AI Paper Atlas

Research map / Probabilistic and scientific modeling / Molecules and proteins

Measuring Compositionality in Representation Learning

Jacob Andreas

ICLR 2019 · Probabilistic and scientific modeling › Molecules and proteins · approach: Probabilistic embeddings and relational inference · 24 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
Measure phrase compositionality with TRE rather than assuming it from word vectorsFastText CBOW phrase embeddings · Noun-noun compound human judgmentsSpearman correlation with human compositionality ratings: ρ=−0.34, p<0.01Improved
Compare TRE across communication protocolsSpeaker-listener RNNs · Reference game, held-out referentscorrelation with generalization error: r=0.50, p<1e!−6Mixed
Compare TRE across communication protocolsSpeaker-listener RNNs · Reference game, all training runscorrelation with absolute reward: r=0.57, p<1e!−9Hurt
Track TRE alongside mutual information during learningFew-shot classifier · Color MNIST validation, training runscorrelation between TRE and mutual information: r=0.70, p < 1e!−10; r=0.71, p < 1e!−8Mixed
Restrict analysis to successful communication runsSpeaker-listener RNNs · Reference game, successful runs (held-out reward > 0.5)correlation between TRE and generalization error: r=0.6, p<1e!−3Mixed
Restrict analysis to successful communication runsSpeaker-listener RNNs · Reference game, successful runs (held-out reward > 0.5)correlation between TRE and absolute reward: r=0.38, p<0.05Hurt

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