PaperlayerView exact revision on arXiv

Source-linked AI summary

How broad is that claim? Mapping Generalisation in NLP Research

Chenxin Diao, Nataliya Stepanova, Emily Allaway

arXiv:2609.14770v2cs.CLcs.AI

TL;DR

Generalisations can extend scientific findings but may obscure variability and exceed what the evidence supports, while existing approaches do not readily scale or cover diverse scientific claims. This paper introduces a taxonomy and LLM-powered framework, then uses them to build a labelled NLP-paper corpus. The corpus analyses show that hedging and vague descriptors are associated with generalisations, while the taxonomy is scoped to NLP outcome claims.

  • Problem

    Generalisations can obscure variability and suggest conclusions beyond the evidence, while their complexity makes automated identification difficult and prior annotation approaches have scale or coverage limits.

    Source: §1 · p. 1

  • Method

    The paper defines five generalisation categories, operationalises them in an LLM annotation framework, validates annotations with humans, and builds a corpus of labelled NLP-paper sentences.

    Source: §1 · p. 2

  • Results

    Hedging and vague descriptors are associated with generalisations in NLP papers, and the framework produced a corpus of 180K labelled sentences from 4.5K documents.

    Source: §1 · p. 2

  • Takeaways & Limitations

    NLPGENS supports analysis of generalisation patterns across venues and subdomains and their associations with linguistic features and citation counts.

    Source: §1 · p. 2

    The taxonomy covers outcome claims in NLP papers, not background claims about other authors, and may miss edge cases such as nested claims or pronoun ambiguity.

    Source: Limitations · p. 10

Abstract

from arXiv · show

Generalisations are common in scientific communication, even though they are semantically ambiguous. An automated method is needed to identify and categorise claims according to their level of generalisation, in order help detect an over-reliance on generalisations and possible misrepresentations of scientific findings. We introduce a comprehensive taxonomy of generalisations in the scientific domain, NLPGenX, which labels claims according to their level of generality and framing within the text. We operationalise this taxonomy with an LLM-powered framework, NLPGenA, that automatically classifies sentences from scientific articles into 5 different generalisation classes. We validate our framework with human annotators and use the framework to construct a large-scale dataset of NLP papers annotated according to generality, with auxiliary labels for hedging and vague descriptors (NLPGens). We use NLPGens to analyse the use of generalisations in NLP papers across multiple venues and subdomains, and to examine associations with citation counts, hedging, and vague descriptors.

1 Introduction

Generalisations help researchers extend findings but can obscure variability and invite interpretations beyond the evidence. The paper introduces a taxonomy, an automated annotation framework, and a dataset for studying generalisations in NLP papers.

1 Introduction

  • Generalisations can extend concrete findings to broader implications, but may obscure variability and exaggerate what evidence supports.The authors warn that this can lead to potential misinterpretations.Source: §1 · p. 1
  • Prior definitions of generalisations are too abstract or syntactically limited to transfer directly to scientific claims, motivating a new taxonomy.The taxonomy draws on studies of generalisations in science and captures outcome claims and their framing.Source: §1 · p. 1
  • NLPGENA uses LLMs to classify outcome-reporting sentences by generalisation type, addressing the scale limits of manual annotation and the narrow coverage of heuristic systems.The framework is validated against human annotations, with substantial agreement reported.Source: §1 · p. 2
  • NLPGENS contains 180K filtered and labelled sentences from 4.5K NLP papers, with labels for generality, hedging, and vague descriptors.The source papers span 2017–2024; the corpus supports analyses across venues and NLP subdomains, and exploratory analyses of citation counts.Source: §1 · p. 2
  • Hedging and vague descriptors are both associated with generalisations in NLP papers.The paper also examines how generalisation types vary across venues and subdomains and their association with citation counts.Source: §1 · p. 2

2 Prior Work

Prior work studies generalisations at different linguistic levels and scientific claims through several discourse lenses. This paper distinguishes its approach by labelling generality at the sentence level and providing a taxonomy intended for scientific claims.

