Documentation

Linglib.Phenomena.Numerals.Compare

Bridge Theorems: Numeral Salience across Frameworks #

@cite{blok-2015} @cite{claus-walch-2024} @cite{cummins-2015} @cite{cummins-franke-2021} @cite{lasersohn-1999} @cite{woodin-etal-2023}

Connects the graded roundness model (k-ness) to ten existing modules:

  1. NSAL ↔ RSA cost: OT NSAL violations as normalized RSA utterance cost
  2. Woodin frequency ↔ RSA prior: weighted roundness as utterance prior
  3. k-ness ↔ PrecisionMode: roundness score grounds Kao et al.'s binary switch
  4. k-ness ↔ NumeralModifiers: tolerance modifiers pair with high roundness
  5. k-ness ↔ C&F enrichment: wider enrichment for rounder numerals
  6. OT ↔ RSA parameter map: constraint-to-parameter correspondence
  7. Evaluative valence ↔ framing: @cite{claus-walch-2024} framing predictions
  8. maxMeaning ↔ HasDegree: degree bridge theorems
  9. NumeralTheory ↔ RSA L1: lower-bound semantics + RSA derives exact readings
  10. Kennedy alternatives ↔ RSA: Class A/B competition via RSAConfig + rsa_predict

Architecture #

Phenomena.Gradability.Imprecision.Numerals (k-ness core)
    ↑ ↑ ↑
    |              |                |
Phenomena. NeoGricean. Semantics.Montague.
NumberUse. Constraints. Domain.Degree
WoodinEtAl2024 NumericalExprs (extended)
    ↑ ↑ ↑
    +--------------+-------+--------+
                           |
               Phenomena.Numerals.Compare (this file)

NSAL as RSA Utterance Cost #

In RSA, cost : U → ℚ penalizes certain utterances. The OT constraint NSAL provides a principled grounding: cost(u) = nsalViolations(u) / 6.

Round numerals (100, 1000) have cost ≈ 0; non-round (7, 99) have cost ≈ 1. This connects @cite{cummins-2015}'s constraint-based account to the Bayesian RSA framework via the cost field of RSAScenario.

Weighted Roundness as Utterance Prior #

In RSA, a uniform utterance prior is standard. But Woodin et al.'s frequency data suggests round numerals are a priori more available to speakers. weightedRoundnessScore provides an empirically-grounded utterance prior: rounder numerals are more likely to be chosen, all else being equal.

Roundness Grounds Precision Mo@cite{kao-etal-2014-hyperbole} use a binary PrecisionMode (.exact/.approximate) with #

Goal.approxPrice using fixed base := 10. The k-ness model provides a principled threshold: score ≥ 2 → .approximate, else .exact.

This means:

Precision mode agrees with Kao et al.'s implicit assumptions. Round numbers (multiples of 10) get approximate mode.

Fixed base-10 rounding and adaptive precision mode agree on round numbers: if n is round (divisible by 10), inferPrecisionMode gives.approximate.

Roundness and Tolerance Modifiers #

Numeral modifiers like "approximately" and "around" interact with roundness:

The k-ness model predicts this: tolerance modifiers combine naturally with high-roundness numerals because the pragmatic halo is already wide.

"approximately" does not require roundness but pairs better with it. The requiresRound field is false, but naturalness correlates with score.

Roundness Predicts Enrichment Width #

@cite{cummins-franke-2021} show that "more than M" undergoes pragmatic enrichment to "between M and M+δ" for some δ. The enrichment width δ depends on the roundness of M:

The wider enrichment for 100 admits more non-goal worlds, weakening its argumentative strength — explaining C&F's pragmatic reversal.

OT Constraint ↔ RSA Parameter Correspondence #

OT ConstraintRSA ParameterConnection
INFOφ (agreement)Both measure truth-conditional informativeness
GranularityQUD / GoalBoth encode contextual precision requirements
QSIMPcost (additive)Modifier complexity as utterance cost
NSALcost (roundness)k-ness violations as utterance cost

This mapping is not formal isomorphism but conceptual correspondence: both frameworks explain the same empirical patterns (round number preference, context-sensitivity) through different mechanisms.

Evaluative Valence Predicts Framing Direction #

@cite{claus-walch-2024} show that "at most" and "up to" have the same truth conditions but opposite framing effects. The evaluativeValence field in NumeralModifierEntry predicts this:

ModifierValencePredicted framingObserved framing
"at most"negativereversedreversed
"up to"positivestandardstandard

The prediction: negative valence → reversed framing, positive/neutral → standard.

"at most" has negative evaluative valence, which predicts reversed framing.

