Source-linked AI summary
Hey, guess what: CUES lead to inferences about typicality
Alison Chi, Sanghyeok Choi, Hannah Rohde, Emily Allaway
arXiv:2610.33239v1cs.CL
TL;DR
Event typicality can shape how listeners interpret meaning beyond what speakers say, but evidence on these inferences in humans and LLMs has been limited. The paper builds two event corpora and tests humans and LLMs. Humans’ inferences track three cues, while LLMs often lack the commonsense knowledge needed to follow them reliably.
Problem
Evidence on event-typicality inferences has been limited, although these inferences draw on commonsense knowledge and reasoning in interpreting conversation.
Source: §1 · p. 1Method
Results
Takeaways & Limitations
The findings identify commonsense event knowledge as a major hurdle for LLMs’ event-typicality inferences in the tested settings.
Source: §1 · p. 2, §8 · p. 8The experiments tested only one form of utterance emphasis and a single minimal dialogue context, used textual data that excludes prosody, and recruited US-based participants for English-developed materials.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Throughout life, we accumulate knowledge of how typical certain events are. We use this knowledge in daily conversation, often interpreting meaning beyond what has been said explicitly. In this work, we take steps to better understand this meaning recovery, investigating both atypicality and typicality inferences. First, we construct two corpora, EventTyp and the human-validated EventTypGold. We conduct large-scale human and LLM experiments and find that for humans, three CUES predict inferences about typicality: Commonsense Knowledge, Utterance Emphasis, and Snark. However, in most experimental settings, we find that LLMs lack the prerequisite Commonsense Knowledge necessary to accurately assess an event's general typicality, and therefore struggle to follow these CUES.
1 Introduction
The paper studies how event typicality shapes meaning beyond what speakers explicitly say, extending limited prior human and LLM research. It introduces two corpora and reports that human typicality inferences track three cues, while LLMs often lack the commonsense knowledge needed to follow them.
1 Introduction
- Mentioning a typically inferable event can seem uninformative, inviting an atypicality inference; mentioning an atypical event can invite a typicality inference.The authors distinguish these: only atypicality inferences are claimed to require pragmatic reasoning; typicality inferences are treated as semantic belief updates.Source: §1 · p. 1
- Prior work has examined these inferences only in limited human experiments, with even less research on LLMs.The paper frames the task as challenging because it draws on both commonsense knowledge and reasoning ability.Source: §1 · p. 1
- The authors create EventTyp, with 2,354 labeled event tuples, and EventTypGold, a human-validated subset of 858 tuples.Their experiments involve about 900 human participants and 10 LLMs tested with three inference methods.Source: §1 · p. 2
- Humans’ typicality inferences are invited by Commonsense Knowledge, Utterance Emphasis, and Snark.Snark is described as an atypicality inference that helps a listener learn negative information about someone.Source: §1 · p. 2
- LLMs mostly disagree with humans about commonsense event typicality, and cue-following is limited mainly to large models with Chain-of-Thought prompting.The authors identify insufficient commonsense knowledge as a hurdle for LLM event-typicality inferences.Source: §1 · p. 2
2 Related Work
Prior research has studied event plausibility and pragmatic reasoning, but work on fine-grained event typicality and LLM inferences remains limited.
2 Related Work
- Prior event-likelihood studies distinguish plausibility levels, but the authors identify no prior work comparing the typicality of two highly plausible events.Examples include plausible versus implausible events and a probable-to-inconceivable scale.Source: §2 · p. 2
- Earlier LLM research on event-typicality inferences used two model families and 24 utterance tuples, and observed no atypicality inferences.The paper contrasts this limited scale with its own experiments.Source: §2 · p. 2
- Experiments asking participants to rate event typicality in dialogue snippets found atypicality inferences for typical events, but none for atypical events; subsequent work confirms these inferences reflect Quantity-driven reasoning.Source: §2 · p. 2
- Prior work reports a persistent human–LLM gap in pragmatic reasoning (Hu et al., 2023; Ma et al., 2025).Research has also examined LLM event knowledge and argument distinctions across different levels of event plausibility.Source: §2 · p. 2
3 The EventTyp Corpus
EventTyp pairs a context with typical and atypical follow-up events, using filtering, generation, and validation steps to build a 2,354-tuple corpus.
3 The EventTyp Corpus
- EventTyp contains 2,354 (Context, TypicFollowUp, AtypicFollowUp) tuples, with follow-up events generated by an LLM and atypical follow-ups dependent on both the context and typical follow-up.The final corpus incorporates 24 tuples from Kravtchenko and Demberg (2015).Source: §3 · p. 3, §3.4 · p. 4
3.1 Context Event Extraction and Filtering
- The authors extracted 33,924 utterances from ATOMIC isBefore and isAfter event pairs, then filtered these to 12,978 Context events.The extracted pairs were limited to utterances with up to two entities and filtered for noisy or unsuitable events.Source: §3.1 · p. 3
3.2 Typical Follow-ups (TypicFollowUps)
- An LLM generated up to three typical follow-ups per Context, producing 19,827 candidate pairs.The resulting candidates were subsequently evaluated against manually developed criteria.Source: §3.2 · p. 3, §3.3 · p. 3
3.3 Atypicality Inference Event Criteria
- Manual annotation yielded two criteria for atypicality-inference triggers: follow-ups must be natural dialogue and naturally typical without being entailed by the context.LLM filtering used 27 true/false questions, rejected a candidate if any rule was violated, and favored precision over recall.Source: §3.3 · p. 3
3.4 Atypical Follow-ups (AtypicFollowUps)
- LLM filtering retained 4,155 candidate pairs; after manual validation and correction, the authors kept 4,113 tuples and subsampled 2,330 to limit the corpus to one follow-up pair per Context.The human-annotated sample showed modest LLM–human agreement (Cohen’s κ = .4532); the final total includes 24 prior tuples.Source: §3.4 · p. 4, §3.3 · p. 4
4 Human and LLM Experiment Details
The experiments collect likelihood judgments for context–follow-up pairs from humans and LLMs, using surprisal or prompted ratings to assess responses.
4 Human and LLM Experiment Details
- Likelihood ratings for follow-up events were collected for both the Commonsense Knowledge and Event-Typicality Inference experiments.Human participants rated batches of context–follow-up pairs, with three ratings per pair in the first experiment and five in the second.Source: §4 · p. 4, §4.1 · p. 4
4.1 Human Participant Details
- The study recruited 330 participants for the Commonsense Knowledge experiment and 575 for the Event-Typicality Inference experiment.Participants were recruited in the United States because the source data contained US-specific references.Source: §4.1 · p. 4
4.2 LLM Details
- The LLM evaluation covered 28 configurations and three inference methods: surprisal, direct prompting, and Chain-of-Thought prompting.Surprisal was computed over completion tokens given the input; prompting elicited likelihood ratings, with instruction-tuned models used for the prompting methods.Source: §4.2 · pp. 4–5
5 Commonsense Knowledge Experiment
The experiment compares human and LLM likelihood judgments of typical and atypical follow-ups, finds generally low human–LLM agreement, and uses strict human ratings to form EventTypGold.
5.1 Experimental Setup
- For every EventTyp tuple, humans and LLMs rated generated, manually checked questions for both follow-ups; four ratings were collapsed to likely/unlikely and Cohen’s κ calculated separately.Source: §5 · p. 5, §5.1 · p. 5
5.2 RQ1: Do humans and LLMs share Commonsense Knowledge about event typicality?
- Only GPT-5.4 with CoT exceeded 0.4 agreement on both follow-ups (κTypicFollowUp = 0.4336; κAtypicFollowUp = 0.5059), while agreement was generally low.Source: Figure 3 · p. 5, §5.2 · p. 5
5.3 The EventTypGold Corpus
- The TypicFollowUp was rated more likely than the AtypicFollowUp for 80.04% of generated tuples and 95.83% of K&D tuples.Source: §5.3 · p. 5
- Table 3 defines six experiment conditions, each showing a participant a context and one follow-up setting.Source: Table 3 · p. 6
- With at least two-thirds agreement on LIKELY TypicFollowUp and UNLIKELY AtypicFollowUp, 36.09% (841) generated and 70.83% (17) K&D tuples qualified; 858 formed EventTypGold.Source: §5.3 · p. 6
6 Event-Typicality Inference Experiment
The experiment compared human and LLM likelihood judgments for dialogue follow-ups when they were unmentioned, mentioned plainly, or introduced with emphasis. Humans showed different typicality-inference patterns by event type and emphasis, while few LLM configurations produced both inference types significantly.
6.1 Experimental Setup
- The study collected likelihood ratings from humans and LLMs on 450 EventTypGold dialogue scenarios, comparing unmentioned, plain-mention, and HeyGW follow-up conditions.HeyGW preceded the follow-up with “And hey, guess what”; NONE served as the baseline likelihood given the context.Source: §6 · p. 6, §6.1 · p. 6
6.2 RQ2: When do humans make inferences about event typicality?
- For TypicFollowUps, humans made consistent atypicality inferences only with Utterance Emphasis; without it, typicality inferences were more frequent.The authors note that this differs from Kravtchenko and Demberg (2022), potentially because their verbose backstories explicitly mentioned typicality, whereas these scenarios used short dialogue.Source: §6.2 · p. 7, Figure 4 · p. 7
- For TypicFollowUps in HeyGW, humans made more atypicality inferences in snarky scenarios than in non-snarky ones.Snarky scenarios were manually annotated where an atypicality inference could reveal a negative trait and support interpreting the speaker as snarky.Source: §6.1 · p. 6, §6.2 · p. 7, Figure 5 · p. 7
- Humans made typicality inferences for AtypicFollowUps with or without emphasis, and for TypicFollowUps without emphasis.The authors describe the AtypicFollowUp pattern as belief updating after learning an initially unexpected event happened, rather than a pragmatic inference.Source: §6.2 · p. 7, Figure 4 · p. 7
6.3 RQ3: Do LLMs make inferences about event typicality in the contexts where humans do?
- 6 of 28 LLM configurations produced statistically significant typicality and atypicality inferences; 4 of those used Chain-of-Thought prompting on larger models.The analysis focused on HeyGW, where humans made both inference types in the expected event contexts.Source: Figure 6 · p. 8, §6.3 · p. 8
- Among larger LLM configurations that showed atypicality inferences, most also showed increased atypicality inferences for snarky scenarios.Source: §6.3 · p. 8
7 Discussion
LLMs sometimes follow the same CUES as humans, but inconsistent commonsense judgments complicate that pattern. Higher commonsense knowledge is associated with more event-typicality inferences, while ratings of the same events vary across prompt framings.
7 Discussion
- Higher commonsense knowledge significantly correlates with more event-typicality inferences (p<0.01), indicating that it is an essential component of these inferences.Source: §7 · p. 8
- No LLM configuration except GPT-5.4 with CoT scored above 80% consistency on both TypicFollowUps and AtypicFollowUps, and LLM consistency was low compared with humans.Fig. 7 compares likelihood ratings for the same follow-up events across the commonsense-knowledge and event-typicality-inference experiments in the comparable NONE setting.Source: §7 · p. 8
- For the same follow-up events, adding dialogue and named entities coincided with inconsistent LLM ratings, suggesting sensitivity to prompt framing rather than event typicality.The passage relates this pattern to prior findings of inconsistent responses when prompt wording changes.Source: §7 · p. 8
8 Conclusion
The paper presents two corpora and large-scale experiments, finding three cues that predict human event-typicality inferences. LLMs sometimes respond to those cues, but commonsense knowledge about general event typicality remains a major hurdle.
