Noisy memory encoding explains negative polarity illusions

Technical Summary: Noisy Memory Encoding Explains Negative Polarity Illusions

Problem Statement
The paper addresses the "negative polarity illusion," a behavioral phenomenon where ungrammatical sentences containing negative polarity items (NPIs) like ever are rated as more acceptable than their strictly ungrammatical controls, despite lacking a proper licensing environment. Specifically, in sentences like The authors that no critics recommended have ever received acknowledgment..., the NPI ever is structurally unlicensed because the negation no is embedded within a relative clause rather than in the main clause. While existing theories attribute this to failed memory retrieval or noisy pragmatic inference, they struggle to explain two critical empirical constraints: (1) the illusion is restricted to negative quantifiers (e.g., no, few) as licensors, and (2) the effect is sensitive to time pressure, appearing in speeded tasks but disappearing in untimed ones. The authors propose that these phenomena are better explained by the lossy-context surprisal theory (Futrell et al., 2020; Hahn et al., 2022), which posits that comprehenders maintain imperfect, "lossy" memory representations of complex sentences and rationally reconstruct missing or forgotten elements based on linguistic statistics.

Methodology
The study tests a novel prediction derived from the lossy-context theory: that the strength of the illusion is modulated by the similarity of determiners in the main-clause and embedded-clause subject positions. The hypothesis is that more similar determiners are more prone to being misremembered or exchanged in memory, thereby creating a non-veridical representation that licenses the NPI.

• Stimuli Design: The authors conducted six untimed offline acceptability judgment experiments. Each experiment tested a specific pair of determiners in the canonical illusion-prone structure: {few, many}, {few, most}, {few, all}, {few, the}, {no, many}, and {no, all}. The control condition for all experiments used the canonical pair {no, the}.
• Similarity Metric: Determiner similarity was quantified using cosine similarity scores derived from word embeddings (GloVe, fastText, and BERT). The authors calculated the arithmetic mean of cosine similarities across these three models to create a robust similarity score for each pair.
• Experimental Procedure: Participants (native English speakers from the US) rated sentence naturalness on a 7-point Likert scale. Each experiment included 36 critical items and 72 filler items. Crucially, unlike previous studies that relied on speeded tasks, these experiments were untimed, allowing participants ample time to encode the input.
• Statistical Analysis: The authors employed Bayesian multilevel cumulative ordinal regression models (using the brms package in R) to estimate the illusion effect. The illusion effect was defined as the posterior probability that the illusion-prone condition received higher ratings than the ungrammatical control condition.

Key Results

1. Illusion Strength and Determiner Similarity: The experiments revealed that determiner pairs with higher cosine similarity triggered significantly stronger illusion effects. Specifically, the pairs {few, many} and {few, most} (the most similar pairs) produced robust illusion effects even in the absence of time pressure. In contrast, the canonical control pair {no, the} (the least similar) did not produce a significant illusion effect, replicating previous findings that the effect disappears in untimed tasks for this specific structure.
2. Correlation: A descriptive Pearson correlation of 0.877 was observed between the cosine similarity of the determiner pairs and the estimated posterior illusion strength. This positive correlation supports the hypothesis that memory interference increases with the similarity of the linguistic elements being encoded.
3. Untimed Validation: The finding that strong illusions occur in untimed tasks for similar determiner pairs challenges theories that rely solely on time pressure or rapid retrieval failure, suggesting instead that the encoding process itself is inherently noisy and prone to interference regardless of processing speed.

Key Contributions

• Mechanism Identification: The paper provides empirical support for the lossy-context memory encoding account of syntactic illusions. It demonstrates that the illusion arises not merely from retrieval failure but from the initial, imperfect encoding of determiners, which are then rationally reconstructed based on context and statistical likelihood.
• Predictive Power: The study successfully validates a novel prediction of the lossy-context theory: that illusion strength is a function of the similarity between the determiners in the main and embedded clauses. This offers a quantitative metric (cosine similarity) for predicting illusion strength.
• Differentiation from Competing Theories: The results distinguish the lossy-context approach from retrieval-based theories (e.g., Vasishth et al., 2008) and contrastive inference accounts. The failure of retrieval theories to explain the specificity of negative quantifiers and the success of the similarity-based prediction in untimed conditions suggest that the source of the error lies in the encoding/maintenance phase of working memory.

Significance and Claims
The authors claim that these findings offer further evidence that human language processing is imperfect and resource-rational. In the face of working memory limitations, comprehenders do not simply fail; they rationally reconstruct the most probable representation from noisy linguistic input to facilitate downstream processing. The negative polarity illusion is thus framed not as an isolated grammatical anomaly, but as a revealing case study of how rational inference operates over noisy linguistic representations within a noisy-channel framework. The study suggests that the "illusion" is a byproduct of the brain's efficient strategy to handle complex dependencies by forgetting predictable function words and reconstructing them, a process that becomes more error-prone when the elements to be remembered are highly similar.