2 Prior Work

  • Earlier generalisation schemas moved from noun phrases to clauses, events, and predicates, whereas this work annotates generalisations at the sentence level.The cited approaches include event and coreference taxonomies and classifications of events as episodic or habitual.Source: §2 · p. 2
  • Earlier automated methods used grammatical features or fine-tuned language models, which require labelled training data; this work uses zero-shot prompting with an instruction-tuned LLM.Source: §2 · p. 2
  • Research on scientific claims has addressed discourse roles, citation purposes, argumentation, and news headlines; this paper instead labels claim semantics through generality.These prior aspects are described as focusing primarily on pragmatics and claim relationships.Source: §2 · pp. 2–3
  • Prior scientific-article annotation schemas overlap but often use domain-specific definitions or cover limited claim sets, motivating a taxonomy with concrete, actionable definitions.The cited manual work concerns psychology, medicine, and experimental philosophy.Source: §2 · p. 3

3 NLPGENX: A Taxonomy of Scientific Outcome Generalisations

NLPGENX classifies outcome generalisations by their scope and framing, drawing together definitions from several fields. Its five categories distinguish claims that are specific from broader claims with different forms of framing and genericity.

3 NLPGENX: A Taxonomy of Scientific Outcome Generalisations

  • NLPGENX focuses on outcome generalisations and combines definitions from psychology, medicine, and experimental philosophy; transfer beyond NLP requires domain-specific validation.The taxonomy represents genericity (how broadly a claim applies) and framing (how it relates to the study).Source: §3 · p. 3

3.1 Dimensions of Generalisation

  • Framing links a generalisation to a particular study, while genericity concerns whether language explicitly qualifies how broadly a claim applies.Generic claims lack cues such as quantification; non-generic claims may qualify scope or refer to specific instances.Source: §3.1 · p. 3

3.2 Taxonomic Categories

  • The taxonomy’s five categories separate non-generalised claims from generalised claims, then distinguish generalised claims by framing and genericity.The three distinctions are scope, whether the generalisation is study-framed, and whether it is generic or non-generic.Source: §3.2 · p. 3
  • Bare generic claims express broad generalisations without an overt study frame or explicit scope cues.Source: §3.2 · p. 3
  • Framed generic claims embed a broadly expressed finding within explicit study framing; framed non-generic claims qualify broader application within that study context.Source: §3.2 · p. 3
  • Bare non-generic claims constrain scope without presenting the claim as a finding of the current study, while non-generalised claims report specific results without broader conclusions.Source: §3.2 · p. 3

4 NLPGENA: A Framework for Annotating Generalisations in NLP

NLPGENA applies NLPGENX through a sequence of classifiers that identify outcome claims and assign generalisation labels, with separate annotations for hedging and vague descriptors. The framework also distinguishes study framing and genericity using explicit criteria.

4 NLPGENA: A Framework for Annotating Generalisations in NLP

  • NLPGENA uses nine annotation tasks: seven filter outcome sentences and derive taxonomy labels, while two independently identify hedging and vague descriptors.The taxonomy covers outcome claims, and the framework applies its steps in sequence.Source: §4 · p. 4
  • A rule-based counterpart performs less well on tasks requiring semantic judgement, so the experiments use LLM-based classification consistently.Source: §4 · p. 4

4.1 Is it about outcomes?

  • Outcome sentences include descriptions or analyses of results, introductions of scientific artefacts, and claims about contributions to field knowledge.Source: §4.1 · p. 4

4.2 Is it generalised?

  • A claim is generalised when it has underspecified quantification, expresses causality, or changes little in meaning when “usually” or “generally” is inserted.Quantifiers with specific thresholds are excluded from the underspecified-quantification criterion.Source: §4.2 · p. 4

4.3 Is it framed?

  • A generalised claim is framed when it refers to the study population or related domain, or to the study through its authors, results, tables, or figures.Claims without a reference to the study are unframed.Source: §4.3 · pp. 4–5

4.4 Is it generic?

  • A generic claim must be present tense, lack quantification, and avoid reference to the study population; framed claims are assessed for genericity after removable framing is separated.A complete embedded sentence can remain generic after its study framing is removed.Source: §4.4 · p. 5

4.5 Hedging and Vague Descriptors

  • Hedging and vague descriptors are auxiliary labels that do not contribute to the taxonomy label; vague descriptors are adjectives or adverbs with unclear degree.Examples include “significant”, “good”, or “efficient” without an explicit threshold.Source: §4.5 · p. 5

5 NLPGENS: A Large-Scale Dataset of Generalisations in NLP Research

The corpus was assembled by extracting sentences from five standard paper sections and labeling outcome sentences across the taxonomy and auxiliary attributes. Verification found high classifier accuracy and agreement, while independent annotation was less stable for some semantic modules.