8 Conclusion
- The study created EventTyp and EventTypGold and reports investigating event-related typicality inferences on a larger scale than previous work.The corpus-creation process is recounted to enable reproducibility.Source: §8 · p. 8
- Human experiments identified Commonsense Knowledge, Utterance Emphasis, and Snark as cues that predict event-typicality inferences.Source: §8 · p. 8
- LLMs sometimes respond to the same cues, but predicting an event's general typicality with commonsense knowledge is a bigger hurdle than expected.Source: §8 · p. 8
- Future work could examine verb aspect and prosodic cues in humans and investigate why LLMs fail to display commonsense knowledge consistently and how to mitigate this.Source: §8 · p. 8
Limitations
This section outlines limits on generalizing the experiments across contexts, event types, speakers, and LLM families, and notes that text-only stimuli leave prosody untested. It also describes constraints and validation issues in generating and filtering typical and atypical follow-ups, including documented removals from the released data.
Limitations
- The experiments tested only HeyGW emphasis and minimal dialogue contexts, leaving other emphasis and context interventions and prosodic effects unexamined.The authors specifically note that adding “!” and other forms of context remain to be tested at scale for humans and LLMs.Source: Limitations · p. 9
- Test events were limited to relatively frequent events in an average US person’s life that finish within a day, so generalization to other event types is unknown.The authors leave open whether other types of events trigger typicality-related inferences in the same patterns.Source: Limitations · p. 9
- Human experiments sampled US-based participants and English-developed criteria, limiting representation of English speakers’ diversity and leaving cross-language inference patterns uncertain.Event-typicality inferences may differ for speakers of languages other than English.Source: Limitations · p. 9
- LLM experiments covered five model families and only one closed-source model, GPT-5.4, so behavior of other families and closed-source models remains unknown.The authors note varied behavior across tested models and name Mistral and DeepSeek as untested families.Source: Limitations · p. 9
A.1 Cooperativity & Quantity
- Cooperative-communication expectations make apparent underinformativity invite inferences, including that an explicitly reported event is atypical for the individual.Atypicality reasoning treats a seemingly redundant contribution as newsworthy and infers that the person’s usual behavior differs from the reported act.Source: §A.1 · p. 12, §A.2 · p. 12
A.3 Typicality inferences
- Typicality inferences are framed as belief updates rather than pragmatic inferences: learning an atypical event may make it seem more typical for that individual.The example describes learning that someone cried while thanking her boss, prompting an update about how typical crying is for her.Source: §A.3 · p. 12
B.1 Context Event Extraction
- The corpus drew context events from ATOMIC, but its event relations could not directly supply the desired follow-ups, so events were pooled to generate follow-ups.The authors cite insufficient suitable typical follow-ups, uneven follow-up-type mappings, and the goal of pairing each context with both follow-up types.Source: §B.1 · p. 13
B.2 Context Event Filtering
- Rules-based and GPT-4o filtering removed malformed, vague, context-dependent, and extreme events, yielding 12,978 contexts.The easy and hard passes applied complementary exclusion criteria to the ATOMIC data.Source: §B.2 · p. 13, §B.2.1 · p. 13, §B.2.2 · p. 14
B.3 Typical Follow-ups (TypicFollowUps)
- TypicFollowUps prompts GPT-4o to generate a state of being, subaction, and next action, increasing diversity over Kravtchenko and Demberg’s (2015) corpus; each item may be None.The authors use “event” for simplicity even though some state-of-being utterances may not be events.Source: §B.3 · p. 15
- The LLM assessed causality especially poorly and inconsistently, so Explanation and Result prompts had no filtering effect because their answers usually contradicted one another.The authors therefore removed the answers to those prompts.Source: §B.3 · p. 16
- For filtering typical follow-ups, the authors used a “When..., how often...?” question because its frequency interpretation made questions about events without typicality sound infelicitous.This differed from the “In general, when...how likely” format used in later experiments.Source: §B.3 · p. 16
- Typical follow-up questions use a “When...how often” format to ask how often a described event occurs.Source: §B.3 · p. 17
B.4 Atypical Follow-ups (AtypicFollowUps)
- Errors in some state and subaction generation examples may have degraded LLM performance through tokenization, but manual filtering was expected to prevent downstream problems.The errors involved examples not being joined properly with newline characters.Source: §B.4 · p. 17
- Atypical-follow-up prompts constrained generated states to be temporary and specific, and generated subactions to be more interesting while staying close to the original length.The state prompt allowed the state of knowing as a notable exception to the temporary-state rule.Source: §B.4 · p. 17
- Atypical follow-ups were manually validated to limit error propagation, with one author reviewing nearly 4,000 generated events versus over 12,000 typical follow-ups.During this process, some previously filtered TypicFollowUps were corrected; tuples with problematic contexts that could not be corrected were removed, including cases found in later validation, and the released data records filtering stages and retroactive removals.Source: §B.4 · p. 18
B.5 Data Sample
The paper details how it sampled EventTyp materials, recruited human participants, selected LLM configurations, and standardized prompting. These choices aimed to support comparable human and model assessments of event typicality.
