TSM | Paraphrase generation · Quora Question Pairs, 50K | BLEU-4: 16.5 -> 26.24 | Improved |
| Integrate TSM gaze predictions into task attention | Sentence compression · Google Sentence Compression | F1: 85.1 vs. 85.0 | No clear effect |
| Remove TSM fixation guidance (ablation) | Joint task models · Quora Question Pairs, 50K and 100K | BLEU-4: 26.24 -> 24.62; 28.82 -> 27.81 | Hurt |
| Replace hybrid-pretrained TSM initialization with random initialization (ablation) | Joint task models · Quora Question Pairs and Google Sentence Compression | BLEU-4 / F1: 26.24 -> 25.26; 28.82 -> 27.11; 85.0 -> 83.7 | Hurt |
| Swap TSM weights between the upstream tasks (ablation) | Joint task models · Quora Question Pairs and Google Sentence Compression | BLEU-4 / F1: 26.24 -> 23.43; 28.82 -> 27.60; 85.0 -> 83.8 | Hurt |
| Freeze TSM weights during joint task training (ablation) | Joint task models · Quora Question Pairs and Google Sentence Compression | BLEU-4 / F1: 26.24 -> 25.73; 28.82 -> 28.26; 85.0 -> 83.9 | Hurt |
| Pre-train the TSM on E-Z Reader synthetic data before human-gaze training (ablation) | Text Saliency Model · Provo | Spearman's ρ: 0.83* -> 0.88* | Improved |
| Pre-train the TSM on E-Z Reader synthetic data before human-gaze training (ablation) | Text Saliency Model · MQA-RC | Spearman's ρ: 0.76* -> 0.94* | Improved |
| Remove pre-training on E-Z Reader synthetic data (ablation) | Text Saliency Model · Dundee, Provo + Geco, Geco | Spearman's ρ: 0.99* -> 0.99*; 1.00* -> 0.99*; 0.99* -> 0.98* | No clear effect |
| Remove fine-tuning on human gaze data (ablation) | Text Saliency Model · Dundee, Provo + Geco, Provo, Geco, MQA-RC | Spearman's ρ: 0.99* -> -0.68; 1.00* -> 0.10; 0.88* -> -0.25; 0.99* -> -0.10; 0.94* -> -0.05 | Hurt |
| Add E-Z Reader synthetic-data pre-training (ablation) | Text Saliency Model · Provo | JSD: 0.44 to 0.24 | Improved |
Closest papers by meaning
- Punctuation-aware Hybrid Trainable Sparse Attention for Large Language Models (NeurIPS 2026 · similarity 0.85)
- Head Pursuit: Probing Attention Specialization in Multimodal Transformers (NeurIPS 2025 · similarity 0.84)
- MVA: Linear Attention with High-order Query-Keys Integration and Multi-level Vocabulary Decomposition (ICML 2025 · similarity 0.84)
- Seeing Eye to AI: Human Alignment via Gaze-Based Response Rewards for Large Language Models (ICLR 2025 · similarity 0.84)
- SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking (NeurIPS 2026 · similarity 0.84)
- Hymba: A Hybrid-head Architecture for Small Language Models (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.
