Skip to content

Novel approach to analyze how different layers of BERT encode linguistic information

Notifications You must be signed in to change notification settings

Kash6/ProbelessProbing

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 

Repository files navigation

ProbelessProbing

🧠 Probe-less Probing of BERT’s Layer-Wise Linguistic Knowledge

This project explores a novel approach called Probe-less Probing to analyze how different layers of BERT encode linguistic information. Rather than relying on external classifiers, we use masked word prediction as a behavioral signal to probe BERT’s internal understanding.

🔍 What We Did

Reimplemented the study using the STREUSLE 4.4 dataset, which includes rich lexical-semantic and syntactic annotations.

Ran layer-wise masked token prediction using BERT-base (12 layers).

Evaluated BERT’s ability to:

Predict the correct word.

Predict the correct part-of-speech (POS).

Analyzed accuracy trends across layers and calculated expected layers for different POS tags and multiword expressions using differential scoring.

📊 Key Findings

BERT achieves ~43.27% accuracy for masked word prediction and ~67.08% accuracy for POS tagging.

POS and word prediction accuracies increase with layer depth, peaking around layer 6–7.

Multiword expressions (MWEs) show different behavior, with later layers often performing better.

Our reproduction sometimes outperformed the original results, possibly due to model or tokenizer updates.

⚙️ Technologies Used

Hugging Face Transformers

Google Colab

Python (NumPy, pandas)

Stanza (for POS tagging)

Custom differential gain analysis

📁 Contributions

Akash Mehta: Analysis and computation of expected layers for various linguistic features across layers; result comparison.

Rithvik Reddy Sama: POS, word match accuracy, and layer gain calculations.

About

Novel approach to analyze how different layers of BERT encode linguistic information

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors