Garden Path Traversal in GPT-2

William Jurayj, William Rudman, Carsten Eickhoff


Abstract
In recent years, large-scale transformer decoders such as the GPT-x family of models have become increasingly popular. Studies examining the behavior of these models tend to focus only on the output of the language modeling head and avoid analysis of the internal states of the transformer decoder. In this study, we present a collection of methods to analyze the hidden states of GPT-2 and use the model’s navigation of garden path sentences as a case study. To enable this, we compile the largest currently available dataset of garden path sentences. We show that Manhattan distances and cosine similarities provide more reliable insights compared to established surprisal methods that analyze next-token probabilities computed by a language modeling head. Using these methods, we find that negating tokens have minimal impacts on the model’s representations for unambiguous forms of sentences with ambiguity solely over what the object of a verb is, but have a more substantial impact of representations for unambiguous sentences whose ambiguity would stem from the voice of a verb. Further, we find that analyzing the decoder model’s hidden states reveals periods of ambiguity that might conclude in a garden path effect but happen not to, whereas surprisal analyses routinely miss this detail.
Anthology ID:
2022.blackboxnlp-1.25
Volume:
Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP
Month:
December
Year:
2022
Address:
Abu Dhabi, United Arab Emirates (Hybrid)
Editors:
Jasmijn Bastings, Yonatan Belinkov, Yanai Elazar, Dieuwke Hupkes, Naomi Saphra, Sarah Wiegreffe
Venue:
BlackboxNLP
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
305–313
Language:
URL:
https://aclanthology.org/2022.blackboxnlp-1.25
DOI:
10.18653/v1/2022.blackboxnlp-1.25
Bibkey:
Cite (ACL):
William Jurayj, William Rudman, and Carsten Eickhoff. 2022. Garden Path Traversal in GPT-2. In Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 305–313, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.
Cite (Informal):
Garden Path Traversal in GPT-2 (Jurayj et al., BlackboxNLP 2022)
Copy Citation:
PDF:
https://aclanthology.org/2022.blackboxnlp-1.25.pdf
Video:
 https://aclanthology.org/2022.blackboxnlp-1.25.mp4