B.5 Data Sample
- After filtering, the sample contains 2,330 tuples with unique contexts and actions, plus 24 tuples from Kravtchenko and Demberg (2015).The authors note that some contexts may still cover similar topics because action extraction was not strict.Source: §B.5 · p. 18
C.1 Human Participant Details
- Human participants were US-based English speakers with strong Prolific histories; inference-experiment participants also had to pass the pragmatics pre-test.The commonsense experiment recruited 354 participants and required 330 complete submissions.Source: §C.1 · p. 20
C.2.1 LLM version selection
- The LLM selection aimed for comparable open-weight configurations, including disabling Qwen3 thinking and using Olmo-3.1-Instruct for the larger Olmo prompting model.The authors sought to make CoT prompting available across models.Source: §C.2.1 · p. 20
C.2 LLM Details
- For LLM likelihood ratings, the study used both direct and Chain-of-Thought prompting with instruction-tuned versions of the models used for surprisal.The four response options were appended to prompts to avoid a bias toward “Likely.”Source: §C.2.2 · p. 21
C.2.2 Detailed Setup of Prompting
- The prompting experiment sampled 20 answers per prompt, or 10 for GPT-5.4, and reversed response-option order for half the samples.The study used direct and Chain-of-Thought prompting settings.Source: §C.2.2 · p. 21
D.1 Experimental Setup
The experimental setup turns event pairs into validated likelihood questions and narrows the inference study to selected scenarios. It also defines the snark label and screens human participants for pragmatic reasoning ability.
D.1 Experimental Setup
- The study converted event pairs into likelihood questions and had humans validate each question and correct errors such as pronoun ambiguity.Questions followed the format “When {Context}, how likely is it that {Followup}?”Source: §D.1 · pp. 21–22
E Event-Typicality Inference Experiment
- The human-facing interface presented likelihood-rating questions drawn from EventTypGold.Figure 9 gives an example of the format participants saw.Source: E Event-Typicality Inference Experiment · p. 22
E.1 Data Sample
- The experiment sampled 450 tuples, restricting contexts to events that take a day or less and occur at least yearly.The authors set aside other tuples while describing them as high quality for future follow-up studies.Source: §E.1 · p. 22
E.2 Dialogue Scenarios: Name selection
- Rare names from non-English-speaking countries were chosen to reduce the chance that models had seen the exact sentences in training data.Names were balanced by gender and sampled from 198 names.Source: §E.2 · p. 22
E.3 More explanation of the snarky label
- Snark is an atypicality inference that becomes ironic when the event concerns meeting a social convention or minimal standard.The paper describes snark as a critical or mocking attitude behind a seemingly positive utterance.Source: §E.3 · p. 23
E.4 Pragmatics Pre-test for Human Participants
- The 18-question pragmatics pre-test covered social conventions and commonsense-based coherence, and 676 of 1,196 approved participants passed.Passing required at least 5 of 6 correct in each question group and completion within seven minutes.Source: §E.4 · pp. 23–24
E.5 Ordinal Regressions
Human responses showed the predicted inference patterns across follow-up and setting conditions, while only a subset of LLM configurations consistently matched those patterns. Greater commonsense agreement also correlated with more typicality inferences.
