TL;DR
This paper investigates whether modern language models (LMs) exhibit "semantic attraction," a phenomenon where sentence processing is facilitated by a semantically plausible but syntactically unrelated word. The study finds that LMs, measured by surprisal, are sensitive to semantic attraction, reducing their prediction difficulty, though not always mirroring human processing patterns.
What Problem Does This Paper Address?
In psycholinguistics, "semantic attraction" describes how our brains can be misled by semantically plausible words that are not directly related to the grammatical structure of a sentence. This can lead to an "illusion of grammaticality," where a sentence that should be difficult to process becomes easier because of this attraction. For example, in "Julia saw the beer that the lady with the meal quite happily ate during an expensive night out," the word "meal" is semantically related to "ate" and "lady," making the sentence, which is otherwise implausible (beer isn't typically eaten), easier to process than if it were "Julia saw the beer that the lady with the wine quite happily ate..."
While previous research has shown this effect in human readers using eye-tracking, it's unclear if large language models (LMs), which are increasingly used to model human language processing, also exhibit this behavior. This paper aims to bridge this gap by using LMs to analyze a dataset of sentences designed to elicit semantic attraction and see if their "surprisal" scores (a measure of prediction difficulty) reflect this phenomenon.
Methods
The researchers employed a surprisal-based analysis using three autoregressive Transformer-based language models: GPT-2 Base, DistilGPT-2, and GPT-Neo. Surprisal, a measure of how unexpected a word is given its context, is calculated by LMs. The paper computed both sentence-level surprisal (SentSurp) and target phrase-level surprisal (Surp) for the critical sentences in the dataset. They then used linear mixed-effects models to analyze how sentence plausibility, attractor plausibility, and their interaction affected these surprisal scores.
Datasets
The study utilized a dataset derived from Experiment 1 of the eye-tracking study by Cunnings and Sturt (2018). This dataset consists of 128 stimuli, each comprising an introduction, a critical sentence, and a wrap-up sentence. For their analysis, the researchers focused solely on the critical sentences, which were constructed with four conditions by varying the plausibility of the direct object (e.g., "cake" vs. "beer") and the plausibility of an attractor noun (e.g., "meal" vs. "wine").
Key Insights and Ideas
- LMs are sensitive to semantic attraction: The models consistently showed reduced surprisal scores when a semantically plausible attractor was present, indicating they are influenced by this phenomenon, similar to humans.
- Attractor plausibility matters, especially at the target level: While sentence-level surprisal was less affected by the attractor, target phrase-level surprisal was significantly reduced when the attractor was plausible.
- Human-like interaction is rare: Unlike human readers, most LMs did not show a significant interaction between sentence plausibility and attractor plausibility. This suggests that while LMs can detect semantic attraction, they don't always replicate the nuanced way humans integrate this information, particularly in implausible sentences.
- Smallest model shows most human-like behavior: Interestingly, the smallest model, DistilGPT-2, exhibited a marginal interaction effect, suggesting that sometimes smaller models might capture certain human-like linguistic phenomena more effectively than larger ones, a concept related to "inverse scaling."
Key Results
- All models successfully distinguished between plausible and implausible sentences, assigning higher surprisal scores to implausible ones.
- For target phrase-level surprisal (Surp), models showed sensitivity to attractor plausibility, with plausible attractors leading to lower surprisal.
- For sentence-level surprisal (SentSurp), models were generally insensitive to attractor plausibility, with the exception of a marginal effect in DistilGPT-2.
- No significant interaction between sentence and attractor plausibility was found in most models for either Surp or SentSurp.
Limitations and Future Work
The authors acknowledge that none of the models perfectly matched human behavioral patterns, particularly regarding the interaction effect between sentence and attractor plausibility. They suggest that future work could explore techniques like contrastive explanations to understand how attractors concretely influence LM predictions, potentially shedding light on which tokens contribute to the prediction of the target verb. Analyzing how attractors concretely affect predictions, for example, using techniques like contrastive explanations, could provide deeper insights.
Summary updated