Research map / Vision-language and multimodal / Human and embodied modeling
SemLa: A Visual Analysis System for Fine-Grained Text Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SemLa, a visual analysis system with novel visualizations (interactive local words, stacked-bar safeguarded explanations, Sankey-style contrastive explanations) for interpreting fine-grained text classifiers, evaluated via case studies and expert interviews.
Problem
Interpreting fine-grained text classification models with many similar classes is hard; existing explanation tools and embedding visualizations are laborious or incompatible with many fine-grained classes and do not help spot disagreements between feature-importance methods.
Approach
Computes per-word minimal hypercube localities in the embedding space to show interactive local words that update with navigation; combines multiple feature importance metrics in a stacked bar (VIFI) to reveal disagreements; and provides example-based contrastive explanations against a foil class using Sankey-style links.
Key results
- Three industry experts strongly agreed SemLa was useful overall and helps understand individual predictions and identify sub-clusters within classes.
- For identifying model weaknesses and decision boundaries, two of three experts strongly agreed and one somewhat agreed the system was helpful.
- On ease of understanding, one expert strongly agreed visualizations were intuitive while two somewhat agreed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SemLa for visual analysis of fine-grained text classification | SemLa · Intent recognition | overall usefulness | Helped |
| Add example-based contrastive explanations | SemLa · Intent recognition | understanding individual predictions | Helped |
Limitations
- Evaluation limited to three experts and a single intent recognition use case.
- Local words currently use only lexical features; other feature types not explored.
- Usability ratings for ease of understanding were mixed among experts.
Experimental setup
Metrics: expert Likert-scale questionnaire (5-point, eight items)
Baselines: DeepNLPVis FIND LIME
Builds on: LIME integrated gradients DeepNLPVis FIND
Closest papers by meaning
- Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.85)
- The Devil is in the Fine-Grained Details: Evaluating Open-Vocabulary Object Detectors for Fine-Grained Understanding (CVPR 2024 · similarity 0.83)
- Advancing Fine-Grained Classification by Structure and Subject Preserving Augmentation (NeurIPS 2024 · similarity 0.82)
- Democratizing Fine-grained Visual Recognition with Large Language Models (ICLR 2024 · similarity 0.82)
- FG-CLIP: Fine-Grained Visual and Textual Alignment (ICML 2025 · similarity 0.82)
- Analyzing and Boosting the Power of Fine-Grained Visual Recognition for Multi-modal Large Language Models (ICLR 2025 · similarity 0.82)
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.