E.5 Ordinal Regressions
- All interactions in the human follow-up-by-setting and snark models were significant, motivating the subset comparisons in Table 10.The reported human analysis used 13,500 rating responses.Source: §E.5 · pp. 24–25, Table 8 · p. 24, Table 9 · p. 25
- Four of five CoT-prompted LLM configurations consistently made human-like atypicality and typicality inferences and showed the snark effect.The snark effect is present only when atypicality inferences are made.Source: §E.5 · p. 25
E.6 Additional Figures
- Higher commonsense knowledge agreement correlated with more event-typicality inferences (R2 = 0.235, p < 0.01).The correlation used commonsense knowledge agreement to predict the inference score.Source: §E.6 · p. 25, Figure 13 · p. 25
F Additional Discussion
Matching EventTypGold labels to LLM-voted typicality did not improve inference scores. The authors suggest that sensitivity to question wording, rather than underlying event knowledge, could explain why relabeling made little difference.
F Additional Discussion
- Relabeling events to match LLM-voted typicality left inference scores about the same as using human labels.Figures 14 and 15 compare human and LLM typicality-inference rates for events assigned the corresponding follow-up labels.Source: F Additional Discussion · p. 26, Figure 14 · p. 26, Figure 15 · p. 26
- The authors suggest that if LLMs respond to question framing rather than underlying event knowledge, changing event labels would not help.They connect this possibility to the lack of consistency shown in Figure 7.Source: F Additional Discussion · p. 26
G Metric definitions
This appendix section defines the metrics and annotation prompts used to operationalize judgments about responses, agreement, consistency, inference, and entailment or plausibility. It also specifies criteria for event and utterance properties, including narration, frequency, temporal relations, wording, coherence, typicality-question suitability, and standalone status.
G Metric definitions
- Responses use four integer ratings mapped to Very Unlikely, Unlikely, Likely, and Very Likely; majority voting labels each follow-up Likely when most ratings are 3 or 4, and Unlikely when most are 1 or 2.The dataset has 2,354 EventTyp datapoints and 450 EventTypGold datapoints.Source: G Metric definitions · pp. 26–27
- The appendix defines three agreement scores using Cohen’s Kappa and abbreviates Combined as Comb.Source: G Metric definitions · p. 27
- Inference score measures the change in average response from the NONE baseline to a Plain or HeyGW setting; positive, negative, and zero values indicate atypicality, typicality, and no inference, respectively.Consistency measures the proportion of datapoints where the Commonsense Knowledge vote matches the NONE-setting vote in the Event-Typicality Inference experiment.Source: G Metric definitions · p. 27
H.1 Abstract
- The prompts ask whether a statement concerns an abstract action, defined as an action that may encompass a series of undefined, highly contextual, and potentially abstract steps.Source: §H.1 · p. 27
H.2 Entailment or near-entailment
- Entailment prompts test direct entailment, whether falsifying the second statement would require death or apocalypse, and whether it is part of the first action’s definition.The prompts are described as applying to all Context-Event pairs.Source: §H.2 · p. 28
- Entail-Para asks whether statement B can be seen as a paraphrase of statement A, distinguishing a ticket being taken after it is received from an unrelated event.The examples label the ticket paraphrase Yes and the event example No.Source: §H.2 · p. 29
- Entail-Plausible asks whether it would take mental gymnastics to imagine statement B being false given statement A, with examples covering inferred plans and event outcomes.Examples mark Mary having plans after inviting Linda as Yes and Linda accepting as No.Source: §H.2 · p. 29
H.3 Not sufficiently likely
- The Listing 19 prompt is described as the result of iterative prompt engineering after more intuitive prompts led the LLM to reject almost everything or almost nothing.Source: §H.3 · p. 29
- Not-sufficiently-likely prompts judge surprise directly, flag unexpected added information such as new entities, or identify interpretations that depart from a default reading.The direct-surprise prompt accepts surprisal of 4 or less on a 0–10 scale; the photo example treats storing the photo as a non-default camera interpretation.Source: §H.3 · pp. 29–30
