Summary of Reverse-engineering the Reader, by Samuel Kiegeland et al.
Reverse-Engineering the Reader
by Samuel Kiegeland, Ethan Gotlieb Wilcox, Afra Amini, David Robert Reich, Ryan Cotterell
First submitted to arxiv on: 16 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI); 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 This paper investigates whether language models, trained on natural language text, can be directly optimized to better mimic human cognition. The authors propose a novel technique that fine-tunes a language model to predict humans’ reading times of linguistic units (e.g., phonemes, morphemes, or words) using surprisal estimates derived from the language model. They evaluate this approach across various model sizes and datasets, finding improved psychometric predictive power for language models. However, they also observe an inverse relationship between psychometric performance and downstream NLP task performance as well as perplexity on held-out test data, which has been previously observed but not induced through alignment to psychometric data. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper tries to figure out if computers can be trained to think like humans. The researchers want to see if they can make a computer model better at understanding language by making it predict how long it takes for people to read certain words or phrases. They try this with different sizes and types of models, and find that it works! However, they also notice that the more accurate the model is at predicting reading times, the worse it does on other tasks like recognizing sentences. |
Keywords
» Artificial intelligence » Alignment » Language model » Nlp » Perplexity