5.1 Implementation

  • Keywords identify sentences in Abstract, Introduction, Results, Discussion, and Conclusion sections of the NLPeer NLP subsets.Separate prompts query the LLM three times; majority voting and a decision tree produce final labels, with preceding sentences supplied as context for selected decisions.Source: §5.1 · p. 5

5.2 Dataset Statistics

  • 180,426 fully labelled outcome sentences span five taxonomy classes; 54.6% are gen/framed/non-generic and 31.6% are non-gen.The corpus also includes annotations for hedging and vague descriptors.Source: §5.2 · pp. 5–6

5.3 Human Validation Setup

  • Verification asks annotators whether framework outputs are correct, while independent annotation asks them to assign labels without seeing those outputs.For each protocol, three annotators review sampled items; applicable module samples balance positive and negative framework labels.Source: §5.3 · p. 6
  • Verification accuracy was at least 0.970 for every classifier, but independent annotation agreement varied widely across modules, with α from 0.065 to 0.933.The authors therefore use verification as the primary reliability result and treat independent annotation as diagnostic; the evaluated annotations come from one off-the-shelf openweight model.Source: §5.3 · p. 6
  • Each final taxonomy label receives a data-weighted reliability estimate composed from classifier accuracies along the decision-tree paths that produce it.The authors describe these estimates as usable to weight or filter annotated data in downstream analyses.Source: §5.3 · pp. 6–7

6 Analysis

Across five analyses of about_outcome sentences, the paper compares generalisation classes across sections and tracks and examines their associations with citations, hedging, and vague descriptors.

6.1 Trends Across Sections

  • Across venues, gen/* and non-gen claims are about 60:40; gen/framed/non-generic is most abundant, with most generalisations in discussion and most */generic claims in introductions.The fewest */generic claims appear in abstracts and results sections.Source: §6.1 · p. 7, Figure 3 · p. 8

6.2 Trends Across Tracks

  • Nearly half of Summar-Simpl sentences are non-gen, versus 15% in CogMod; Applications and Summar-Simpl have the most non-gen sentences, while Applications has among the lowest bare-generic rates.The authors leave a thorough investigation of track-level differences for future work.Source: §6.2 · p. 8, Figure 4 · p. 8

6.3 Generalisations and Citation Count

  • For ACL ’17, gen/framed/non-generic sentences correlate significantly with citation counts (Spearman ρ almost 0.3); similar trends elsewhere are not significant.Citation-count prediction is difficult using generalisability alone without semantic content.Source: Table 5 · p. 8, §6.3 · pp. 8–9

6.4 Hedging tracks generalisation

  • Hedging is 2.6 times more frequent in generalised than non-generalised outcome sentences (37.9% vs. 14.6%); 84.9% of hedged claims are generalised.The authors describe hedging as a strong, but imperfect, surface cue for generalisation.Source: §6.4 · p. 9
  • Among generalised labels, hedging is highest for gen/bare/non-generic (49.2%) and lowest for gen/framed/non-generic (36.7%).The authors suggest explicit framing may accompany less hedging, while bare generalisations attract more.Source: §6.4 · p. 9

6.5 Vague descriptors track generalisation

  • Vague descriptors occur in 80.2% of generalised sentences versus 64.8% of non-generalised sentences; the odds are 2.20 times higher, with a similar association across venues.Across venues, 72.8% of outcome sentences containing a vague descriptor are labelled generalised.Source: §6.5 · p. 9
  • The vague-descriptor association is exploratory: independent agreement for is_vague is low (α = 0.090), and shared LLM pipeline artefacts may contribute.The vague-descriptor and generalisation labels come from separate prompts in the same LLM pipeline.Source: §6.5 · p. 9

7 Conclusion

The paper introduces a taxonomy and automated framework for classifying how generalised claims about NLP findings are. It uses them to build a large annotated corpus and report preliminary patterns in NLP research.

7 Conclusion

  • NLPGENX classifies the generality of sentences communicating scientific findings in NLP research.It captures five generalisation categories and study-specific framing.Source: §7 · p. 9
  • NLPGENA operationalises the taxonomy with an automated ensemble of linguistics-inspired classifiers for assigning sentences to its five generalisation categories.The authors describe it as the first fully automated instruction-tuned-LLM solution for annotating generalisation in language, to their knowledge.Source: §7 · p. 9
  • NLPGENS contains 180K NLP claims labelled for generalisation, hedging, and vague descriptors, and its analyses reveal preliminary patterns across time, subdomains, citation counts, and claim attributes.The reported patterns include a growing relative percentage of generalisations and associations between generalised claims, hedging, and vague descriptors.Source: §7 · p. 9

Limitations

The taxonomy is validated for NLP outcome sentences, not all scientific claims. The authors also flag unresolved linguistic edge cases, uncertain transfer to other domains, and limitations of LLM-based annotation.

