Research map / Label-efficient and robust learning / Recognition and long-tail
Training Temporal Word Embeddings with a Compass
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train slice-specific context embeddings against frozen atemporal target embeddings | CBOW temporal embeddings · Testset1 (NAC-S), All analogies | MRR: 0.422 -> 0.481 | Helped |
| Train slice-specific context embeddings against frozen atemporal target embeddings | CBOW temporal embeddings · Testset2 (NAC-L), All analogies | MRR: 0.449 -> 0.484 | Helped |
| Use compass-based training instead of static embeddings | Word embeddings on Testset1 (NAC-S) · Testset1 (NAC-S), dynamic analogies | MRR: 0.148 -> 0.394 | Helped |
| Use compass-based training instead of static embeddings | Word embeddings on Testset2 (NAC-L) · Testset2 (NAC-L), dynamic analogies | MRR: 0.102 -> 0.367 | Helped |
| Use compass-based training instead of TW2V | Temporal embeddings on held-out MLPC · MLPC held-out | posterior log probability: -2.83 -> -1.75 | Helped |
| Initialize the second training phase with both static weight matrices rather than only the frozen target matrix (ablation) | TWEC · Held-out evaluation | held-out performance | Helped |
| Initialize the second training phase with both static weight matrices rather than only the frozen target matrix (ablation) | TWEC · Temporal analogy tests | analogy performance | Hurt |
| Use compass-based training rather than pairwise alignment | Temporal embeddings trained on the small NAC-S corpus · Testset1 (NAC-S) | MRR: 0.481 vs 0.143; 0.481 vs 0.135 | Helped |
| Use compass-based training rather than pairwise alignment | Temporal embeddings trained on the larger NAC-L corpus · Testset2 (NAC-L) | MRR: 0.484 vs 0.444; 0.484 vs 0.449 | Helped |
| Use compass-based training rather than pairwise alignment | Temporal embeddings by analogy category · Testset2 (NAC-L), President of the USA and Super Bowl Champions | MP1: 0.8833 vs 0.9905; 0.1452 vs 0.2524 | Hurt |
Closest papers by meaning
- Dynamic Word Embeddings (ICML 2017 · similarity 0.88)
- Word2Fun: Modelling Words as Functions for Diachronic Word Representation (NeurIPS 2021 · similarity 0.87)
- Closed-Form Training Dynamics Reveal Learned Features and Linear Structure in Word2Vec-like Models (NeurIPS 2025 · similarity 0.85)
- What the Vec? Towards Probabilistically Grounded Embeddings (NeurIPS 2019 · similarity 0.85)
- A Simple but Tough-to-Beat Baseline for Sentence Embeddings (ICLR 2017 · similarity 0.85)
- Meaning Representations from Trajectories in Autoregressive Models (ICLR 2024 · 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.