Research map / Label-efficient and robust learning / Recognition and long-tail
Non-Linear Similarity Learning for Compositionality
Masashi Tsubaki, Kevin Duh, Masashi Shimbo, Yuji Matsumoto
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 |
|---|---|---|---|
| replace cosine similarity with polynomial kernel (ablation) | SUM + STR SEM · SICK test (SemEval 2014 Task 1) | Pearson r: 0.7647 -> 0.8480 | Improved |
| replace cosine similarity with RBF kernel (ablation) | SUM (bag of words) · SICK test (SemEval 2014 Task 1) | Pearson r: 0.7588 -> 0.8339 | Improved |
| add STR SEM structure features to SUM (ablation) | Poly kernel · SICK test (SemEval 2014 Task 1) | Pearson r: 0.8332 -> 0.8480 | Improved |
| add STR POS structure features to SUM (ablation) | Poly kernel · SICK test (SemEval 2014 Task 1) | Pearson r: 0.8332 -> 0.8301 | No clear effect |
| add STR POS structure features to SUM (ablation) | Cosine kernel · SICK test (SemEval 2014 Task 1) | Pearson r: 0.7588 -> 0.7510 | No clear effect |
| use tanh non-linear composition instead of linear (ablation) | Poly (SUM + STR SEM) · SICK test (SemEval 2014 Task 1) | Pearson r: 0.8480 -> 0.8392 | Hurt |
| use tanh non-linear composition instead of linear (ablation) | RBF (SUM + STR SEM) · SICK test (SemEval 2014 Task 1) | Pearson r: 0.8447 -> 0.8393 | Hurt |
| use tanh non-linear composition instead of linear (ablation) | Cosine (SUM + STR SEM) · SICK test (SemEval 2014 Task 1) | Pearson r: 0.7647 -> 0.7717 | Improved |
| polynomial kernel instead of cosine on MSE (ablation) | SUM · SICK test (SemEval 2014 Task 1) | MSE: 0.4820 -> 0.3205 | Improved |
| initialize word vectors with LSA instead of GloVe or NLM (ablation) | Poly (SUM) · SICK test (SemEval 2014 Task 1) | Pearson r (Figure 2) | Hurt |
| random initialization instead of pre-trained GloVe/NLM (ablation) | Poly (SUM) · SICK test (SemEval 2014 Task 1) | Pearson r (Figure 2) | No clear effect |
| proposed model vs feature-engineered SemEval systems | Poly (SUM + STR SEM) · SICK test (SemEval 2014 Task 1) | Pearson r: 0.8280 -> 0.8480 | Improved |
Closest papers by meaning
- Measuring Compositionality in Representation Learning (ICLR 2019 · similarity 0.85)
- Joint Word Representation Learning Using a Corpus and a Semantic Lexicon (AAAI 2016 · similarity 0.85)
- Narrowing the Gap between Supervised and Unsupervised Sentence Representation Learning with Large Language Model (AAAI 2024 · similarity 0.84)
- Siamese Recurrent Architectures for Learning Sentence Similarity (AAAI 2016 · similarity 0.84)
- On Affine Homotopy between Language Encoders (NeurIPS 2024 · similarity 0.83)
- Consistency of Compositional Generalization Across Multiple Levels (AAAI 2025 · 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.