Limitations

  • The taxonomy was developed and validated for outcome sentences in NLP papers, so it does not cover background claims about other authors’ work.Nested claims with components from different categories and pronoun ambiguity may also fall outside its account.Source: Limitations · p. 10
  • The taxonomy’s scope leaves out background claims and may not capture nested claims or pronoun ambiguity.The authors describe generic semantics and complex scientific sentences as sources of these edge cases.Source: Limitations · p. 10
  • Development using computer-graphics papers suggests possible applicability to other empirical domains, but does not establish cross-domain validity.Differences in terminology and discourse conventions, especially in less constrained genres such as broad social-science discourse, may require adaptation and separate validation.Source: Limitations · p. 10
  • LLM-based annotation limits NLPGENA’s explainability, reproducibility, and efficiency.The rule-based baseline approaches NLPGENA on some surface-oriented decisions but has substantial gaps on semantic-judgement tasks and end-to-end classification.Source: Limitations · p. 10

A LLM Annotation Details

The annotation pipeline filters and processes paper sentences before assigning taxonomy labels, while keeping hedging and vagueness as auxiliary attributes. Human verification audits the pipeline’s decisions using sampled rows and accounts for the conditional extraction step.

A LLM Annotation Details

  • The extraction process retains sentences from sections matching specified headings and quarantines papers that fail the Introduction coverage check.Paragraphs are segmented into sentences, and every non-empty Introduction token must reappear in the extracted sentences after punctuation-and-whitespace normalisation.Source: A LLM Annotation Details · p. 12
  • about_outcome gates the pipeline; five taxonomic classifiers feed a rule-based final label, while hedging and vagueness are measured separately on the raw claim.get_target_sent runs only when the reporting-clause and embedded-sentence conditions hold; its output and, for two classifiers, two-sentence context inform downstream classification.Source: A LLM Annotation Details · p. 12, Figure 5 · p. 13
  • Only 13 cells remained missing and 94 were tie-broken after majority voting across three LLM annotation passes.The authors report 2,019,550 majority-voted classifier cells and conclude that the corpus is effectively complete at scale.Source: A LLM Annotation Details · p. 13

B Human Verification Details

  • Three annotators independently audited whether each sampled LLM decision was correct, using the same inputs available to the model.For Boolean modules, the authors reconstructed verifier-implied task labels before computing nominal Krippendorff’s α; this reconstruction was unavailable for get_target_sent.Source: B Human Verification Details · pp. 13, 15
  • The verification sample used 200 rows per module; taxonomic and auxiliary exports drew on a shared 100-row about_outcome-True core and were balanced by each module’s label.get_target_sent was sampled separately from eligible rows, with additional about_outcome-True rows used to meet quotas where needed.Source: B Human Verification Details · p. 14
  • The reliability accounting treats get_target_sent’s conditional invocation asymmetrically, distinguishing a wrong extraction from a missed call when both trigger conditions are truly met.An over-call is not counted as a failure because the prompt is designed to return the original sentence in that case; adding a separate skip-veto factor on non-trigger paths would double-count.Source: B Human Verification Details · p. 15

C Independent Annotation Details

Independent annotation tests whether people reproduce NLPGENA’s labels without seeing its answers, while verification tests whether its answers are defensible. The section also reports a merged removable-framing evaluation, where independent agreement is lower than verification agreement.

