How do readers selectively hold frame-related words in mind?
This explores how the human mind, while reading, locks onto words that fit a coherent frame and tunes out words that are nearby but irrelevant — and what that selective grip reveals that mechanical similarity-matching misses.
This explores how the human mind, while reading, locks onto words that fit a coherent frame and tunes out the linguistically adjacent ones that don't belong — a selective grip, not a sum of all the words on the page. The corpus's most direct answer is that meaning-making is *resonant*: the mind holds frame-related words in tight mutual activation while suppressing words that merely sit next to them in the sentence, and crucially this selectivity tracks frame-coherence rather than how often words co-occur Does the mind selectively activate frames from only some words?. A companion note pushes this further — meaning isn't compositional aggregation but the *live detection* of which subsets of words light up a shared frame, an operation that is selective, non-additive, and non-monotonic (adding a word can shut a frame down rather than enrich it) How do readers actually build meaning from words?.
What makes this interesting is that the holding is never just lexical — it runs on several layers at once. Discourse comprehension demands that a reader simultaneously track the linguistic segments, the speaker's intentions, and what's currently salient in attention; these three constrain each other in parallel, so which words you hold in mind is partly governed by which purpose is active and what's foregrounded How do readers track segments, purposes, and salience together?. That attentional layer is why the same trigger word can carry different weight in different spots: projection strength turns out to be gradient, rising or falling depending on whether the content speaks to the live Question Under Discussion rather than on any fixed property of the word itself Does projection strength vary by context or by word type?.
The selectivity also has a social dimension you might not expect. Which frame a reader resonates with isn't uniform across people — interpretations of socially loaded sentences split legitimately by the reader's position, so "holding frame-related words" is partly an act of where you're standing, not a single correct parse Why do readers interpret the same sentence so differently?. And the holding can be deliberately shifted: positive reframing neutralizes negativity while preserving the underlying content, which only works because it swaps the active frame without destroying meaning — unlike sentiment transfer, which flips both Does positive reframing preserve meaning better than sentiment transfer?.
The sharpest payoff comes from the contrast with machines, which is probably the thing you didn't know you wanted to know. Human selective resonance is precisely what current language models *can't* do. They struggle to ignore irrelevant material — needing explicit training to resist topical distractors, because they learn what-to-do instructions but not what-to-ignore instructions Why do language models engage with conversational distractors?. They can't comfortably hold multiple live interpretations at once, disambiguating only a third of cases where humans manage ninety percent Can language models recognize when text is deliberately ambiguous?. Their grip even weakens with sheer volume — reasoning degrades sharply just from padding the input, far below any context limit Does reasoning ability actually degrade with longer inputs?. One strand of the corpus argues this is structural: models trained on form alone never acquire the relation between expression and communicative intent that lets a human know which words matter Can language models learn meaning from text patterns alone?. So the human ability to hold only the frame-related words — to suppress, not just attend — looks less like a feature and more like the missing ingredient.
Sources 10 notes
Human meaning-making operates through selective frame activation: the mind holds frame-related words in tight resonance while ignoring linguistically adjacent but frame-unrelated words. This selectivity tracks frame-coherence, not co-occurrence frequency, and represents a cognitive operation that standard similarity computation cannot capture.
Meaning-making is the live detection of which word subsets activate shared frames, not compositional aggregation of individual word meanings. This operation is selective, non-additive, and non-monotonic, fundamentally different from how current AI processes language.
Discourse processing demands parallel recognition of linguistic segments, intentional structure, and attentional salience—not sequential processing. These three layers constrain each other during comprehension, and failures in any single layer disrupt overall understanding.
Across 19 English expressions, projectivity varies continuously based on whether content addresses the Question Under Discussion. The same presupposition trigger projects more or less depending on context, not on fixed lexical properties.
Interpretation Modeling research shows that disagreement on socially embedded sentences reflects valid differences in reader perspective, not annotation failure. Structured human disagreement in NLI benchmarks confirms that interpretation distributions carry meaningful information.
The POSITIVE PSYCHOLOGY FRAMES benchmark demonstrates that reframing neutralizes negativity while keeping original content intact, whereas sentiment transfer reverses both polarity and meaning. Reframing is semantically constrained and requires genuine understanding of complementary perspectives.
Fine-tuning on just 1,080 synthetic dialogues with distractor turns significantly improves topic resilience, revealing that the gap is not model capacity but absent training signal. Models learn to follow what-to-do instructions but not what-to-ignore instructions.
AMBIENT benchmark shows GPT-4 correctly disambiguates only 32% of cases versus 90% for humans. This failure spans lexical, structural, and scope ambiguity—revealing that LLMs cannot hold multiple interpretations simultaneously, a fundamental gap hidden by standard benchmarks.
FLenQA shows reasoning accuracy drops from 92% to 68% at just 3000 tokens of padding, far below context window capacity. The degradation is task-agnostic, uncorrelated with language modeling performance, and persists even with chain-of-thought prompting.
Bender & Koller argue that meaning requires the relation between expressions and communicative intents. Since LLMs are trained only on form-to-form prediction with no access to shared attention or intent, they cannot reconstruct the meaning that grounds language.