/on-isotropy-of-contrastive-SRL

Primary LanguageJupyter Notebook

On Isotropy, Contextualization and Learning Dynamics of Contrastive-based Sentence Representation Learning

"On Isotropy, Contextualization and Learning Dynamics of Contrastive-based Sentence Representation Learning"
ACL 2023 Findings

  1. metrics.py defines our implementations on computing 1) self-similarity 2) intra-sentence/document similarity 3)anisotropy estimation 4) rouge dimension 5) embedding norm

  2. run analysis_pipeline.py with these defined metrics to analyze your model at once. All results will be saved in a analysis_output.txt.

Run with customized arguments. For instance, run analysis on all-minilm-l6-v2, starting from layer 5 to the last layer: python analysis_pipeline.py --model_name sentence-transformers/all-MiniLM-L6-v2 --analyze_layer_start 5

  1. we provide the corresponding noteboook scripts analysis_pipeline.ipynb to play with these defined metrics step-by-step in to analyze all-mpnet-base-v2 as an example. Change MODEL_NAME in the notebook to analyze your model.

  2. In fun_example.ipynb you can see some intuitive and fun examples of our paper's findings:

"NLP" in "NLP is about" and "is" in "NLP is about" are similar (high intra-sentence similarity)
"NLP" in "NLP is about" and "NLP" in "What is NLP?" are similar (high self-similarity for semantic tokens)
"is" in "NLP is good" and "is" in "The album is fire" are not even close (low self-similarity for function tokens)