C Independent Annotation Details

  • Independent agreement varied by module: α was 0.065 for is_generalised and 0.090 for is_vague, but 0.933 for is_present_tense under both protocols.Agreement was also lower independently for ref_study_population (0.428) and sentence_embedded (0.415).Source: C Independent Annotation Details · p. 16
  • For target-sentence extraction, the human majority matched NLPGENA on 173 of 200 items (Acc = 0.865), while three normalized strings were identical on 138 items.The authors report raw pairwise agreement rather than label-level α for this free-text operation.Source: C Independent Annotation Details · pp. 15–16
  • The final framework label matched independent human-majority decisions on 48 of 100 items, versus 96 under verification; the two human-implied labels agreed on 61.Verification asks whether a proposed output is defensible, whereas independent annotation tests whether annotators reproduce it without seeing it.Source: C Independent Annotation Details · pp. 16–17
  • Independent annotation gives a less stable reference for some semantic distinctions, which contextualizes why deterministic surface rules may not reproduce them reliably.The authors contrast this with highly reproducible surface-oriented modules such as present-tense detection.Source: C Independent Annotation Details · pp. 16–17

D Composite Label: removable_framing

  • The composite evaluation merges reporting-clause detection, sentence embedding, and target-sentence extraction; extraction is required for correctness only when the framework invokes it.Only 28 of the 100 evaluated rows had both framing conditions true and therefore received a target-sentence human vote.Source: D Composite Label: removable_framing · p. 17
  • 97 of 100 composite decisions matched NLPGENA under verification, compared with 87 under independent annotation (24 of 28 positive and 63 of 72 negative outputs).The composite accuracy includes target-sentence extraction when the framework invokes it; the reported Krippendorff’s α measures only the Boolean framing gate.Source: D Composite Label: removable_framing · p. 17

E Rule-based Baseline

The rule-based baseline translates NLPGENA’s prompts into deterministic rules and runs them through the same module order and decision tree. It approximates some components, but its end-to-end accuracy and the small, imbalanced audit data constrain what can be concluded about replacement or supervised alternatives.

E Rule-based Baseline

  • The baseline converts prompt instructions into regular-expression, lexicon, and spaCy feature rules, then applies the same module order and final decision tree as NLPGENA.The rules were frozen before evaluation and did not use LLM predictions, rationales, or human annotations.Source: E Rule-based Baseline · p. 18
  • The rule baseline operationalizes outcome detection, reporting clauses, embedded sentences, study references, population references, tense, quantification, and generalisation with surface patterns and linguistic features.For example, the target-sentence rule extracts a finite complement only when both reporting-clause and sentence-embedded checks pass.Source: E Rule-based Baseline · p. 18
  • The rules approach NLPGENA on ref_study and is_present_tense, suggesting these components may be candidates for deterministic replacement.The authors note that such replacement could reduce LLM calls and improve reproducibility.Source: E Rule-based Baseline · pp. 18–19
  • 0.120 final-category accuracy shows that wholesale replacement of NLPGENA by prompt-derived rules is not viable.The authors attribute the result to difficulty capturing semantic distinctions such as generality and study-population references, with component errors propagating through the pipeline.Source: E Rule-based Baseline · p. 19, Table 8 · p. 19
  • A supervised baseline was not meaningfully comparable because the non-shared training rows were severely imbalanced and the independent round was too small.Examples include 99:1 imbalance for outcome detection and 89:11 for present-tense detection; larger independently labelled data are left for future work.Source: E Rule-based Baseline · p. 19

F Dataset Examples and Error Analysis for Annotation

The error analysis identifies recurring boundary problems in the annotation decision tree, while the venue and track analyses describe variation in outcome-sentence rates and generalisation rates. Track-level patterns differ across venues, and the authors caution that the soft assignments are imperfect.

F Dataset Examples and Error Analysis for Annotation

  • Errors cluster around the generalisation boundary, study framing versus population references, and genericity judgments after reporting frames are removed.Examples include treating specific numerical results as generalisations, mistaking an outcome name for a study population, and evaluating the reporting clause instead of the embedded proposition.Source: F Dataset Examples and Error Analysis… · p. 19

G about_outcome Across Sections

  • Most paper sections have relatively consistent about_outcome rates across tested venues, with the largest range in conclusions.Table 10 reports each section’s share of all sentences in that venue, not an average per paper.Source: G about_outcome Across Sections · p. 19, Table 10 · p. 21

H Track Assignment

  • Generalisation rates differ across NLP subdomains, but track orderings are inconsistent across venues and years.Tracks for papers without track metadata were assigned to the closest COLING-2020 track centroid using title-and-abstract embeddings; the authors describe this soft labelling as naive and say more gold track labels may be needed for robust analysis.Source: H Track Assignment · p. 21

I Citation Analysis

The citation analysis relates generalisation classes to paper citation counts using Spearman correlations. Outside ACL ’17, the reported correlations are not statistically significant, and the authors note that citation distributions and metadata coverage constrain interpretation.

