Entropy- and Distance-Based Predictors From GPT-2 Attention Patterns Predict Reading Times Over and Above GPT-2 Surprisal

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9324 - 9334 December 7-11, 2022 ©2022 Association for Computational Linguistics Entropy- and Distance-Based Predictors From GPT-2 Attention Patterns Predict Reading Times Over and Above GPT-2 Surprisal Byung-Doh Oh Department of Linguistics The Ohio State University oh.531@osu.edu William Schuler Department of Linguistics The Ohio State University schuler.77@osu.edu Abstract Transformer-based large language models are trained to make predictions about the next word by aggregating representations of previous to- kens through their self-attention mechanism. In the field of cognitive modeling, such attention patterns have recently been interpreted as em- bodying the process of cue-based retrieval, in which attention over multiple targets is taken to generate interference and latency during re- trieval. Under this framework, this work first defines an entropy-based predictor that quan- tifies the diffuseness of self-attention, as well as distance-based predictors that capture the incremental change in attention patterns across timesteps. Moreover, following recent studies that question the informativeness of attention weights, we also experiment with alternative methods for incorporating vector norms into attention weights. Regression experiments us- ing predictors calculated from the GPT-2 lan- guage model show that these predictors deliver a substantially better fit to held-out self-paced reading and eye-tracking data over a rigorous baseline including GPT-2 surprisal. Addition- ally, the distance-based predictors generally demonstrated higher predictive power, with ef- fect sizes of up to 6.59 ms per standard devia- tion on self-paced reading times (compared to 2.82 ms for surprisal) and 1.05 ms per standard deviation on eye-gaze durations (compared to 3.81 ms for surprisal). 1 Introduction Much work in broad-coverage sentence processing has focused on studying the role of expectation op- erationalized in the form of surprisal (Hale, 2001; Levy, 2008) using language models (LMs) to de- fine a conditional probability distribution of a word given its context (Smith and Levy, 2013; Goodkind and Bicknell, 2018). Recently, this has included Transformer-based language models (Wilcox et al., 2020; Merkx and Frank, 2021; Oh et al., 2022). Figure 1: The predictors and attention weight formu- lations examined in this work. The entropy-based pre- dictor (i.e. NAE) quantifies the diffuseness of atten- tion over previous tokens at a given timestep, while the distance-based predictors (i.e. ∆NAE, MD, EMD) cap- ture the change in attention patterns across consecutive timesteps (top row: weights at ‘journey,’ bottom row: weights at ‘to’). These predictors can be calculated from attention weights formulated using different methods (i.e. ATTN-W, ATTN-N, ATTNRL-N). However, expectation-based accounts have em- pirical shortcomings, such as being unable to fully account for garden-path effects (van Schijndel and Linzen, 2021) or predict the timing of delays in certain constructions (Levy et al., 2013). For this reason, some research has begun to focus on the effects of memory and attention using predictors calculated from language model representations. For example, Ryu and Lewis (2021) recently drew connections between the self-attention patterns of Transformers (Vaswani et al., 2017) and cue-based retrieval models of sentence comprehension (e.g. Lewis et al., 2006). Their attention entropy, …

Content truncated to 500 words