Summary of An Empirical Comparison Of Vocabulary Expansion and Initialization Approaches For Language Models, by Nandini Mundra et al.
An Empirical Comparison of Vocabulary Expansion and Initialization Approaches for Language Models
by Nandini Mundra, Aditya Nanda Kishore, Raj Dabre, Ratish Puduppully, Anoop Kunchukuttan, Mitesh M. Khapra
First submitted to arxiv on: 8 Jul 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper presents a novel approach to initializing language models for multilingual tasks, addressing the challenge of limited vocabulary coverage in tokenizers. By establishing that initializing within the convex hull of existing embeddings is a good starting point, the authors propose Constrained Word2Vec (CW2V), a simple method that doesn’t require cross-lingual embeddings. The study evaluates different initialization methods for expanding RoBERTa and LLaMA 2 across four languages and five tasks, showing that CW2V performs equally well or even better than more advanced techniques. Additionally, simpler approaches like multivariate initialization perform on par with these advanced methods, indicating the potential for efficient large-scale multilingual continued pretraining. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us understand how to make language models work better for different languages. Currently, these models do well in English but struggle in other languages. To fix this, people often fine-tune them for those languages. However, there’s a problem: the original model doesn’t have enough words to represent new languages correctly. The authors of this paper came up with a simple solution that works just as well as more complicated methods. They also tested their idea on several tasks and languages, showing that it really does work. |
Keywords
» Artificial intelligence » Llama » Pretraining » Word2vec