I Citation Analysis

  • Citation-count analysis excludes EMNLP ’23 because its papers lacked the ACL Anthology URLs needed to retrieve counts, and other venues’ citation distributions may be too clustered or unstable to reveal strong effects.The authors note a spike near zero citations for EMNLP ’24 and more spread in ACL ’17.Source: I Citation Analysis · pp. 21–22
  • For COLING ’20, ARR ’22, and EMNLP ’24, none of the correlations between generalisation classes and citation counts are statistically significant.The gen/framed/* class is among the more positively correlated classes in COLING ’20 and ARR ’22, consistent with the ACL ’17 observation, but the effects are not statistically strong.Source: I Citation Analysis · p. 21

J Data Access, Licensing, and Intended Use

This section describes NLPeer access, licensing, and bounded research use, including privacy and offensive-content limitations. It also outlines prompt templates for outcome and generalisation decisions and claim-level checks for complement clauses, reported content, hedging, and under-specified modifiers.

J Data Access, Licensing, and Intended Use

  • NLPeer v17/v28 were accessed through official TUdatalib requests, and NLPGENS annotates parsed, section-filtered manuscript text.The source subsets carry varying Creative Commons terms; the authors conduct non-commercial academic research with attribution, do not redistribute the raw corpus, and release NLPGENS under CC BY-NC-SA 4.0 with source licences retained. They report no separate data-use agreement.Source: J Data Access, Licensing, and Intended… · pp. 22–23
  • NLPGENS fits NLPeer’s intended research scope by analyzing linguistic properties of manuscript claims, not peer reviews, profiling, review generation, or scoring.NLPeer is intended for computational study of peer review and NLP-based reviewing assistance; the authors attribute this intended-use framing to Dycke et al. (2023).Source: J Data Access, Licensing, and Intended… · p. 23

K Personal Data and Offensive Content

  • The study excludes F1000-22, avoids profiling or demographic inference, and reports no separate systematic offensive-content screen beyond venue publication norms.NLPeer had stripped identifying metadata from review reports in other subsets; the authors judged offensive-content risk low because inputs were accepted NLP/CL papers.Source: K Personal Data and Offensive Content · p. 24

L Prompt Templates

  • The outcome prompt treats introductions, results, discussion, contributions, conclusions, and limitations as outcomes, while excluding background, setup, methods, navigation, errors, and future work.It classifies mixed outcome/non-outcome claims as outcomes, and says hedging or opinion does not change the decision.Source: L Prompt Templates · p. 25
  • The generalisation prompt ignores hedging, marks explicit causality and non-specific frequency as generalised, and uses frequency checks or a usually/generally insertion test otherwise.Specific, 100%, or 0% frequencies count as not generalised; mixed generalised and non-generalised claims are labelled generalised. The insertion test allows collective, distributive, or temporal readings as context fits.Source: L Prompt Templates · p. 26
  • Other prompts identify study references, experimental populations, present-tense main clauses, non-numeric quantifiers, reporting verbs, and complement clauses.The instructions define quantifiers as words limiting extent, frequency, or proportion, and reporting verbs as the main verbs in clauses attributing speech or thought.Source: L Prompt Templates · pp. 27–29
  • The complement-clause prompt distinguishes subordinate clauses from noun phrases by testing whether the clause can be replaced with a relative wh-pronoun such as “who” or “which.”The examples label complement clauses Yes and noun phrases No.Source: L Prompt Templates · pp. 29–30
  • The reporting-clause prompt extracts complete reported sentences, including relevant subordinate clauses, but returns the original sentence when the content is incomplete or the sentence is not a reporting clause.Examples retain “whereas” and “by incorporating” clauses, while “Table 4 shows the experimental results” remains unchanged because its reported content is not a complete sentence.Source: L Prompt Templates · p. 30
  • The hedging prompt counts only wording that explicitly signals uncertainty, possibility, approximation, limited generality, or reduced commitment, and instructs the model not to infer hedging.It asks the model to state the hedging expression before answering Yes or No.Source: L Prompt Templates · p. 31
  • The under-specified-modifier prompt asks whether “how <adjective/adverb>?” is a valid question, answers No when no adjective or adverb is present, and counts a sentence as under-specified if any modifier is under-specified.Examples treat “how significant?” as valid but “how true?” as invalid.Source: L Prompt Templates · p. 31
Loading 2609.14770v2…