The formal valence field aligns with C&W's experimental observation that "at most" is endorsed more in negative contexts.

"up to" has positive evaluative valence, which predicts standard framing.

The formal valence field aligns with C&W's experimental observation that "up to" is endorsed more in positive contexts.

Degree Bridge Theorems #

Connect the pure-Nat maxMeaning comparisons to HasDegree.degree (ℚ-valued). The CardinalityDegree instance maps Nat to via cast; these theorems prove the two representations agree for all five OrderingRel variants.

A type with a natural-number cardinality measure.

Equations

Exact bare numeral meaning equals degree equality (@cite{bylinina-nouwen-2020} §4).

Bridge 9: NumeralTheory ↔ RSA L1 #

Standard RSA (L0→S1→L1) with bare numeral utterances over a 0-3 cardinality domain. Under lower-bound semantics (≥), RSA pragmatic reasoning derives the exact reading as a scalar implicature:

Reimplements NumeralTheory.runL1 using RSAConfig + rsa_predict.

Finite cardinality type (worlds 0-3).

Instances For
    Equations
    • One or more equations did not get rendered due to their size.
    Instances For
      Equations
      • One or more equations did not get rendered due to their size.

      Utterance type for standard numeral words.

      Instances For
        Equations
        • One or more equations did not get rendered due to their size.
        Instances For
          Equations
          • One or more equations did not get rendered due to their size.

          Lower-bound meaning: "one" = ≥1, "two" = ≥2, "three" = ≥3. Inlined for rsa_predict reification (avoids maxMeaning indirection).

          Equations
          Instances For

            Lower-bound numeral RSA: bare numerals under ≥ semantics with belief-based S1 (score = L0^α).

            Equations
            • One or more equations did not get rendered due to their size.
            Instances For

              Under lower-bound semantics, RSA strengthens "two" from ≥2 to the exact reading: L1 assigns more probability to w=2 than w=3.

              RSA strengthens "one" analogously: L1("one", w=1) > L1("one", w=2).

              "Three" trivially peaks at w=3 (only compatible world in the 0-3 range).

              The inlined meaning agrees with LowerBound.meaning (grounding).

              Bridge 10: Kennedy Alternative Sets through RSA #

              @cite{kennedy-2015}

              @cite{kennedy-2015}'s alternative sets for modified numerals through RSA L1. Lower alternatives for n=3: {bare 3, more than 3, at least 3}. Upper alternatives for n=3: {bare 3, fewer than 3, at most 3}.

              Under bilateral (exact) bare-numeral semantics, RSA predicts:

              Reimplements kennedyLowerL1, kennedyUpperL1, and all Kennedy alternative set theorems using RSAConfig + rsa_predict.

              Wider cardinality range (0-5) for modified numeral competition.

              Instances For
                Equations
                • One or more equations did not get rendered due to their size.
                Instances For
                  Equations
                  • One or more equations did not get rendered due to their size.

                  Lower-bound Kennedy alternatives for n=3.

                  Instances For
                    Equations
                    • One or more equations did not get rendered due to their size.
                    Instances For
                      Equations
                      • One or more equations did not get rendered due to their size.

                      Upper-bound Kennedy alternatives for n=3.

                      Instances For
                        Equations
                        • One or more equations did not get rendered due to their size.
                        Instances For
                          Equations
                          • One or more equations did not get rendered due to their size.

                          Kennedy lower-bound alternatives through RSA L1 (bilateral bare semantics).

                          Equations
                          • One or more equations did not get rendered due to their size.
                          Instances For

                            Kennedy upper-bound alternatives through RSA L1 (bilateral bare semantics).

                            Equations
                            • One or more equations did not get rendered due to their size.
                            Instances For

                              Class B competition at boundary: at w=3, "bare 3" beats "at least 3" in L1. The speaker who knew exactly 3 would use "bare 3" (more informative), so a listener hearing "at least 3" infers w≥4 is more likely.

                              Class A excludes the boundary: "more than 3" is false at w=3, so L1(w=4 | "more than 3") > L1(w=3 | "more than 3").

                              Bare numeral is peaked: L1("bare 3", w=3) > L1("bare 3", w=4). Under exact semantics, "bare 3" is only true at w=3.

                              Class B strengthened above bare: L1("at least 3", w=4) > L1("at least 3", w=3). The pragmatic listener hearing "at least 3" infers the speaker didn't know exactly 3 (ignorance implicature pushes probability above the boundary).

                              Upper Class B strengthened below bare: L1("at most 3", w=2) > L1("at most 3", w=3). Hearing "at most 3" pushes probability below the boundary (ignorance).