livnlp/cosine-discounting

Solving Cosine Similarity Underestimation between High Frequency Words by ℓ2 Norm Discounting

• Proceedings of The 61st Annual Meeting of the Association for Computational Linguistics, Toronto, Canada, 2023
• Supervised by Professor Danushka Bollegala, head of NLP and machine learning research group at the University of Liverpool
• Co-author: Yi Zhou

research paper

• Link to the paper (arXiv)

oblectives:

• Investigate the impact of word frequency on cosine similarity of contextualized word embedding of masked language models such as BERT
• Mitigate the impact of word frequency on cosine similarity underestimation of contextualized word embedding in high-frequency words

Publication: ACL 2023

1. prepare BookCorpus:

• download and save BookCorpus dataset

2. word embedding and BERT properties collection:

• collect word embedding, word frequency, and l2-norm from the entire BookCorpus

3. word frequency VS l2_norm:

• plots between word frequency and l2-norm

4. WIC preprocess:

• Process WiC by adding word frequency, default BERT cosine similarity, and BERT word embedding

5. parameters of the classifier:

• Bayesian optimization to find the theta of the classifier

6. WIC 5 cross validation dataset prep:

• Prepare 5 cross-validations dataset from WiC

7. l2_norm discounting:

• Cosine similarity prediction of the l2-norm discounting method compared to default BERT

University of Liverpool thesis

part 1: data collection

• word embedding and BERT properties collection
• distributions of words

part 2: relationships between word frequency and BERT's properties

• word frequency VS local l2-norm
• word frequency VS global l2-norm
• word frequency VS l2-norm variance
• word frequency VS global isotropy
• word frequency VS global self-similarity

part3: approaches to mitigate the impact of word frequency on under and over cosine similarity estimation of BERT

• BERT baseline
• approach1: a single linear adjustment
• approach2: two-way linear adjustment
• approach3: word type-dependent adjustment
• approach4: probabilistic word type-dependent adjustment
• approach5: z-score normalization

cosine-discounting's Issues

request for some word properties file

Hey author!

I want to load word properties file in "Pulication/ano3 word frequency VS l2_norm.ipynb/Dataset preparation", but i have some problem, can you provided the file of word properties file ?

Thank you so much!