Summary of Interpretability Of Language Models Via Task Spaces, by Lucas Weber et al.
Interpretability of Language Models via Task Spaces
by Lucas Weber, Jaap Jumelet, Elia Bruni, Dieuwke Hupkes
First submitted to arxiv on: 10 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 This paper presents an alternative approach to interpreting language models (LMs) by focusing on their processing quality and linguistic abilities. Instead of relying solely on benchmark performances, the authors create “linguistic task spaces” that represent an LM’s conceptualization of language, revealing connections between language phenomena. The similarity probing method is used to assess interactions between learning signals from different linguistic phenomena. A fine-tuning via gradient differentials (FTGD) approach is also introduced to disentangle these learning signals. The authors apply their methods to LMs of varying scales and find that larger models generalize better to overarching concepts, utilizing shared structure more effectively. Additionally, the distributedness of linguistic processing increases with pre-training, as parameter sharing between related tasks grows. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Language models are usually tested on different benchmarks to understand how they process language. But what if we looked at how well they can connect ideas and concepts? This paper takes a new approach by creating special spaces that show how language models think about language. By looking at these “linguistic task spaces”, researchers can learn more about the connections between different language phenomena, like grammar and syntax. The authors also develop a way to see what learning signals are driving these connections. They test their methods on three different-sized language models and find that larger ones are better at understanding broader concepts. This could be because they share more ideas and knowledge with each other. |
Keywords
» Artificial intelligence » Fine tuning » Syntax