H.4 Malformed context
- Malformed-A-Base asks whether a context statement is ungrammatical, misspelled, malformed, or otherwise nonsensical.Examples label “Mary loses Linda arm” malformed and “Mary loses Linda’s keys” not malformed.Source: §H.4 · p. 30
- The malformed-context check asks whether a statement is awkward because it redundantly states information already implied by the event.Examples include specifying that shoes go on feet or that grabbing uses a hand.Source: §H.4 · p. 30
H.5 Malformed event utterance
- The malformed-event checks ask whether utterances are ungrammatical or nonsensical and whether both statements use simple present tense.A separate coherence check asks whether the statements make sense and are related to each other.Source: §H.5 · p. 31
- The malformed-event coherence check asks whether statements A and B make sense and are related to each other.Source: §H.5 · p. 31
H.6 Utterance not brief
- The brevity check asks whether statement B can lose minor descriptive modifiers without changing its meaning, but not whether it can be deleted entirely.If the whole statement could be removed, the prescribed answer is No.Source: §H.6 · p. 31
H.7 Strange use of definites
- The definiteness check asks whether an indefinite or possessive object in statement B can sensibly be replaced by “the” or a possessive form.It applies only to subaction events.Source: §H.7 · p. 32
H.8 Narration
- The narration check asks whether paired statements seem more suited to a novel than a text message or online forum.Its clues include colorful or literary verbs, minute actions, and information resembling internal monologue.Source: §H.8 · p. 32
- Narration uses paired context-event statements with yes/no judgments; the examples distinguish spill-and-cleanup sequences and other event pairs.The narration listing applies to all context-event pairs.Source: §H.8 · pp. 32–33
H.9 Context event not sufficiently frequent
- The frequency criteria flag abnormal or hyperspecific events and lifetime milestones, while distinguishing milestones from events that recur in jobs or hobbies.The examples identify graduation as a once-in-a-lifetime event and childbirth as significant even if it can happen more than once.Source: §H.9 · p. 33
H.10 Time-gap
- The time-gap metric asks whether statements A and B have a very wide temporal or conceptual gap, and applies to next-action events.Examples contrast watering plants with their staying healthy versus a person thanking the waterer.Source: §H.10 · p. 33
H.11 Time-scope
- The time-scope metric asks whether statement A is much wider in scope than statement B, and applies only to subaction events.Examples contrast a broad activity such as playing chess with a specific subaction such as planning the next move.Source: §H.11 · p. 34
H.12 Passive
- The passive metric classifies whether a context utterance uses passive phrasing, contrasting “gets told off” with “tells off.”The listing says this metric applies to all context utterances.Source: §H.12 · p. 34
H.13 Not standalone
- The standalone metric asks whether a statement makes sense without extra context or reference to a prior context.Examples contrast “Mary leaves the shop” with “Mary goes back to the shop.”Source: §H.13 · p. 34
- The H.13 listing includes a “Standalone” prompt applied to all context utterances, with “No” shown for “Mary finds the keys at home.”The supplied excerpt does not include the prompt’s full criterion.Source: §H.13 · p. 34
H.14 Too vague
- The H.14 prompt labels statements extremely underspecified only when a core part of their meaning seems missing, distinguishing this from merely general wording.For instance, “Mary reacts” is too vague after seeing “the look” but not after seeing “the angry look.”Source: §H.14 · p. 34
H.15 No typicality
- The H.15 prompt accepts a “When…, how often…?” question only when it is meaningful for a situational relationship, not an inherent trait.Examples distinguish petting a stray dog from being an animal lover, and taking a year to build a house from another person moving into it.Source: §H.15 · p. 35
H.16 Prompts that had no effect in the end
- The H.16 prompts tested because, result-in-either-direction, and habituality judgments; the subsection identifies them as having no effect in the end.The prompts define because as more likely than “as a result,” ask whether one statement directly results from another, and test whether a statement is explicitly habitual.Source: §H.16 · pp. 35–36